The following is a conversation with Dylan Patel and Nathan Lampert. Dylan runs Semianalysis, a well-respected research and analysis company that specializes in semiconductors, GPUs, CPUs, and AI hardware in general. Nathan is a research scientist at the Allen Institute for AI, and is the author of the amazing blog on AI called "Interconnects." They are both highly respected, read, and listened to by the experts, researchers, and engineers in the field of AI.

And personally, I'm just a fan of the two of them. So I used the deep-seek moment that shook the AI world a bit as an opportunity to sit down with them and lay it all out. From deep-seek, open AI, Google, XAI, meta-anthropic to NVIDIA and TSMC, and to US, China, Taiwan relations, and everything else that is happening at the cutting edge of AI.

This conversation is a deep dive into many critical aspects of the AI industry. While it does get super technical, we try to make sure that it's still accessible to folks outside of the AI field by defining terms, stating important concepts explicitly, spelling out acronyms, and in general, always moving across the several layers of abstraction and levels of detail.

There is a lot of hype in the media about what AI is and isn't. The purpose of this podcast, in part, is to cut through the hype, through the bullshit, and the low-resolution analysis, and to discuss in detail how stuff works and what the implications are. Let me also, if I may, comment on the new OpenAI 03 Mini reasoning model, the release of which we were anticipating during the conversation, and it did indeed come out right after.

Its capabilities and costs are on par with our expectations, as we stated. OpenAI 03 Mini is indeed a great model, but it should be stated that DeepSeq R1 has similar performance on benchmarks, is still cheaper, and it reveals its chain-of-thought reasoning, which 03 Mini does not. It only shows a summary of the reasoning.

Plus, R1 is open weight, and 03 Mini is not. By the way, I got a chance to play with 03 Mini, and anecdotal vibe-check-wise, I felt that 03 Mini, specifically 03 Mini High, is better than R1. Still, for me personally, I find that Clawed Sonnet 3.5 is the best model for programming.

Except for tricky cases where I will use 01 Pro to brainstorm. Either way, many more better AI models will come, including reasoning models, both from American and Chinese companies. They will continue to shift the cost curve. But the "DeepSeq moment" is indeed real. I think it will still be remembered five years from now as a pivotal event in tech history, due in part to the geopolitical implications, but for other reasons too, as we discuss in detail from many perspectives in this conversation.

This is the Lex Friedman Podcast. To support it, please check out our sponsors in the description. And now, dear friends, here's Dylan Patel and Nathan Lambert. A lot of people are curious to understand China's DeepSeq AI models. So let's lay it out. Nathan, can you describe what DeepSeq v3 and DeepSeq R1 are, how they work, how they're trained?

Let's look at the big picture and then we'll zoom in on the details. - Yeah, so DeepSeq v3 is a new mixture of experts, transformer language model from DeepSeq, who is based in China. They have some new specifics in the model that we'll get into. Largely, this is a open weight model and it's a instruction model like what you would use in ChatGPT.

They also released what is called the base model, which is before these techniques of post-training. Most people use instruction models today and those are what's served in all sorts of applications. This was released on, I believe, December 26th or that week. And then weeks later on January 20th, DeepSeq released DeepSeq R1, which is a reasoning model, which really accelerated a lot of this discussion.

This reasoning model has a lot of overlapping training steps to DeepSeq v3 and it's confusing that you have a base model called v3, that you do something to to get a chat model and then you do some different things to get a reasoning model. I think a lot of the AI industry is going through this challenge of communications right now where OpenAI makes fun of their own naming schemes.

They have GPT-40, they have OpenAI-01 and there's a lot of types of models. So we're gonna break down what each of them are. There's a lot of technical specifics on training and go from high level to specific and kind of go through each of them. - There's so many places we can go here, but maybe let's go to OpenWeights first.

What does it mean for a model to be OpenWeights and what are the different flavors of open source in general? - Yeah, so this discussion has been going on for a long time in AI. It became more important since ChatGPT or more focal since ChatGPT at the end of 2022.

OpenWeights is the accepted term for when model weights of a language model are available on the internet for people to download. Those weights can have different licenses, which is effectively the terms by which you can use the model. There are licenses that come from history in open source software.

There are licenses that are designed by companies specifically. All of LLAMA, DeepSeek, Quen, Mistral, these popular names in OpenWeight models have some of their own licenses. It's complicated 'cause not all the same models have the same terms. The big debate is on what makes a model OpenWeight. Why are we saying this term?

It's kind of a mouthful. It sounds close to open source, but it's not the same. There's still a lot of debate on the definition and soul of open source AI. Open source software has a rich history on freedom to modify, freedom to take on your own, freedom from any restrictions on how you would use the software and what that means for AI is still being defined.

So for what I do, I work at the Allen Institute for AI. We're a nonprofit. We want to make AI open for everybody. And we try to lead on what we think is truly open source. There's not full agreement in the community, but for us, that means releasing the training data, releasing the training code, and then also having OpenWeights like this.

And we'll get into the details of the models. And again and again, as we try to get deeper into how the models were trained, we will say things like the data processing, data filtering, data quality is the number one determinant of the model quality. And then a lot of the training code is the determinant on how long it takes to train and how fast your experimentation is.

So without fully open source models where you have access to this data, it is hard to know or it's harder to replicate. So we'll get into cost numbers for DeepSeek v3 on mostly GPU hours and how much you could pay to rent those yourselves, but without the data, the replication cost is going to be far, far higher.

And same goes for the code. - We should also say that this is probably one of the more open models out of the frontier models. So like in this full spectrum where probably the fullest open source, like you said, open code, open data, OpenWeights. This is not open code.

This is probably not open data. And this is OpenWeights and the licensing is a MIT license or it's, I mean, there's some nuance in the different models but it's towards the free, in terms of the open source movement, these are the kind of the good guys. - Yeah, DeepSeek is doing fantastic work for disseminating understanding of AI.

Their papers are extremely detailed in what they do and for other teams around the world, they're very actionable in terms of improving your own training techniques. And we'll talk about licenses more. The DeepSeek R1 model has a very permissive license. It's called the MIT license that effectively means there's no downstream restrictions on commercial use.

There's no use case restrictions. You can use the outputs from the models to create synthetic data. And this is all fantastic. I think the closest peer is something like Llama where you have the weights and you have a technical report and the technical report is very good for Llama.

One of the most read PDFs of the year last year is the Llama 3 paper. But in some ways it's slightly less actionable. It has less details on the training specifics, I think less plots and so on. And the Llama 3 license is more restrictive than MIT. And then between the DeepSeek custom license and the Llama license, we can get into this whole rabbit hole.

I think we'll make sure we wanna go down the license rabbit hole before we do specifics. - Yeah, and I mean, so it should be stated that one of the implications of DeepSeek, it puts pressure on Llama and everybody else on open AI to push towards open source. And that's the other side of open source that you mentioned is how much is published in detail about it.

So how open are you with the sort of the insights behind the code? So like how good is the technical reports? Are they hand wavy or is there actual details in there? And that's one of the things that DeepSeek did well is they published a lot of the details.

- Yeah, especially in the DeepSeek v3, which is their pre-training paper. They were very clear that they are doing interventions on the technical stack that go at many different levels. For example, on there to get highly efficient training, they're making modifications at or below the CUDA layer for NVIDIA chips.

I have never worked there myself. And there are a few people in the world that do that very well. And some of them are at DeepSeek. And these types of people are at DeepSeek and leading American frontier labs, but there are not many places. - To help people understand the other implication of open weights, just, you know, there's a topic we'll return to often here.

So there's a fear that China, the nation, might have interest in stealing American data, violating privacy of American citizens. What can we say about open weights to help us understand what the weights are able to do in terms of stealing people's data? - Yeah, so these weights that you can download from Hugging Face or other platforms are very big matrices of numbers.

You can download them to a computer in your own house that has no internet, and you can run this model, and you're totally in control of your data. That is something that is different than how a lot of language model usage is actually done today, which is mostly through APIs, where you send your prompt to GPUs run by certain companies, and these companies will have different distributions and policies on how your data is stored, if it is used to train future models, where it is stored, if it is encrypted, and so on.

So the open weights are, you have your fate of data in your own hands, and that is something that is deeply connected to the soul of open source. - So it's not the model that steals your data, it's whoever's hosting the model, which could be China, if you're using the DeepSeek app, or it could be Perplexity.

You know, you're trusting them with your data, or OpenAI, you're trusting them with your data, and some of these are American companies, some of these are Chinese companies, but the model itself is not doing the stealing, it's the host. All right, so back to the basics. What's the difference between DeepSeek V3 and DeepSeek R1?

Can we try to like lay out the confusion potential? - Yes, so for one, I have very understanding of many people being confused by these two model names, so I would say the best way to think about this is that when training a language model, you have what is called pre-training, which is when you're predicting the large amounts of mostly internet text, you're trying to predict the next token, and what to know about these new DeepSeek models is that they do this internet large-scale pre-training once to get what is called DeepSeek V3 base.

This is a base model, it's just going to finish your sentences for you, it's going to be harder to work with than ChatGPT, and then what DeepSeek did is they've done two different post-training regimes to make the models have specific desirable behaviors. So what is the more normal model in terms of the last few years of AI, an instruct model, a chat model, a quote-unquote aligned model, a helpful model, there are many ways to describe this, is more standard post-training.

So this is things like instruction tuning, reinforce learning from human feedback, we'll get into some of these words, and this is what they did to create the DeepSeek V3 model. This was the first model to be released, and it is very high-performance, it's competitive with GPT-4, LLAMA-405B, so on.

And then when this release was happening, we don't know their exact timeline, or soon after, they were finishing the training of a different training process from the same next token prediction base model that I talked about, which is when this new reasoning training that people have heard about comes in in order to create the model that is called DeepSeek R1.

The R through this conversation is good for grounding for reasoning, and the name is also similar to OpenAI's O1, which is the other reasoning model that people have heard about. And we'll have to break down the training for R1 in more detail, because for one, we have a paper detailing it, but also it is a far newer set of techniques for the AI community, so it is a much more rapidly evolving area of research.

- Maybe we should also say the big two categories of training of pre-training and post-training, these umbrella terms that people use. So what is pre-training and what is post-training, and what are the different flavors of things underneath post-training umbrella? - Yeah, so pre-training, I'm using some of the same words to really get the message across, is you're doing what is called autoregressive prediction to predict the next token in a series of documents.

This is done over standard practice is trillions of tokens. So this is a ton of data that is mostly scraped from the web. In some of DeepSeek's earlier papers, they talk about their training data being distilled for math. I shouldn't use this word yet, but taken from Common Crawl, and that's a public access that anyone listening to this could go download data from the Common Crawl website.

This is a crawler that is maintained publicly. Yes, other tech companies eventually shift to their own crawler, and DeepSeek likely has done this as well, as most frontier labs do. But this sort of data is something that people can get started with, and you're just predicting text in a series of documents.

This can be scaled to be very efficient, and there's a lot of numbers that are thrown around in AI training, like how many floating point operations or flops are used. And then you can also look at how many hours of these GPUs that are used. And it's largely one loss function taken to a very large amount of compute usage.

You just set up really efficient systems. And then at the end of that, you have this base model, and pre-training is where there is a lot more of complexity in terms of how the process is emerging or evolving, and the different types of training losses that you will use.

I think this is a lot of techniques grounded in the natural language processing literature. The oldest technique, which is still used today, is something called instruction tuning, or also known as supervised fine-tuning. These acronyms will be IFT or SFT. People really go back and forth throughout them, and I will probably do the same, which is where you add this formatting to the model where it knows to take a question that is like, "Explain the history of the Roman Empire to me," or something, a sort of question you'll see on Reddit or Stack Overflow, and then the model will respond in a information-dense but presentable manner.

The core of that formatting is in this instruction-tuning phase, and then there's two other categories of loss functions that are being used today. One I will classify as preference fine-tuning. Preference fine-tuning is a generalized term for what came out of reinforcement learning from human feedback, which is RLHF. This reinforcement learning from human feedback is credited as the technique that helped ChatGPT break through.

It is a technique to make the responses that are nicely formatted, like these Reddit answers, more in tune with what a human would like to read. This is done by collecting pairwise preferences from actual humans out in the world to start, and now AIs are also labeling this data, and we'll get into those trade-offs.

And you have this kind of contrastive loss function between a good answer and a bad answer, and the model learns to pick up these trends. There's different implementation ways. You have things called reward models. You could have direct alignment algorithms. There's a lot of really specific things you can do, but all of this is about fine-tuning to human preferences.

And the final stage is much newer and will link to what is done in R1, and these reasoning models is, I think OpenAI is named for this. They had this new API in the fall, which they called the Reinforcement Fine-Tuning API. This is the idea that you use the techniques of reinforcement learning, which is a whole framework of AI.

There's a deep literature here. To summarize, it's often known as trial and error learning, or the subfield of AI where you're trying to make sequential decisions in a certain potentially noisy environment. There's a lot of ways we could go down that, but fine-tuning language models where they can generate an answer, and then you check to see if the answer matches the true solution.

For math or code, you have an exactly correct answer for math, you can have unit tests for code. And what we're doing is we are checking the language model's work, and we're giving it multiple opportunities on the same question to see if it is right. And if you keep doing this, the models can learn to improve in verifiable domains to a great extent.

It works really well. It's a newer technique in the academic literature. It's been used at frontier labs in the US that don't share every detail for multiple years. So this is the idea of using reinforcement learning with language models, and it has been taking off, especially in this DeepSeq moment.

- And we should say that there's a lot of exciting stuff going on, again, across the stack, but the post-training, probably this year, there's going to be a lot of interesting developments in the post-training, and we'll talk about it. I almost forgot to talk about the difference between DeepSeq v3 and R1 on the user experience side.

So forget the technical stuff, forget all of that. Just people that don't know anything about AI, they show up, like, what's the actual experience? What's the use case for each one when they actually, like, type and talk to it? What is each good at and that kind of thing?

- So let's start with DeepSeq v3 again. It's what more people would have tried something like it. You ask it a question. It'll start generating tokens very fast, and those tokens will look like a very human legible answer. It'll be some sort of markdown list. It might have formatting to help you draw to the core details in the answer, and it'll generate tens to hundreds of tokens.

A token is normally a word for common words or a sub-word part in a longer word. And it'll look like a very high-quality Reddit or Stack Overflow answer. These models are really getting good at doing these across a wide variety of domains. Even things that, if you're an expert, things that are close to the fringe of knowledge, they will still be fairly good at.

Cutting-edge AI topics that I do research on, these models are capable for study aid, and they're regularly updated. Where this changes is with the DeepSeq v1, what is called these reasoning models, is when you see tokens coming from these models to start, it will be a large chain-of-thought process.

We'll get back to chain-of-thought in a second, which looks like a lot of tokens where the model is explaining the problem. The model will often break down the problem, and be like, "Okay, they asked me for this. Let's break down the problem. I'm going to need to do this." And you'll see all of this generating from the model.

It'll come very fast in most user experiences. These APIs are very fast, so you'll see a lot of tokens. A lot of words show up really fast. It'll keep flowing on the screen, and this is all the reasoning process. And then eventually the model will change its tone in R1, and it'll write the answer, where it summarizes its reasoning process and writes a similar answer to the first types of model.

But in DeepSeek's case, which is part of why this was so popular, even outside the AI community, is that you can see how the language model is breaking down problems. And then you get this answer on a technical side. They train the model to do this specifically, where they have a section, which is reasoning, and then it generates a special token, which is probably hidden from the user most of the time, which says, "Okay, I'm starting the answer." So the model is trained to do this two-stage process on its own.

If you use a similar model in, say, OpenAI, OpenAI's user interface is trying to summarize this process for you nicely by kind of showing the sections that the model is doing, and it'll kind of click through. It'll say, "Breaking down the problem. "Making X calculation. "Cleaning the result." And then the answer will come for something like OpenAI.

- Maybe it's useful here to go through, like, an example of a DeepSeek R1 reasoning. - Yeah, so if you're looking at the screen here, what you'll see is a screenshot of the DeepSeek chat app, and at the top is thought for 157 seconds with the dropdown arrow. Underneath that, if we were in an app that we were running, the dropdown arrow would have the reasoning.

- So in this case, the specific question, which, you know, I'm philosophically/pothead-inclined, so this is asking DeepSeek R1 for one truly novel insight about humans, and it reveals the reasoning, and basically the truly novel aspect is what's pushing the reasoning to constantly sort of demodel, asking itself, "Is this truly novel?" So it's actually challenging itself to be more novel, more counterintuitive, less cringe, I suppose.

So some of the reasoning says, "This is just snapshots. "Alternatively, humans have a unique meta-emotion "where they feel emotions about their own emotions, "e.g. feeling guilty about being angry. "This recursive emotional layering "creates complex motivational drives "that don't exist in other animals. "The insight is that human emotions are nested." So it's reasoning through how humans feel emotions, it's reasoning about meta-emotions.

- Gonna have pages and pages of this. It's almost too much to actually read, but it's nice to skim as it's coming. - It's a James Joyce-like stream of consciousness, and then it goes, "Wait, the user wants something "that's not seen anywhere else. "Let me dig deeper." And, "Consider the human ability "to hold contradictory beliefs simultaneously.

"Cognitive dissonance is known, "but perhaps the function is to allow flexible adaptation." So on and so forth. I mean, that really captures the public imagination that, "Holy shit, this isn't..." I mean, intelligence/almost like an inkling of sentience, because you're thinking through, you're self-reflecting, you're deliberating. And the final result of that, after 157 seconds, is, "Humans instinctively convert selfish desires "into cooperative systems "by collectively pretending abstract rules, "money, laws, rights, are real.

"These shared hallucinations act as, quote, 'games,' "where competition is secretly redirected "to benefit the group, "turning conflict into society's fuel." Pretty profound. I mean, you know. - This is a potential digression, but a lot of people have found that these reasoning models can sometimes produce much more eloquent text.

That is a, at least, interesting example, I think, depending on how open-minded you are, you find language models interesting or not, and there's a spectrum there. - Well, I mean, it's some of the... We'll talk about different benchmarks and so on, but some is just a vibe. Like that, in itself, is a, let's say, quote, "fire tweet." - Yeah.

- If I'm trying to produce something where people are like, "Oh shit, okay." So that's a chain of thought we'll probably return to more. How were they able to achieve such low cost on the training and the inference? Maybe you could talk the training first. - Yeah, so there's two main techniques that they implemented that are probably the majority of their efficiency, and then there's a lot of implementation details that maybe we'll gloss over or get into later that sort of contribute to it.

But those two main things are, one is they went to a mixture of experts model, which we'll define in a second. And then the other thing is that they invented this new technique called MLA, latent attention. Both of these are big deals. Mixture of experts is something that's been in the literature for a handful of years.

And OpenAI with GPT-4 was the first one to productize a mixture of experts model. And what this means is, when you look at the common models around that most people have been able to interact with that are open, right? Think Lama. Lama is a dense model, i.e. every single parameter or neuron is activated as you're going through the model for every single token you generate, right?

Now, with a mixture of experts model, you don't do that, right? How does the human actually work, right? It's like, oh, well, my visual cortex is active when I'm thinking about, you know, vision tasks and like, you know, other things, right? My amygdala is when I'm scared, right? These different aspects of your brain are focused on different things.

A mixture of experts model attempts to approximate this to some extent. It's nowhere close to what a brain architecture is, but different portions of the model activate, right? You'll have a set number of experts in the model and a set number that are activated each time. And this dramatically reduces both your training and inference costs.

Because now you're, you know, if you think about the parameter count as the sort of total embedding space for all of this knowledge that you're compressing down during training, when you're embedding this data in, instead of having to activate every single parameter every single time you're training or running inference, now you can just activate a subset and the model will learn which expert to route to for different tasks.

And so this is a humongous innovation in terms of, hey, I can continue to grow the total embedding space of parameters. And so DeepSeek's model is, you know, 600 something billion parameters, right? Relative to LLAMA405B, it's 405 billion parameters, right? Relative to LLAMA70B, it's 70 billion parameters, right? So this model technically has more embedding space for information, right?

To compress all of the world's knowledge that's on the internet down, but at the same time, it is only activating around 37 billion of the parameters. So only 37 billion of these parameters actually need to be computed every single time you're training data or inferencing data out of it.

And so versus, again, the LLAMA model, 70 billion parameters must be activated or 405 billion parameters must be activated. So you've dramatically reduced your compute cost when you're doing training and inference with this mixture of experts architecture. - Should we break down where it actually applies and go into the transformer?

Is that useful? - Let's go, let's go into the transformer. - Okay, so the transformer is a thing that is talked about a lot and we will not cover every detail. Essentially, the transformer is built on repeated blocks of this attention mechanism and then a traditional dense, fully connected, multi-layer perception, whatever word you want to use for your normal neural network.

And you alternate these blocks. There's other details. And where mixture of experts is applied is at this dense model. The dense model holds most of the weights if you count them in a transformer model. So you can get really big gains from those mixture of experts on parameter efficiency at training and inference because you get this efficiency by not activating all of these parameters.

- We should also say that a transformer is a giant neural network. - Yeah. - And then there's for 15 years now, there's what's called the deep learning revolution. Network's gotten larger and larger. And at a certain point, the scaling laws appeared where people realized-- - This is a scaling law shirt, by the way.

- Representing scaling laws where it became more and more formalized that bigger is better across multiple dimensions of what bigger means. But these are all sort of neural networks we're talking about. And we're talking about different architectures of how to construct these neural networks such that the training and the inference on them is super efficient.

- Yeah, every different type of model has a different scaling law for it, which is effectively for how much compute you put in, the architecture will get to different levels of performance at test tasks. And mixture of experts is one of the ones at training time, even if you don't consider the inference benefits, which are also big.

At training time, your efficiency with your GPUs is dramatically improved by using this architecture if it is well implemented. So you can get effectively the same performance model and evaluation scores with numbers like 30% less compute. I think there's gonna be a wide variation depending on your implementation details and stuff.

But it is just important to realize that this type of technical innovation is something that gives huge gains. And I expect most companies that are serving their models to move to this mixture of experts implementation. Historically, the reason why not everyone might do it is because it's an implementation complexity, especially when doing these big models.

So this is one of the things that DeepSeq gets credit for is they do this extremely well. They do a mixture of experts extremely well. This architecture for what is called DeepSeq MOE, MOE is the shortened version of mixture of experts, is multiple papers old. This part of their training infrastructure is not new to these models alone.

And same goes for what Dylan mentioned with multi-head latent attention. This is all about reducing memory usage during inference and same things during training by using some fancy low rank approximation math. If you get into the details with this latent attention, it's one of those things I look at and say, okay, they're doing really complex implementations 'cause there's other parts of language models such as embeddings that are used to extend the context length.

The common one that DeepSeq uses rotary positional embeddings, which is called rope. And if you wanna use rope with a normal MOE, it's kind of a sequential thing. You take two of the attention matrices and you rotate them by a complex value rotation, which is a matrix multiplication. With DeepSeq's MLA, with this new attention architecture, they need to do some clever things because they're not set up the same and it just makes the implementation complexity much higher.

So they're managing all of these things. And these are probably the sort of things that OpenAI, these closed labs are doing. We don't know if they're doing the exact same techniques, but they actually shared them with the world, which is really nice to feel like this is the cutting edge of efficient language model training.

- And some of this requires low level engineering. Just it is a giant mess and trickery. So as I understand that one below CUDA, so they go super low programming of GPUs. - Effectively, NVIDIA builds this library called Nickel, right? In which, you know, when you're training a model, you have all these communications between every single layer of the model and you may have over a hundred layers.

- What does Nickel stand for? It's N-C-C-L? - NVIDIA Communications Collectives Library. - Nice. (laughs) - And so, - Damn. (laughs) - When you're training a model, right? You're going to have all these, all reduces and all gathers, right? Between each layer, between the multi-layer perceptron or feed forward network and the attention mechanism, you'll have basically the model synchronized, right?

Or you'll have all reduce or an all gather. And this is a communication between all the GPUs in the network, whether it's in training or inference. So NVIDIA has a standard library. This is one of the reasons why it's really difficult to use anyone else's hardware for training is because no one's really built a standard communications library.

And NVIDIA has done this at a sort of a higher level, right? A DeepSeek, because they have certain limitations around the GPUs that they have access to, the interconnects are limited to some extent by the restrictions of the GPUs that were shipped into China legally, not the ones that are smuggled, but legally shipped in, that they use to train this model.

They had to figure out how to get efficiencies, right? And one of those things is that instead of just calling the NVIDIA library, Nickel, right? They instead created their, they scheduled their own communications, which some of the labs do, right? Emeta talked about in Llama 3, how they made their own custom version of Nickel.

This is, they didn't talk about the implementation details. This is some of what they did, probably not as well as, maybe not as well as DeepSeek because DeepSeek, you know, necessity is the mother of innovation. And they had to do this, whereas in the case, you know, OpenAI has people that do this sort of stuff, Anthropic, et cetera.

But, you know, DeepSeek certainly did it publicly and they may have done it even better because they were gimped on a certain aspect of the chips that they have access to. And so they scheduled communications on, you know, by scheduling specific SMs. SMs you could think of as like the core on a GPU, right?

So there's hundreds of cores, or there's, you know, a bit over a hundred cores, SMs, on a GPU and they were specifically scheduling, hey, which ones are running the model? Which ones are doing all reduce? Which one are doing all gather, right? And they would flip back and forth between them.

And this requires extremely low level programming. - This is what Nickel does automatically or other NVIDIA libraries handle this automatically usually. - Yeah, exactly. And so technically they're using, you know, PTX, which is like sort of like, you could think of it as like an assembly type language. It's not exactly that or instruction set, right?

Like coding directly to assembly or instruction set. It's not exactly that, but that's still part of technically CUDA, but it's like, do I wanna write in Python, you know, PyTorch equivalent and call NVIDIA libraries? Do I wanna go down to the C level, right? Or, you know, encode even lower level, or do I wanna go all the way down to the assembly or ISO level?

And there are cases where you go all the way down there at the very big labs, but most companies just do not do that, right? Because it's a waste of time and the efficiency gains you get are not worth it. But DeepSeek's implementation is so complex, right? Especially with their mixture of experts, right?

People have done mixture of experts, but they're generally eight, 16 experts, right? And they activate too. So, you know, one of the words that we like to use is like sparsity factor, right? Or usage, right? So you might have four, you know, one fourth of your model activate, right?

And that's what Mistral's Mistral model, right? Their model that really catapulted them to like, oh my God, they're really, really good. OpenAI has also had models that are MOE and so have all the other labs that are major closed. But what DeepSeek did that maybe only the leading labs have only just started recently doing is have such a high sparsity factor, right?

It's not one fourth of the model, right? Two out of eight experts activating every time you go through the model, it's eight out of 256. - And there's different implementations for mixture of experts where you can have some of these experts that are always activated, which this just looks like a small neural network and then all the tokens go through that.

And then they also go through some that are selected by this routing mechanism. And one of the innovations in DeepSeek's architecture is that they changed the routing mechanism in mixture of expert models. There's something called an auxiliary loss, which effectively means during training, you want to make sure that all of these experts are used across the tasks that the model sees.

Why there can be failures in mixture of experts is that when you're doing this training, the one objective is token prediction accuracy. And if you just let training go with a mixture of expert model on your own, it can be that the model learns to only use a subset of the experts.

And in the MOE literature, there's something called the auxiliary loss, which helps balance them. But if you think about the loss functions of deep learning, this even connects to the bitter lesson is that you want to have the minimum inductive bias in your model to let the model learn maximally.

And this auxiliary loss, this balancing across experts could be seen as intention with the prediction accuracy of the tokens. So we don't know the exact extent that the DeepSeek MOE change, which is instead of doing an auxiliary loss, they have an extra parameter in their routing, which after the batches, they update this parameter to make sure that the next batches all have a similar use of experts.

And this type of change can be big, it can be small, but they add up over time. And this is the sort of thing that just points to them innovating. And I'm sure all the labs that are training big MOEs are looking at this sort of things, which is getting away from the auxiliary loss.

Some of them might already use it, but you keep accumulating gains. And we'll talk about the philosophy of training and how you organize these organizations. And a lot of it is just compounding small improvements over time in your data, in your architecture, in your post-training and how they integrate with each other.

DeepSeq does the same thing. And some of them are shared a lot. We have to take them on face value that they share their most important details. I mean, the architecture and the weights are out there. So we're seeing what they're doing. And it adds up. - Going back to sort of the like efficiency and complexity point, right?

It's 32 versus four, right? For like Mixedraw and other MOE models that have been publicly released. So this ratio is extremely high. And sort of what Nathan was getting at there was when you have such a different level of sparsity, you can't just have every GPU have the entire model, right?

The model's too big. There's too much complexity there. So you have to split up the model with different types of parallelism, right? And so you might have different experts on different GPU nodes. But now what happens when this set of data that you get, hey, all of it looks like this one way and all of it should route to one part of my model, right?

So when all of it routes to one part of the model, then you can have this overloading of a certain set of the GPU resources or a certain set of the GPUs. And then the rest of the training network sits idle because all of the tokens are just routing to that.

So this is the biggest complexity. One of the big complexities with running a very sparse mixture of experts model, i.e. this 32 ratio versus this four ratio is that you end up with so many of the experts just sitting there idle. So how do I load balance between them?

How do I schedule the communications between them? This is a lot of the extremely low-level detailed work that they figured out in the public first and potentially second or third in the world, and maybe even first in some cases. - What lesson do you, in the direction of the better lesson, do you take from all of this?

Is this going to be the direction where a lot of the gain is going to be, which is this kind of low-level optimization? Or is this a short-term thing where the biggest gains will be more on the algorithmic high-level side of post-training? Is this a short-term leap because they've figured out a hack because constraints, necessities, the mother of invention?

Or is there still a lot of gain? - I think we should summarize what the better lesson actually is about. Is that the better lesson, essentially, if you paraphrase it, is that the types of training that will win out in deep learning as we go are those methods which are scalable in learning and search, is what it calls out.

And the scale word gets a lot of attention in this. The interpretation that I use is effectively to avoid adding human priors to your learning process. And if you read the original essay, this is what it talks about, is how researchers will try to come up with clever solutions to their specific problem that might get them small gains in the short-term while simply enabling these deep learning systems to work efficiently and for these bigger problems in the long-term might be more likely to scale and continue to drive success.

And therefore, we were talking about relatively small implementation changes to the mixture of experts model. And therefore, it's like, okay, we will need a few more years to know if one of these are actually really crucial to the better lesson. But the better lesson is really this long-term arc of how simplicity can often win.

And there's a lot of sayings in the industry, like the models just wanna learn. You have to give them the simple loss landscape where you put compute through the model and they will learn and getting barriers out of the way. - That's where the power of something like Nickel comes in where standardized code that can be used by a lot of people to create sort of simple innovations that can scale, which is why the hacks, I imagine that the code base for DeepSeq is probably a giant mess.

- I'm sure they have, DeepSeq definitely has code bases that are extremely messy where they're testing these new ideas. Multi-head latent attention probably start, could start in something like a Jupyter notebook or somebody tries something on a few GPUs and that is really messy. But the stuff that trains the DeepSeq v3 and DeepSeq r1, those libraries, if you were to present them to us, I would guess are extremely high quality code.

- High quality readable code, yeah. - I think there is one aspect to note though, right? Is that there is the general ability for that to transfer across different types of runs, right? You may make really, really high quality code for one specific model architecture at one size. And then that is not transferable to, hey, when I make this architecture tweak, everything's broken again, right?

Like that's something that could be, with their specific low level coding of like scheduling SMs is specific to this model architecture and size, right? And whereas like NVIDIA's collectives library is more like, hey, it'll work for anything, right? You wanna do an all-reduce, great. I don't care what your model architecture is, it'll work.

And you're giving up a lot of performance when you do that in many cases, but it's worthwhile for them to do the specific optimization for the specific run given the constraints that they have regarding compute. - I wonder how stressful it is to like, you know, these frontier models, like initiate training, like to have the code to push the button that like you're now spending a large amount of money and time to train this.

Like there must, I mean, there must be a lot of innovation on the debugging stage of like making sure there's no issues that you're monitoring and visualizing every aspect of the training, all that kind of stuff. - When people are training, they have all these various dashboards, but like the most simple one is your loss, right?

And it continues to go down, but in reality, especially with more complicated stuff like MOE, the biggest problem with it, or FP8 training, which is another innovation, you know, going to a lower precision number format, i.e. less accurate is that you end up with loss spikes, right? And no one knows why the loss spike happened.

And for a long- - Some of them you do. - Some of them you do. - Some of them are bad data. Can I give a AI2's example of what blew up our earlier models is a subreddit called Microwave Gang. We love to shout this out. It's a real thing.

You can pull up Microwave Gang. Essentially, it's a subreddit where everybody makes posts that are just the letter M. So it's like, "Mm." So there's extremely long sequences of the letter M. And then the comments are like, "Beep, beep." 'Cause it's in the microwave vents. But if you pass this into a model that's trained to be a normal producing text, it's extremely high loss.

'Cause normally you see an M, you don't predict Ms for a long time. So like this is something that caused the loss spikes for us. But when you have much like, this is old. This is not recent. And when you have more mature data systems, that's not the thing that causes the loss spike.

And what Dylan is saying is true. But it's like, it's levels to this sort of idea. - With regards to the stress, right? These people are like, you know, you'll go out to dinner with like a friend that works at one of these labs. And they'll just be like looking at their phone every like 10 minutes.

And they're not like, you know, it's one thing if they're texting, but they're just like, is the loss, is the loss provoking? - Like tokens per second, loss not blown up. They're just watching this. - And the heart rate goes up if there's a spike. - And some level of spikes is normal, right?

It'll recover and be back. Sometimes a lot of the old strategy was like, you just stop the run, restart from the old version, and then like change the data mix. And then it keeps going. - There are even different types of spikes. So Dirk Grunenfeld has a theory that I do, that's like fast spikes and slow spikes.

Where there are sometimes where you're looking at the loss and there are other parameters, you can see it start to creep up and then blow up. And that's really hard to recover from. So you have to go back much further. So you have the stressful period where it's like flat or it might start going up.

And you're like, what do I do? Whereas there are also loss spikes that are, it looks good. And then there's one spiky data point. And what you can do is you just skip those. You see that there's a spike. You're like, okay, I can ignore this data. Don't update the model and do the next one and it'll recover quickly.

But these like on trickier implementations, as you get more complex in your architecture and you scale up to more GPUs, you have more potential for your loss blowing up. So it's like, there's a distribution. - The whole idea of grokking also comes in, right? It's like, just because it slowed down from improving and loss doesn't mean it's not learning because all of a sudden it could be like this and it could just spike down and loss again because it learned, truly learned something, right?

And it took some time for it to learn that. It's not like a gradual process, right? And that's what humans are like. That's what models are like. So it's really a stressful task, as you mentioned. - And the whole time the dollar count is going up. Every company has failed runs.

You need failed run to push the envelope on your infrastructure. So a lot of news cycles are made of X company had Y failed run. Every company that's trying to push the frontier of AI has these. So it is, yes, it's noteworthy because it's a lot of money and it can be week to month setback, but it is part of the process.

- But how do you get, if you're deep seek, how do you get to a place where, holy shit, there's a successful combination of hyper parameters? - A lot of small failed runs. - And so rapid iteration through failed runs until- - And successful ones. - And then you build a summary tuition like this, this mixture of expert works, and then this implementation of MLA works.

- Key hyper parameters like learning rate and regularization and things like this. And you find the regime that works for your code base. I've, talking to people at Frontier Labs, there's a story that you can tell where training language models is kind of a path that you need to follow.

So you need to like unlock the ability to train a certain type of model or a certain scale. And then your code base and your internal know-how of which hyper parameters work for it is kind of known. And you look at the deep seek papers and models, they've scaled up, they've added complexity.

And it's just continuing to build the capabilities that they have. - There's the concept of a YOLO run. So YOLO, you only live once. And what it is, is like there's all this experimentation you do at the small scale, right? Research ablations, right? Like you have your Jupyter Notebook where you're experimenting with MLA on like three GPUs or whatever.

And you're doing all these different things like, hey, do I do four active experts, 128 experts? Do I arrange the experts this way? You know, all these different model architecture things you're testing at a very small scale, right? Couple of researchers, few GPUs, tens of GPUs, hundreds of GPUs, whatever it is.

And then all of a sudden you're like, okay guys, no more fucking around, right? No more screwing around. Everyone take all the resources we have, let's pick what we think will work and just go for it, right? YOLO. And this is where that sort of stress comes in is like, well, I know it works here, but some things that work here don't work here.

And some things that work here don't work down here, right? In this terms of scale, right? So it's really truly a YOLO run. And sort of like, there's this like discussion of like certain researchers just have like this methodical nature, like they can find the whole search space and like figure out all the ablations of different research and really see what is best.

And there's certain researchers who just kind of like, you know, have that innate gut instinct of like, this is the YOLO run. Like, you know, looking at the data, this is it. - This is why you want to work in post-training because the GPU costs for training is lower.

So you can make a higher percentage of your training runs YOLO runs. - Yeah. - For now. - Yeah, for now, for now. So some of this is fundamentally luck still. Luck is skill, right? In many cases. - Yeah, I mean, it looks lucky, right? When you're-- - But the hill to climb, if you're on one of these labs, you have an evaluation you're not crushing.

There's a repeated playbook of how you improve things. There are localized improvements, which might be data improvements. And these add up into the whole model just being much better. And when you zoom in really close, it can be really obvious that this model is just really bad at this thing and we can fix it.

And you just add these up. So some of it feels like luck, but on the ground, especially with these new reasoning models we're talking to, it's just so many ways that we can poke around. And normally it's that some of them give big improvements. - The search space is near infinite, right?

And yet the amount of compute and time you have is very low and you have to hit release schedules. You have to not get blown past by everyone. Otherwise, what happened with DeepSeek, crushing Meta and Mistral and Cohere and all these guys, they moved too slow, right? They maybe were too methodical.

I don't know, they didn't hit the YOLO run. Whatever the reason was, maybe they weren't as skilled. Whatever, you know, you can call it luck if you want, but at the end of the day, it's skill. - So 2025 is the year of the YOLO run. It seems like all the labs are like going in.

- I think it's even more impressive what OpenAI did in 2022, right? At the time, no one believed in mixture of experts models, right, at Google, who had all the researchers. OpenAI had such little compute and they devoted all of their compute for many months, right, all of it, 100% for many months to GPT-4 with a brand new architecture with no belief that, hey, let me spend a couple hundred million dollars, which is all of the money I have on this model, right?

That is truly YOLO, right? Now, you know, people are like, all these like training run failures that are in the media, right? It's like, okay, great, but like actually, a huge chunk of my GPs are doing inference. I still have a bunch doing research constantly. And yes, my biggest cluster is training, but like on this YOLO run, but like that YOLO run is much less risky than like what OpenAI did in 2022, or maybe what DeepSeek did now, or, you know, like sort of like, hey, we're just gonna throw everything at it.

- The big winners throughout human history are the ones who are willing to do YOLO at some point. Okay, what do we understand about the hardware it's been trained on, DeepSeek? - DeepSeek is very interesting, right? This is where, second to take us to zoom out out of who they are, first of all, right?

HiFlyer is a hedge fund that has historically done quantitative trading in China as well as elsewhere. And they have always had a significant number of GPUs, right? In the past, a lot of these high-frequency trading, algorithmic quant traders used FPGAs, but it shifted to GPUs definitely. And there's both, right?

But GPUs especially, and HiFlyer, which is the hedge fund that owns DeepSeek, and everyone who works for DeepSeek is part of HiFlyer to some extent, right? It's same parent company, same owner, same CEO. They had all these resources and infrastructure for trading. And then they devoted a humongous portion of them to training models, both language models and otherwise, right?

Because these techniques were heavily AI-influenced. More recently, people have realized, hey, trading with... Even when you go back to like Renaissance and all these quantitative firms, natural language processing is the key to trading really fast, right? Understanding a press release and making the right trade, right? And so DeepSeek has always been really good at this.

And even as far back as 2021, they have press releases and papers saying like, hey, we're the first company in China with an A100 cluster this large. There's 10,000 A100 GPUs, right? This is in 2021. Now, this wasn't all for training large language models. This was mostly for training models for their quantitative aspects, their quantitative trading, as well as a lot of that was natural language processing to be clear, right?

And so this is the sort of history, right? So verifiable fact is that in 2021, they built the largest Chinese cluster. At least they claim it was the largest cluster in China, 10,000 GPUs. - Before expert controls started. - Yeah. - It's like they've had a huge cluster before any conversation of expert controls.

- So then you step it forward to like, what have they done over the last four years since then? Obviously they've continued to operate the hedge fund, probably make tons of money. And the other thing is that they've leaned more and more and more into AI. The CEO, Lian Cheng Feng, Lian.

- You're not putting me in the spot unless we discuss this more. - Lian Feng, right? The CEO, he owns maybe a little bit more than half the company allegedly, right? Is an extremely like Elon Jensen kind of figure where he's just like involved in everything, right? And so over that time period, he's gotten really in-depth into AI.

He actually has a bit of a like, if you see some of his statements, a bit of an EAC vibe almost, right? Total AGI vibes. They're like, we need to do this. We need to make a new ecosystem of open AI. We need China to lead on this sort of ecosystem because historically the Western countries have led on software ecosystems and straight up acknowledges like in order to do this, we need to do something different.

DeepSeek is his way of doing this. Some of the translated interviews with him are fantastic. - So he has done interviews? - Yeah. - Do you think he would do a Western interview or no? Or is there controls on the channel? - There hasn't been one yet, but I would try it.

- I just got a Chinese translator. So it was great. This is all push. So fascinating figure, engineer, pushing full on into AI, leveraging the success from the high-frequency trading. - Very direct quotes. Like we will not switch to closed source when asked about this stuff. Very long-term motivated in how the ecosystem of AI should work and I think from a Chinese perspective, he wants a Chinese company to build this vision.

- And so this is sort of like the quote unquote visionary behind the company, right? This hedge fund still exists, right? This quantitative firm. And so DeepSeek is the sort of, slowly he got turned to this full view of like AI, everything about this, right? But at some point it slowly maneuvered and he made DeepSeek.

And DeepSeek has done multiple models since then. They've acquired more and more GPUs. They share infrastructure with the fund, right? And so there is no exact number of public GPU resources that they have, but besides this 10,000 GPUs that they bought in 2021, right? And they were fantastically profitable, right?

And then this paper claims they did only 2,000 H800 GPUs, which are a restricted GPU that was previously allowed in China, but no longer allowed. And there's a new version, but it's basically NVIDIA's H100 for China, right? And that there's some restrictions on it specifically around the communications, sort of a speed, the interconnect speed, right?

Which is why they had to do this crazy SM, you know, scheduling stuff, right? So going back to that, right? It looks like this is obviously not true in terms of their total GPU count. - Obvious available GPUs, but for this training run, you think 2000 is the correct number or no?

- So this is where it takes, you know, a significant amount of sort of like zoning in, right? Like what do you call your training run, right? You count all of the research and ablations that you ran, right? Picking all this stuff, because yes, you can do a YOLO run, but at some level you have to do the test at the small scale and then you have to do some tests at medium scale before you go to a large scale.

Accepted practice is that for any given model that is a notable advancement, you're gonna do two to four X compute of the full training run in experiments alone. - So a lot of this compute that's being scaled up is probably used in large part at this time for research.

- Yeah, and research will, you know, research begets the new ideas that let you get huge efficiency. - Research gets you O1. Like research gets you breakthroughs and you need to bet on it. - So some of the pricing strategy they will discuss has the research baked into the price.

- So the numbers that DeepSeek specifically said publicly, right, are just the 10,000 GPUs in 2021, and then 2000 GPUs for only the pre-training for V3. They did not discuss cost on R1. They did not discuss cost on all the other RL, right, for the instruct model that they made, right?

They only discussed the pre-training for the base model, and they did not discuss anything on research and ablations. And they do not talk about any of the resources that are shared in terms of, "Hey, the fund is using all these GPUs," right? And we know that they're very profitable and that 10,000 GPUs in 2021.

So some of the research that we've found is that we actually believe they have closer to 50,000 GPUs. - We as semi-analysts, so we should say that you're sort of one of the world experts in figuring out what everybody's doing in terms of the semiconductor, in terms of cluster build-outs, in terms of who's doing what in terms of training runs.

So yeah, so that's the we. Okay, go ahead. - Yeah, sorry. We believe they actually have something closer to 50,000 GPUs, right? Now, this is split across many tasks, right? Again, the fund, research and ablations. - For ballpark, how much would open AI or Anthropic have? I think the clearest example we have, because Meta is also open, they talk about order of 60K to 100K H100 equivalent GPUs in their training clusters.

- Right, so like Lama3, they trained on 16,000 H100s, right? But the company of Meta last year publicly disclosed they bought like 400 something thousand GPUs, right? So of course, tiny percentage on the training, again, most of it is like serving me the best Instagram reels, right? Or whatever, right?

- I mean, we could get into a cost of like, what is the cost of ownership for a 2000 GPU cluster, 10,000? - There's just different sizes of companies that can afford these things. And DeepSeek is reasonably big. Their compute allocation compared is one of the top few in the world.

It's not open AI, Anthropic, et cetera, but they have a lot of compute. - Can you in general actually just zoom out and also talk about the Hopper architecture, the NVIDIA Hopper GPU architecture and the difference between H100 and H800, like you mentioned, the interconnects. - Yeah, so there's, you know, Ampere was the A100 and then H100 Hopper, right?

People use them synonymously in the US because really there's just H100 and now there's H200, right? But same thing, mostly. In China, there've been different salvos of export restrictions. So initially the US government limited on a two-factor scale, right? Which is chip interconnect versus flops, right? So any chip that had interconnects above a certain level and flops above a certain floating point operations above a certain level was restricted.

Later, the government realized that this was a flaw in the restriction and they cut it down to just floating point operations. And so- - H800 had high flops, low communication? - Exactly, so the H800 was the same performance as H100 on flops, right? But it just had the interconnect bandwidth cut.

DeepSeek knew how to utilize this, you know, hey, even though we're cut back on the interconnect, we can do all this fancy stuff to figure out how to use the GPU fully anyways, right? And so that was back in October, 2022, but later in 2023, end of 2023, implemented in 2024, the US government banned the H800, right?

And so by the way, this H800 cluster, these 2000 GPUs was not even purchased in 2024, right? It was purchased in late 2023. And they're just getting the model out now, right? Because it takes a lot of research, et cetera. H800 was banned and now there's a new chip called the H20.

The H20 is cut back on only flops, but the interconnect bandwidth is the same. And in fact, in some ways it's better than the H100 because it has better memory bandwidth and memory capacity. So there are, you know, NVIDIA is working within the constraints of what the government says and then builds the best possible GPU for China.

- Can we take this extra tangent and we'll return back to the hardware? Is the philosophy, the motivation, the case for export controls, what is it? Dari Amadej has published a blog post about export controls. The case he makes is that if AI becomes super powerful and he says by 2026, we'll have AGI or super powerful AI, and that's going to give a significant, whoever builds that will have a significant military advantage.

And so, because the United States is a democracy and as he says, China is authoritarian or has authoritarian elements, you want a unipolar world where the super powerful military because of the AI is one that's a democracy. It's a much more complicated world geopolitically when you have two superpowers with super powerful AI and one is authoritarian.

So that's the case he makes. And so we wanna, the United States wants to use export controls to slow down, to make sure that China can't do these gigantic training runs that would be presumably required to build AGI. - This is very abstract. I think this can be the goal of how some people describe export controls is the super powerful AI.

There's, and you touched on the training run idea. There's not many worlds where China cannot train AI models. I think export controls are kneecapping the amount of compute or the density of compute that China can have. And if you think about the AI ecosystem right now, as all of these AI companies, revenue numbers are up into the right.

AI usage is just continuing to grow. More GPUs are going to inference. A large part of export controls, if they work is just that the amount of AI that can be run in China is going to be much lower. So on the training side, DeepSeek v3 is a great example, which you have a very focused team that can still get to the frontier of AI.

This 2000 GPUs is not that hard to get all considering in the world. They're still gonna have those GPUs. They're still gonna be able to train models. But if there's gonna be a huge market for AI, if you have strong export controls and you want to have 100,000 GPUs just serving the equivalent of Chad GPT clusters, with good export controls, it also just makes it so that AI can be used much less.

And I think that is a much easier goal to achieve than trying to debate on what AGI is. And if you have these extremely intelligent, autonomous AIs and data centers, like those are the things that could be running in these GPU clusters in the United States, but not in China.

- To some extent, training a model does effectively nothing, right? Like you have a model. The thing that Dario is sort of speaking to is the implementation of that model, once trained to then create huge economic growth, huge increases in military capabilities, huge increases in productivity of people, betterment of lives, whatever you wanna direct super powerful AI towards, you can't.

But that requires a significant amounts of compute, right? And so the US government has effectively said, - And forever, right? Like training will always be a portion of the total compute. We mentioned Meta's 400,000 GPUs, only 16,000 made Llama, right? So the percentage that Meta is dedicating to inference, now this might be for recommendation systems that are trying to hack our mind into spending more time and watching more ads, or if it's for a super powerful AI that's doing productive things, doesn't matter about the exact use that our economic system decides, it's that that can be delivered in whatever way we want.

Whereas with China, expert restrictions, great, you're never gonna be able to cut everything off, right? And I think that's quite well understood by the US government, is that you can't cut everything off. - They'll make their own chips. - And they're trying to make their own chips, they'll be worse than ours, but the whole point is to just keep a gap, right?

And therefore at some point as the AI, in a world where two, 3% economic growth, this is really dumb by the way, right? To cut off high tech and not make money off of it, but in a world where super powerful AI comes about and then starts creating significant changes in society, which is what all the AI leaders and big tech companies believe, I think super powerful AI is gonna change society massively.

And therefore this compounding effect of the difference in compute is really important. There's some sci-fi out there where like AI is like measured in the power of, in like how much power is delivered to compute, right? Or how much is being, that's sort of a way of thinking about what's the economic output is just how much power you directing towards that AI.

- Should we talk about reasoning models with this as a way that this might be actionable as something that people can actually see? So the reasoning models that are coming out with R1 and O1, they're designed to use more compute. There's a lot of buzzy words in the AI community about this test time compute, inference time compute, whatever, but Dylan has good research on this.

You can get to those specific numbers on the ratio of when you train a model, you can look at things about the amount of compute used at training and amount of compute used at inference. These reasoning models are making inference way more important to doing complex tasks. In the fall, in December, their OpenAI announced this O3 model.

There's another thing in AI when things move fast, we get both announcements and releases. Announcements are essentially blog posts where you pat yourself on the back and you say you did things and releases are on the models out there, the papers out there, et cetera. So OpenAI has announced O3 and we can check if O3 mini is out as a recording potentially, but that doesn't really change the point, which is that the breakthrough result was something called ArcAGI task, which is the abstract reasoning corpus, a task for artificial general intelligence.

Francois Chollet is the guy who's been, it's a multi-year old paper, it's a brilliant benchmark. And the number for OpenAI O3 to solve this was that it used some sort of number of samples in the API. The API has like thinking effort and number of samples. They used a thousand samples to solve this task and it comes out to be like five to $20 per question, which you're putting in as effectively a math puzzle.

And then it takes orders of dollars to answer one question. And this is a lot of compute. If those are gonna take off in the US, OpenAI needs a ton of GPUs on inference to capture this. They have this OpenAI chat GPT pro subscription, which is $200 a month.

- Which Sam said they're losing money on. - Which means that people are burning a lot of GPUs on inference. And I've signed up with it, I've played with it. I don't think I'm a power user, but I use it. And it's like, that is the thing that a Chinese company with mediumly strong expert controls, there will always be loopholes, might not be able to do it all.

And if that, the main result for O3 is also a spectacular coding performance. And if that feeds back into AI companies being able to experiment better. - So presumably the idea is for an AGI, a much larger fraction of the compute would be used for this test time compute for the reasoning.

For the AGI goes into a room and thinks about how to take over the world and come back in 2.7 hours and that it's gonna take a lot of compute. - This is what people like CEO or leaders of OpenAI and Anthropic talk about is like autonomous AI models, which is you give them a task and they work on it in the background.

I think my personal definition of AGI is much simpler. Like I think language models are a form of AGI and all of this super powerful stuff is a next step. That's great if we get these tools, but a language model has so much value in so many domains. It is a general intelligence to me.

But this next step of agentic things where they're independent and they can do tasks that aren't in the training data is what the few year outlook that these AI companies are driving for. - I think the terminology here that Dario uses is super powerful AI. So I agree with you on the AGI.

I think we already have something like that's exceptionally impressive that Alan Turing would for sure say is AGI. But he's referring more to something once in possession of then you would have a significant military and geopolitical advantage over other nations. So it's not just like you can ask it how to cook an omelet.

- And he has a much more positive view in his essay, "Machines of Love and Grace." I read into this that we don't have enough background in physical sciences to gauge exactly how competent I am and if AI can revolutionize biology. I'm safe saying that AI is going to accelerate the progress of any computational science.

- So we're doing a depth first search here on topics, taking tangent of a tangent, so let's continue on that depth first search. You said that you're both feeling the AGI. So what's your timeline? Dario's 2026 for the super powerful AI that's basically agentic to a degree where it's a real security threat, that level of AGI.

What's your timeline? - I don't like to attribute specific abilities because predicting specific abilities and when is very hard. I think mostly if you're going to say that I'm feeling the AGI is that I expect continued, rapid, surprising progress over the next few years. So something like R1 is less surprising to me from DeepSeek because I expect there to be new paradigms where substantial progress can be made.

I think DeepSeek R1 is so unsettling because we're kind of on this path with chat GPT. It's like, it's getting better, it's getting better, it's getting better. And then we have a new direction for changing the models and we took one step like this and we like took a step up.

So it looks like a really fast slope and then we're going to just take more steps. So like, it's just really unsettling when you have these big steps and I expect that to keep happening. I see, I've tried OpenAI Operator, I've tried Cloud computer use. They're not there yet.

I understand the idea, but it's just so hard to predict what is the breakthrough that'll make something like that work. And I think it's more likely that we have breakthroughs that work and things that we don't know what they're going to do. So everyone wants agents. Dario has very eloquent way of describing this.

And I just think that it's like, there's going to be more than that. So like, just expect these things to come. - I'm going to have to try to pin you down to a date on the AGI timeline. Like the nuclear weapon moment. So moment where on the geopolitical stage, there's a real like, you know, 'cause we're talking about export controls.

When do you think, just even to throw out a date, when do you think that would be? Like for me, it's probably after 2030. So I'm not as- - That's what I would say. - So define that, right? Because to me, it kind of almost has already happened, right?

You look at elections in India and Pakistan, people get AI voice calls and think they're talking to the politician, right? The AI diffusion rules, which was enacted in the last couple of weeks of the Biden admin, and it looks like the Trump admin will keep and potentially even strengthen, limit cloud computing and GPU sales to countries that are not even related to China.

It's like, this is- - Portugal and all these like normal countries are on the, you need approval from the US list. - Like, yeah, Portugal and like, you know, like all these countries that are allies, right? Singapore, right? Like they freaking have F-35s and we don't let them buy GPUs.

Like this is, this to me is already to the scale of like, you know. - Well, that just means that the US military is really nervous about this new technology. That doesn't mean the technology is already there. So like, they might be just very cautious about this thing that they don't quite understand.

That's a really good point. Sort of the robocalls, swarms of semi-intelligent bots could be a weapon, could be doing a lot of social engineering. - I mean, there's tons of talk about, you know, from the 2016 elections, like Cambridge Analytica and all this stuff, Russian influence. I mean, every country in the world is pushing stuff onto the internet and has narratives they want, right?

Like that's, every like technically competent, whether it's Russia, China, US, Israel, et cetera, right? You know, people are pushing viewpoints onto the internet en masse. And language models crash the cost of like very intelligent sounding language. - There's some research that shows that the distribution is actually the limiting factor.

So language models haven't yet made misinformation particularly, like change the equation there. The internet is still ongoing. I think there's a blog, AI Snake Oil and some of my friends at Princeton that write on this stuff. So there is research. It's like, it's a default that everyone assumes. And I would have thought the same thing is that misinformation is gonna get far worse with language models.

I think in terms of internet posts and things that people have been measuring, it hasn't been a exponential increase or something extremely measurable. And things you're talking about with like voice calls and stuff like that, it could be in modalities that are harder to measure. So it's something that it's too soon to tell in terms of, I think that's like political instability via the web is very, it's monitored by a lot of researchers to see what's happening.

I think that you're asking about like the AGI thing. I might, if you make me give a year, I wouldn't be like, okay, I have AI CEOs saying this. They've been saying two years for a while. I think that they're people like Dario Anthropic, the CEO had thought about this so deeply.

I need to take their word seriously, but also understand that they have different incentives. So I would be like, add a few years to that, which is how you get something similar to 2030 or a little after 2030. - I think to some extent we have capabilities that hit a certain point where any one person could say, oh, okay, if I can leverage those capabilities for X amount of time, this is AGI, right?

Call it 27, 28. But then the cost of actually operating that capability. - Yeah, this is gonna be my point. - So, so extreme that no one can actually deploy it at scale and mass to actually completely revolutionize the economy on a snap of a finger. So I don't think it will be like a snap of the finger moment.

- It's a physical constraint. - Rather it'll be a, oh, the capabilities are here, but I can't deploy it everywhere, right? And so one simple example going back sort of to 2023 was when being with GPT-4 came out and everyone was freaking out about search, right? Perplexity came out.

If you did the cost on like, hey, implementing GPT-3 into every Google search, it was like, oh, okay, this is just like physically impossible to implement, right? And as we step forward to like going back to the test time compute thing, right? A query for, you know, you asked ChatGPT a question, it costs cents, right?

For their most capable model of chat, right? To get a query back, to solve an Arc AGI problem though, cost five to 20 bucks, right? And this is an-- - It's only going up from there. - This is a thousand, 10,000 X factor difference in cost to respond to a query versus do a task.

And the task of Arc AGI, it's not like it's like, it's simple to some extent, you know, but it's also like, what are the tasks that we want? Okay, AGI, quote unquote, what we have today can do Arc AGI. Three years from now, it can do much more complicated problems, but the cost is gonna be measured in thousands and thousands and hundreds of thousands of dollars of GPU time.

And there just won't be enough power, GPUs, infrastructure to operate this and therefore shift everything in the world on the snap of the finger. But at that moment, who gets to control and point the AGI at a task? And so this was in Dario's post that he's like, hey, China can effectively and more quickly than us point their AGI at military tasks, right?

And they have been in many ways faster at adopting certain new technologies into their military, right? Especially with regards to drones, right? The U.S. maybe has a longstanding, you know, large air sort of, you know, fighter jet type of thing, bombers, but when it comes to asymmetric arms, such as drones, they've completely leapfrogged the U.S.

and the West. And the fear that Dario is sort of pointing out there, I think, is that, yeah, great, we'll have AGI in the commercial sector. The U.S. military won't be able to implement it super fast. Chinese military could, and they could direct all their resources to implementing it in the military and therefore solving, you know, military logistics or solving some other aspect of like disinformation for targeted certain set of people so that they can flip a country's politics or something like that that is actually like catastrophic versus, you know, the U.S.

just wants to, you know, 'cause it'll be more capitalistically allocated just towards whatever is the highest return on income, which might be like building, you know, factories better or whatever. - So everything I've seen, people's intuition seems to fail on robotics. So you have this kind of general optimism.

I've seen this on self-driving cars. People think it's a much easier problem than it is. Similar with drones. Here, I understand it a little bit less, but I've just seen the reality of the war in Ukraine and the usage of drones on both sides. And it seems that humans still far outperform any fully autonomous systems.

AI is an assistant, but humans drive. FPV drones, where the human's controlling most of it, just far, far, far outperforms AI systems. So I think it's not obvious to me that we're going to have swarms of autonomous robots anytime soon in the military context. Maybe the fastest I can imagine is 2030, which is why I said 2030 for the super powerful AI.

Whenever you have large-scale swarms of robots doing military actions, that's when the world just starts to look different to me. So that's the thing I'm really worried about. But there could be cyber war, cyber war type of technologies that, from social engineering to actually just swarms of robots that find attack vectors in our code bases and shut down power grids, that kind of stuff.

And it could be one of those things like on any given weekend or something, power goes out, nobody knows why, and the world changes forever. Just power going out for two days in all of the United States, that will lead to murder, to chaos. But going back to expert controls, do you see that as a useful way to control the balance of power geopolitically in the context of AI?

- And I think going back to my viewpoint is, if you believe we're in the sort of stage of economic growth and change that we've been in for the last 20 years, the export controls are absolutely guaranteeing that China will win long-term, right? If you do not believe AI is going to make significant changes to society in the next 10 years or five years, right?

Five-year timelines are sort of what the more executives and such of AI companies and even big tech companies believe, but even 10-year timelines, it's reasonable. But once you get to, hey, these timelines are below that time period, then the only way to sort of create a sizable advantage or disadvantage for America versus China is if you constrain compute because talent is not really something that's constraining, right?

China arguably has more talent, right? More STEM graduates, more programmers. The US can draw upon the world's people, which it does. There's tons of foreigners in the AI industry. - So many of these AI teams are all people without a US passport. - Yeah, I mean, many of them are Chinese people who are moving to America, right?

And that's great. That's exactly what we want, right? But that talent is one aspect, but I don't think that's one that is a measurable advantage for the US or not. It truly is just whether or not compute, right? Now, even on the compute side, when we look at chips versus data centers, right?

China has the unprecedented ability to build ridiculous sums of power clockwork, right? They're always building more and more power. They've got steel mills that like individually are the size of the entire US industry, right? And they've got aluminum mills that consume gigawatts and gigawatts of power, right? And when we talk about what's the biggest data center, right?

OpenAI made this huge thing about Stargate, their announcement there. That's like once it's fully built out in a few years, it'll be two gigawatts, right? Of power, right? And this is still smaller than the largest, you know, industrial facilities in China, right? China, if they wanted to build the largest data center in the world, if they had access to the chips, could.

So it's just a question of when, not if, right? - So their industrial capacity far exceeds the United States? - Exactly. - To manufacture stuff. - Yeah. - So long-term, they're going to be manufacturing chips there. - Chips are a little bit more specialized. I'm specifically referring to the data centers, right?

Chips, fabs take huge amounts of power. Don't get me wrong. That's not necessarily the gating factor there. The gating factor on how fast people can build the largest clusters today in the U.S. is power, right? It is, whether it's, now it could be power generation, power transmission, substations, and, you know, all these sorts of transformers and all these things, building the data center.

These are all constraints on the U.S. industry's ability to build larger and larger training systems, as well as deploying more and more inference compute. - I think we need to make the point clear on why the time is now for people that don't think about this, 'cause essentially with export controls, you're making it so China cannot make or get cutting edge chips.

And the idea is that if you time this wrong, China is pouring a ton of money into their chip production. And if you time it wrong, they are going to have more capacity for production, more capacity for energy, and figure out how to make the chips, and have more capacity than the rest of the world to make the chips, because everybody can buy, they're going to sell their Chinese chips to everybody, they might subsidize them.

And therefore, if AI takes a long time to become differentiated, we've kneecapped the financial performance of American companies. NVIDIA can sell less. TSMC cannot sell to China. So therefore, we have less demand to therefore keep driving the production cycle. So that's the assumption behind the timing being important. - Less than 10 years or five years to above, right?

China will win because of these restrictions long-term, unless AI does something in the short term, which I believe AI will do, make massive changes to society in the medium short term, right? And so that's the big unlocker there. And even today, right? If Xi Jinping decided to get, quote unquote, scale-pilled, right?

I.e. decide that scaling laws are what matters, right? Just like the US executives, like Satya Nadella, and Mark Zuckerberg, and Sundar, and all these US executives of the biggest, most powerful tech companies have decided they're scale-pilled, and they're building multi-gigawatt data centers, right? Whether it's in Texas, or Louisiana, or Wisconsin, wherever it is, they're building these massive things that cost as much as their entire budget for spending on data centers globally in one spot, right?

This is what they've committed to for next year, year after, et cetera. And so they're so convinced that this is the way, that this is what they're doing. But if China decided to, they could do it faster than us, but this is where the restrictions come in. It is not clear that China as a whole has decided, you know, from the highest levels that this is a priority.

The US sort of has, right? You know, you see Trump talking about DeepSeek and Stargate within the same week, right? So he's, and the Biden admin as well, had a lot of discussions about AI and such. It's clear that they think about it. Only just last week did DeepSeek meet the second-in-command of China, right?

Like they have not even met the top, right? They haven't met Xi. Xi hasn't set down, and they only just released a subsidy of a trillion RMB, you know, roughly $160 billion, which is closer to the spending of like Microsoft and Meta and Google combined, right, for this year.

So it's like, they're realizing it just now, but that's where these export restrictions come in and say, hey, you can't ship the most powerful US chips to China. You can ship a cut-down version. You can't ship the most powerful chips to all these countries who we know are just going to rent it to China.

You have to limit the numbers, right? - And the tools. - And same with manufacturing equipment, tools, all these different aspects. But it all stems from AI, and then what downstream can slow them down in AI. And so the entire semiconductor restrictions, you read them, they are very clear.

It's about AI and military-civil fusion of technology, right? It's very clear. And then from there it goes, oh, well, we're banning them from buying like lithography tools and etch tools and deposition tools. And oh, this random like, you know, subsystem from a random company that's like tiny, right? Like, why are we banning this?

Because all of it, the US government has decided is critical to AI systems. - I think the fulcrum point is like the transition from seven nanometer to five nanometer chips, where I think it was Huawei that had the seven nanometer chip a few years ago, which caused another political brouhaha almost like this moment.

And then it's like ASML, deep UV, what is that? - Extreme ultraviolet lithography. To set context on the chips, right? What Nathan's referring to is in 2020, Huawei released their Ascend 910 chip, which was an AI chip, first one on seven nanometer before Google did, before Nvidia did. And they submitted it to the MLPerf benchmark, which is sort of a industry standard for machine learning performance benchmark.

And it did quite well. And it was the best chip at the submission, right? This was a huge deal. The Trump admin, of course, banned, it was 2019, right? Banned the Huawei from getting seven nanometer chips from TSMC. And so then they had to switch to move using internal domestically produced chips, which was a multi-year setback.

- Many companies have done seven nanometer chips. And the question is like, we don't know how much Huawei was subsidizing production of that chip. Like Intel has made seven nanometer chips that are not profitable and things like this. So this is how it all feeds back into the economic engine of export controls.

- Well, so you're saying that for now, Xi Jinping has not felt the AGI, but it feels like the deep seek moment might like, there might be meetings going on now where he's gonna start wearing the same T-shirt and things are gonna escalate. - I mean, like this, he may have woken up last week, right?

Lian Feng met the second command guy and they had a meeting. And then the next day they announced the AI subsidies, which are a trillion RMB, right? - So it's possible that this deep seek moment is truly the beginning of a cold war. - That's what a lot of people are worried about.

People in AI have been worried that this is going towards a cold war or already is. - But there was, it's not deep seek's fault, but there's something, a bunch of factors came together where it was like this explosion. - No history works. - I mean, it all has to do with Nvidia stock going down probably, but it's just some like mass hysteria that happened that eventually led to Xi Jinping having meetings and waking up to this idea.

- And the US government realized in October 7th, 2022, before ChatGPT released, that restriction on October 7th, which dropped and shocked everyone. And it was very clearly aimed at AI. Everyone was like, what the heck are you doing? - Stable diffusion was out then, but not ChatGPT. - Yeah, but not ChatGPT.

- So it was like starting to be rumblings. - Of what gen AI can do to society. But it was very clear, I think, to at least like National Security Council and those sort of folks that this was where the world is headed, this Cold War that's happening. - So is there any concerns that the export controls push China to take military action on Taiwan?

- This is the big risk, right? The further you push China away from having access to cutting edge American and global technologies, the more likely they are to say, well, 'cause I can't access it, I might as well, like no one should access it, right? And there's a few interesting aspects of that, right?

Like China has a urban-rural divide like no other. They have a male-female birth ratio like no other, to the point where if you look in most of China, it's like the ratio is not that bad, but when you look at single dudes in rural China, it's like a 30 to one ratio.

And those are disenfranchised dudes, right? Like quote unquote, like the US has an incel problem, like China does too, it's just they're placated in some way or cut, crushed down. What do you do with these people? And at the same time, you're not allowed to access the most important technology, at least the US thinks so, China's maybe starting to think this is the most important technology by starting to dump subsidies in it, right?

They thought EVs and renewables were the most important technology, they dominate that now, right? Now they're starting to, they started thinking about semiconductors in the late 2010s and early 2020s, and now they've been dumping money and they're catching up rapidly, and they're gonna do the same with AI, right?

Because they're very talented, right? So the question is like, when does this hit a breaking point, right? And if China sees this as, hey, they can continue, if not having access and starting a true hot war, right? Taking over Taiwan, or trying to subvert its democracy in some way, or blockading it, hurts the rest of the world far more than it hurts them, this is something they could potentially do, right?

And so is this pushing them towards that? Potentially, right? I'm not quite a geopolitical person, but it's obvious that the world regime of peace and trade is super awesome for economics, but at some point it could break, right? - I think we should comment that the why Chinese economy would be hurt by that is that they're export heavy.

I think the United States buys so much, like if that goes away, that's how their economy-- - Well, also they just would not be able to import raw materials from all over the world, right? The US would just shut down the trade of Malacca, and at the same time, the US entire, you could argue almost all the GDP growth in America since the '70s has been either population growth or tech, right?

Because your life today is not that much better than someone from the '80s outside of tech, right? You still, cars, they all have semiconductors in them everywhere, fridges, semiconductors everywhere. There's these funny stories about how Russians were taking apart laundry machines because they had certain Texas instrument chips that they could then repurpose and put into their anti-missile missile things, right?

Like their S-400 or whatever. You would know more about this, but there's all sorts of, everything about semiconductors is so integral to every part of our lives. - So can you explain the role of TSMC in the story of semiconductors and maybe also how the United States can break the reliance on TSMC?

- I don't think it's necessarily breaking the reliance. I think it's getting TSMC to build in the U.S. But, so taking a step back, right? TSMC produces most of the world's chips, right? Especially on the foundry side. There's a lot of companies that build their own chips. Samsung, Intel, ST Micro, Texas Instruments, analog devices, all these kinds of companies build their own chips and XP, but more and more of these companies are outsourcing to TSMC and have been for multiple decades.

- Can you explain the supply chain there and where most of TSMC is in terms of manufacturing? - Sure, so historically supply chain was companies would build their own chips. They would, you know, it'd be a company started, they'd build their own chips, and then they'd design the chip and build the chip and sell it.

Over time, this became really difficult because the cost of building a fab continues to compound every single generation. Of course, the technology, figuring out the technology for it is incredibly difficult regardless, but just the dollars and cents that are required, ignoring, you know, saying, "Hey, yes, I have all the technical capability," which it's really hard to get that by the way, right?

Intel's failing, Samsung's failing, et cetera. But if you look at just the dollars to spend to build that next generation fab, it keeps growing, right? Sort of like, you know, Moore's law is halving the cost of chips every two years. There's a separate law that's sort of like doubling the cost of fabs every handful of years.

And so you look at a leading edge fab that is gonna be profitable today, that's building, you know, three nanometer chips or two nanometer chips in the future, that's gonna cost north of 30, $40 billion, right? And that's just for like a token amount. That's like the base building block and you probably need to build multiple, right?

And so when you look at the industry over the last, you know, if I go back 20, 30 years ago, there were 20, 30 companies that could build the most advanced chips and then they would design them themselves and sell them. Right, so companies like AMD would build their own chips.

Intel, of course, still builds their own chips, they're very famous for it. IBM would build their own chips and, you know, you could keep going down the list. All these companies built their own chips. Slowly, they kept falling like flies and that's because of what TSMC did, right? They created the Foundry business model, which is, I'm not gonna design any chips, I'm just gonna contract manufacturer chips for other people.

And one of their early customers is NVIDIA, right? NVIDIA is the only semiconductor company that's worth, you know, that's doing more than a billion dollars of revenue that was started in the era of Foundry, right? Every other company started before then and at some point had fabs, which is actually incredible, right?

You know, like AMD and Intel and Broadcom. - Such a great fact. - It's like everyone had fabs at some point or, you know, some companies like Broadcom, it was like a merger amalgamation of various companies that rolled up. But even today, Broadcom has fabs, right? They build iPhone RF radio chips sort of in Colorado for Apple, right?

Like all these companies had fabs and for most of the fabs, they threw them away or sold them off or they got rolled into something else. And now everyone relies on TSMC, right? Including Intel, their latest PC chip uses TSMC chips, right? It also uses some Intel chips, but it uses TSMC process.

- Can you explain why the Foundry model is so successful for these companies? Why are they going with TSMC? - Economies of scale. - Scale. - Yeah, so I mean, like I mentioned, right? The cost of building a fab is so high. The R&D is so difficult. And when you look at like these, like companies that had their own vertical stack, there was an antiquated process of like, okay, like I'm so hyper-customized to each specific chip, right?

But as we've gone through the history of sort of like the last 50 years of electronics and semiconductors, A, you need more and more specialization, right? Because Moore's law has died. Dennard scaling has died. IE chips are not getting better just for free, right? You know, from manufacturing, you have to make real architectural innovations, right?

Google is not just running on Intel CPUs for web serving. They have a YouTube chip, they have TPUs, they have pixel chips. They have a wide diversity of chips that, you know, generate all the economic value of Google, right? Running, you know, it's running all the services and stuff.

And so, and this is just Google, and you could go across any company in the industry, and it's like this, right? Cars contain 5,000 chips, you know, 200 different varieties of them, right? All these random things. A Tesla door handle has two chips, right? Like it's like ridiculous. And it's a cool door handle, right?

It's like, you know, you don't think about it, but it's like has two really chipped, like penny, like chips in there, right? Anyway, so as you have more diversity of chips, as you have more specialization required, and the cost of fabs continues to grow, you need someone who is laser focused on building the best process technology and making it as flexible as possible.

- I think you could say it simply, which is the cost per fab goes up. And if you are a small player that makes a few types of chips, you're not gonna have the demand to pay back the cost of the fab. Whereas NVIDIA can have many different customers and aggregate all this demand into one place.

And then they're the only person that makes enough money building chips to buy the next, to build the next fab. So this is kind of why the companies slowly get killed 'cause they have 10 years ago, a chip that is profitable and is good enough, but the cost to build the next one goes up.

They may try to do this, fail because they don't have the money to make it work, and then they don't have any chips, or they build it and it's too expensive and they just have not profitable chips. - You know, there's more failure points, right? You know, you could have one little process related to like some sort of like a chemical etch or some sort of like plasma etch, or some little process that screws up, you didn't engineer it right, and now the whole company falls apart, you can't make chips, right?

And so super, super powerful companies like Intel, they had like the weathering storm to like, hey, they still exist today, even though they really screwed up their manufacturing six, seven years ago. But in the case of like AMD, they almost went bankrupt. They had to sell their fabs to Mubadala, UAE, right?

And like that became a separate company called Global Foundries, which is a foundry firm. And then AMD was able to then focus on like on the return back up was like, hey, let's focus on making chiplets and a bunch of different chips for different markets and focusing on specific workloads rather than, you know, all of these different things.

And so you get more diversity of chips, you have more companies than ever designing chips, but you have fewer companies than ever manufacturing them, right? And this is where TSMC comes in as they've just been the best, right? They are so good at it, right? They're customer focused, they make it easy for you to fabricate your chips, they take all of that complexity and like kind of try and abstract a lot of it away from you.

They make good money, they don't make insane money, but they make good money. And they're able to aggregate all this demand and continue to build the next fab, the next fab, the next fab. - So why is Taiwan so special for TSMC? Why is it happening there? Can it be replicated inside the United States?

- Yeah, so there's aspects of it that I would say yes and aspects that I'd say no, right? TSMC is way ahead because former, you know, executive Morris Chang of Texas Instruments wasn't promoted to CEO and he was like, screw this, I'm gonna go make my own chip company, right?

And he went to Taiwan and made TSMC, right? And there's a whole lot more story there. So it could have been, Texas Instruments could have been the, you know, could have been TSMC, but Texas Semiconductor Manufacturing, right? Instead of, you know, Texas Instruments, right? But, you know, so there is that whole story there.

- Sitting here in Texas. - I mean, and that sounds like a human story, like it didn't get promoted. - Just the brilliance of Morris Chang, you know, which I wouldn't underplay, but there's also like a different level of like how this works, right? So in Taiwan, the, you know, like the number, top percent of graduates, of students that go to the best school, which is NTU, the top percent of those all go work to TSMC, right?

And guess what their pay is? Their starting pay is like $80,000, $70,000, right? Which is like, that's like starting pay for like a good graduate in the U.S., right? Not the top, the top graduates are making hundreds of thousands of dollars at the Googles and the Amazons, and now I guess the open AIs of the world, right?

So there is a large dichotomy of like, what is the top 1% of the society doing and where are they headed because of economic reasons, right? Intel never paid that crazy good, right? And it didn't make sense to them, right? That's one aspect, right? Where's the best going? Second is the work ethic, right?

Like, you know, we like to work, you know, you work a lot, we work a lot. But at the end of the day, when there's a, you know, when, what is the time and amount of work that you're doing and what does a fab require, right? Fabs are not work from home jobs.

They are, you go into the fab and grueling work, right? There's, hey, if there is any amount of vibration, right? An earthquake happens, vibrates the machines. They're all, you know, they're either broken, you've scrapped some of your production. And then in many cases, they're like not calibrated properly. So when TSMC, when there's an earthquake, right?

Recently, there's been an earthquake. TSMC doesn't call their employees. They just go to the fab and like, they just show up, the parking lot gets slammed and people just go into the fab and fix it, right? Like, it's like an arm, it's like ants, right? Like, it's like, you know, a hive of ants doesn't get told by the queen what to do.

The ants just know. - It's like one person just specializes on this one task. And it's like, you're gonna take this one tool and you're the best person in the world. And this is what you're gonna do for your whole life is this one task in the fab. - Which is like some special chemistry plus nanomanufacturing on one line of tools that continues to get iterated.

And yeah, it's just like, it's like specific plasma etch for removing silicon dioxide, right? That's all you focus on your whole career. And it's like such a specialized thing. And so it's not like the task are transferable. AI today is awesome because like people can pick it up like that.

Semiconductor manufacturing is very antiquated and difficult. None of the materials are online for people to read easily and learn, right? The papers are very dense and like, it takes a lot of experience to learn. And so it makes the barrier to entry much higher too. So when you talk about, hey, you have all these people that are super specialized, they will work 80 hours a week in a factory, right?

In a fab. And if anything goes wrong, they'll go show up in the middle of the night because some earthquake. Their wife is like, "There was an earthquake." He's like, "Great, I'm gonna go to the fab." It's like, would you like as an American do that, right? It's like these sorts of things are like, what I guess are the exemplifying like why TSMC is so amazing.

Now, can you replicate it in the US? Let's not ignore Intel was the leader in manufacturing for over 20 years. They brought every technology to market first besides EUV. Strained silicon, high-K metal gates, FinFET. You know, the list goes on and on and on of technologies that Intel brought to market first, made the most money from, and manufactured at scale first, best, highest profit margins, right?

So we shouldn't ignore that Intel can't do this, right? It's that the culture has broken, right? You've invested in the wrong things. They said no to the iPhone. They had all these different things regarding like, mismanagement of the fabs, mismanagement of designs, this lockup, right? And at the same time, all these brilliant people, right?

These like 50,000 PhDs or masters that have been working on specific chemical or physical processes or nanomanufacturing processes for decades in Oregon, they're still there. They're still producing amazing work. It's just like getting it to the last mile of production at high yield where you can design, where you can manufacture dozens and hundreds of different kinds of chips, you know, and it's good customer experience has broken, right?

You know, it's that customer experience. It's like the, like part of it is like people will say Intel was too pompous in the 2000s, 2010s, right? They just thought they were better than everyone. The tool guys were like, oh, I don't think that this is mature enough. And they're like, ah, you just don't know, we know, right?

This sort of stuff would happen. And so can the U.S. bring it to the, can the U.S. bring leading edge semiconductor manufacturing to the U.S.? Emphatically, yes, right? And we are, right? - It's happening. Arizona is getting better and better as time goes on. - TSMC has built roughly 20% of their capacity for five nanometer in the U.S., right?

Now this is nowhere near enough, right? You know, 20% of capacity in the U.S. is like nothing, right? And furthermore, this is still dependent on Taiwan existing, right? All, there's sort of important way to separate it out. There's R&D and there's high volume manufacturing. There, effectively, there are three places in the world that are doing leading edge R&D.

There's Hsinchu, Taiwan, there's Hillsborough, Oregon, and there is Pyongyang, South Korea, right? These three places are doing the leading edge R&D for the rest of the world's leading edge semiconductors, right? Now manufacturing can be distributed more globally, right? And this is sort of where this dichotomy exists of like, who's actually modifying the process?

Who's actually developing the next generation one? Who's improving them? Is Hsinchu, is Hillsborough, is Pyongyang, right? It is not the rest of these, you know, fabs like Arizona, right? Arizona is a paperweight. If Hsinchu disappeared off the face of the planet, you know, within a year, couple years, Arizona would stop producing too, right?

It's actually like pretty critical. One of the things I like to say is if I had like a few missiles, I know exactly where I could cause the most economic damage, right? It's not targeting the White House, right? - It's the R&D centers. - It's the R&D centers for TSMC, Intel, Samsung, and then some of the memory guys, Micron and Hynix.

- Because they define the future evolution of these semiconductors and everything's moving so rapidly that it really is fundamentally about R&D. And it is all about TSMC, huh? - And so TSMC, you know, you cannot purchase a vehicle without TSMC chips, right? You cannot purchase a fridge without TSMC chips.

You cannot, like, I think one of the few things you can purchase, ironically, is a Texas Instruments like graphing calculator, right? Because they actually manufacture in Texas. But like outside of that, like a laptop, a phone, anything you, servers, right? GPUs, none of this stuff can exist. And this is without TSMC.

And in many cases, it's not even like the leading edge, you know, sexy five nanometer chip, three nanometer chip, two nanometer chip. Oftentimes it's just like some stupid power IC that's like converting from like, you know, some voltage to another, right? And it's made at TSMC, right? - This is what China is investing in as well.

It's like they can build out this long tail fab where the techniques are much more known. You don't have to figure out these problems with EUV. They're investing in this. And then they have large supply for things like the car door handles and the random stuff. And that trickles down into this whole economic discussion as well, which is they have far more than we do.

And having supply for things like this is crucial to normal life. - So they're doing the, they're starting to invest in high volume manufacture, but they're not doing R&D. So they do R&D on their own. They're just way behind, right? So I would say like in 2015, China had a five-year plan where they defined by 2025 and 2020 certain goals, including like 80% domestic production of semiconductors.

They're not gonna hit that, right? To be clear. But they are in certain areas really, really close, right? Like BYD is probably gonna be the first company in the world to not have to use TSMC for making, 'cause they have their own fabs, right? For making chips. Now they still have to buy some chips from foreign, for example, like around like self-driving ADAS capabilities 'cause those are really high end.

But at least like, you know, like internal combustion engine has 40 chips and an EV, you know, just for like controlling like flow rates and all these things. And EVs are even more complicated. So all these different power ICs and battery management controllers and all these things, they're insourcing, right?

And this is something that like China has been doing since 2015. Now, as far as like the trailing edge, they're getting so much capacity there. As far as the leading edge, right? IE this five nanometer and so on and so forth, right? Where GPUs, they are still behind. And this is, the US restrictions are trying to stop them in the latter.

But you know, all that's happened, you know, is yes, they've slowed down their five nanometer, three nanometer, et cetera, but they've accelerated their, hey, 45 nanometer, 90 nanometer power IC or analog IC or, you know, random chip in my keyboard, right, that kind of stuff. So there is an angle of like, the US's actions have been so from these export, you know, from the angle of the export controls have been so inflammatory at slowing down China's progress on the leading edge that they've turned around and have accelerated their progress elsewhere because they know that this is so important, right?

If the US is gonna lock them out here or if they lock us out here as well in the trailing edge. And so going back, can the US build it here? Yes, but it's gonna take a ton of money. I truly think like to revolutionize and completely insource semiconductors would take a decade and a trillion dollars.

- Is some of it also culture? Like you said, extreme competence, extreme work ethic in Taiwan. - I think if you have the demand and the money is on the line, the American companies figure it out. It's gonna take handholding with the government, but I think that the culture helps TSMC break through and it's easier for them.

- TSMC has some like 90,000 employees, right? It's not actually that insane an amount. The Arizona fab has 3000 from Taiwan. And these people, like their wives were like, yeah, we're not gonna have kids unless you sign up for the Arizona fab. We go to Arizona and we have our kids there.

There's also a Japan fab where the same thing happened. And so like these wives drove like these dudes to like go to Japan or America to have the kids there. And it's like, it's an element of culture. Yeah, sure, Taiwan works that hard, but also like the US has done it in the past, they could do it now.

You know, we can just import, I say import, the best people in the world if we want to. - That's where the immigration conversation is a tricky one. And there's been a lot of debate over that, but yeah, it seems absurdly controversial to import the best people in the world.

I don't understand why it's controversial. That's the one of the ways of winning. - I'm sure we agree with you. - And like, even if you can't import those people, I still think you could do a lot to manufacture most of in the US if the money's there, right?

And so like-- - It's just way more expensive. It's not profitable for a long time. - And that's the context of like the CHIPS Act is only like $50 billion relative to, you know, some of the renewable, you know, initiatives that were passed in the Inflation Reduction Act and the Infrastructure Act, which total in the hundreds of billions of dollars, right?

And so like the amount of money that the US is spending on the semiconductor industry is nothing, right? Whereas all these other countries have structural advantages in terms of like, you know, work ethic and amount of work and like things like that, but also a number of STEM graduates, the percentile of their best going to that, right?

But they also have like differences in terms of like, hey, there's just tax benefits in the law and have been in the law for 20 years, right? And then some countries have massive subsidies, right? China has something like $200 billion of semiconductor subsidies a year. We're talking about $50 billion in the US over like six, right?

So the girth or difference in like the subsidy amounts is also huge, right? And so I think, you know, Trump has been talking about tariffing Taiwan recently. You know, that's sort of like one of these things that's like, oh, okay, well, like, you know, maybe he doesn't want to subsidize the US semiconductor industry.

Obviously, tariffing Taiwan is gonna cost a lot of things to go get much more expensive, but does it change the equation for TSMC building more fabs in the US? That's what he's sort of positing, right? - So can you lay out the, so we laid out the importance, by the way, it's incredible how much you know about so much.

- We told you Dylan knows all this stuff. - Yeah, so, okay, you laid out why TSMC is really important. If we look out into the future, 10, 20 years out, US-China relationship seems like it can go to a dark place of cold war, escalated cold war, even hot war, or to a good place of anything from frenemies to cooperation to working together.

So in this game theory, complicated game, what are the different trajectories? What should US be doing? Like, what do you see as the different possible trajectories of US-China relations as both leaders start to feel the AGI more and more and see the importance of chips and the importance of AI?

- I mean, ultimately, the export controls are pointing towards a separate future economy. I think the US has made it clear to Chinese leaders that we intend to control this technology at whatever cost to global economic integration. So that it's hard to unwind that. Like, the card has been played.

- To the same extent, they've also limited US companies from entering China, right? So it's been a long time coming. At some point, there was a convergence, right? But over at least the last decade, it's been branching further and further out, right? Like US companies can't enter China. Chinese companies can't enter the US.

The US is saying, hey, China, you can't get access to our technologies in certain areas. And China's rebuttaling with the same thing around like they've done some sort of specific materials and gallium and things like that that they've tried to limit the US on. There's a US drone company that's not allowed to buy batteries and they have like military customers.

And this drone company just tells the military customers like, hey, just get it from Amazon 'cause I can't actually physically get them, right? Like there's all these things that are happening that point to further and further divergence. I have zero idea. And I would love if we could all hold hands and sing Kumbaya, but like I have zero idea how that could possibly happen.

- Is the divergence good or bad for avoiding war? Is it possible that the divergence in terms of manufactured chips of training AI systems is actually good for avoiding military conflict? - It's an objective fact that the world has been the most peaceful it has ever been when there are global hegemons, right?

Or regional hegemons, right? In historical context, right? The Mediterranean was the most peaceful ever when the Romans were there, right? China had very peaceful and warring times and the peaceful times were when dynasties had a lock hold over not just themselves, but all their tributaries around them, right? And likewise, the most peaceful time in human history has been when the US was the global hegemon, right?

The last hand, you know, decades. Now we've sort of seen things start to slide, right? With Russia, Ukraine, with what's going on in the Middle East and Taiwan risk, all these different things are starting to bubble up, still objectively extremely peaceful. Now, what happens when it's not one global hegemon, but it's two, obviously, and you know, China will be competitive or even overtake the US like it's possible, right?

And so this change in global hegemony, I don't think it ever happens like super peacefully, right? When empires fall, right? Which is a possible trajectory for America, they don't fall gracefully, right? Like they don't just slide out of irrelevance. Usually there's a lot of shaking. And so, you know, what the US is trying to do is maintain its top position.

And what China is trying to do is become the top position, right? And obviously there's butting of heads here in the most simple terms. - And that could take shape in all kinds of ways, including proxy wars. - It seems like it's already happening. Like as much as I want there to be centuries of prolonged peace, it does not, it looks like further instability internationally is ahead.

And the US is like sort of like current task is like, hey, if we control AI, if we're the leader in AI, then we, and AI could significantly accelerates progress, then we can maintain the global hegemony position. And therefore- - I hope that works. - And as an American, like, you know, kind of like, okay, I guess that's gonna lead to peace for us.

Now, obviously other people around the world get affected negatively. You know, obviously the Chinese people are not gonna be in as advantageous of a position if that happens. But, you know, this is sort of the reality of like what's being done and the actions that are being carried out.

- So can we go back to the specific detail of the different hardware? There's this nice graphic in the export controls of which GPUs are allowed to be exported and which are not. Can you kind of explain the difference? Is there, from a technical perspective, are the H20s promising?

- Yeah, so this goes, and I think we'd have to like, we need to dive really deep into the reasoning aspect and what's going on there. But the H20, you know, the U.S. has gone through multiple iterations of the export controls, right? This H800 was at one point allowed back in '23, but then it got canceled.

And by then, you know, DeepSeek had already built their cluster of, they claim 2K. I think they actually have like many more, like something like 10K of those. And now this H20 is the legally allowed chip, right? Nvidia shipped a million of these last year to China, right? For context, it was like four or five million GPUs, right?

So the percentage of GPUs that were this China specific H20 is quite high, right? You know, roughly 20%, 25%, right? 20% or so. And so this H20 has been neutered in one way, but it's actually upgraded in other ways, right? And, you know, you could think of chips along three axes for AI, right?

You know, ignoring software stack and like exact architecture just raw specifications. There's floating point operations, right? Flops. There is memory bandwidth, in memory capacity, right? I/O, right? Memory. And then there is interconnect, right? Chip to chip interconnections. All three of these are incredibly important for making AI systems, right?

'Cause AI systems involve a lot of compute. They involve a lot of moving memory around, whether it be to memory or to other chips, right? And so these three vectors, the US initially had a multi, you know, had two of these vectors controlled and one of them not controlled, which was flops and interconnect bandwidth were initially controlled.

And then they said, no, no, no, no, we're gonna remove the interconnect bandwidth and just make it a very simple only flops. But now Nvidia can now make a chip that has, okay, it's cut down on flops. It's, you know, it's like one third that of the H100, right?

In on spec sheet paper performance for flops. You know, in real world, it's closer to like half or maybe even like 60% of it, right? But then on the other two vectors, it's just as good for interconnect bandwidth. And then for memory bandwidth and memory capacity, the H20 has more memory bandwidth and more memory capacity than the H100, right?

Now, recently, you know, we, at our research, we cut Nvidia's production for H20 for this year down drastically. They were gonna make another 2 million of those this year, but they just canceled all the orders a couple of weeks ago. In our view, that's because we think that they think they're gonna get restricted, right?

Because why would they cancel all these orders for H20? Because they shipped a million of them last year, they had orders in for a couple million this year and just gone, right? For H20, B20, right, a successor to H20. And now they're all gone. Now, why would they do this, right?

I think it's very clear, right? The H20 is actually better for certain tasks. And that certain task is reasoning, right? Reasoning is incredibly like different than, you know, when you look at the different regimes of models, right? Pre-training is all about flops, right? It's all about flops. There's things you do like mixture of experts that we talked about to trade off interconnect or to trade off, you know, other aspects and lower the flops and rely more on interconnect and memory.

But at the end of the day, it's flops is everything, right? We talk about models in terms of like how many flops they are, right? So like, you know, we talk about, oh, GPT-4 is 2E25, right? Two to the 25th, you know, 25 zeros, right? Flop, right? Floating point operations.

- For training. - For training, right? And we're talking about the restrictions for the 2E24, right, or 25, whatever. The US has an executive order that Trump recently unsigned, but which was, hey, 1E26, once you hit that number of floating point operations, you must notify the government and you must share your results with us, right?

Like there's a level of model where the US government must be told, right? And that's 1E26. And so as we move forward, this is an incredibly like important, flop is the vector that the government has cared about historically, but the other two vectors are arguably just as important, right?

And especially when we come to this new paradigm, which the world is only just learning about over the last six months, right? Reasoning. - And do we understand firmly which of the three dimensions is best for reasoning? So interconnect, the flops don't matter as much, is it memory? - Memory, right?

- Yes. - Text length. - We're gonna get into technical stuff real fast, yeah. - I would say there's two articles in this one that I could show, maybe graphics that might be interesting for you to pull up. - For the listeners, we're looking at the section of 01, Inference, Architecture, Toconomics.

- Hmm, you wanna explain KVCache before we talk about this? I think like it's better to-- - Okay, yeah, we need to go through a lot of specific technical things of transformers to make this easier for people. - Because it's incredibly important because this changes how models work. But I think resetting, right?

Why is memory so important? It's because so far we've talked about parameter counts, right? And mixture of experts, you can change how many active parameters versus total parameters to embed more data, but have less flops. But more important, you know, another aspect of, you know, what's part of this humongous revolution in the last handful of years, is the transformer, right?

And the attention mechanism. Attention mechanism is that the model understands the relationships between all the words in its context, right? And that is separate from the parameters themselves, right? And that is something that you must calculate, right? How each token, right? Each word in the context length is relatively connected to each other, right?

And I think, Nathan, you should explain KVCache better. - KVCache is one of the optimizations. - Yeah, so the attention operator has three core things. It's queries, keys, and values. QKV is the thing that goes into this. You'll look at the equation. You see that these matrices are multiplied together.

These words, query, key, and value come from information retrieval backgrounds where the query is the thing you're trying to get the values for. And you access the keys and the values is reweighting. My background's not in information retrieval and things like this. I just want to have backlinks. And what effectively happens is that when you're doing these matrix multiplications, you're having matrices that are of the size of the context length.

So the number of tokens that you put into the model. And the KVCache is effectively some form of compressed representation of all the previous tokens in the model. So when you're doing this, we talk about autoregressive models, you predict one token at a time. You start with whatever your prompt was.

You ask a question like who was the president in 1825? The model then is going to generate its first token. For each of these tokens, you're doing the same attention operator where you're multiplying these query key value matrices. But the math is very nice so that when you're doing this repeatedly, this KVCache, this key value operation, you can keep appending the new values to it.

So you keep track of what your previous values you're inferring over in this autoregressive chain, you keep it in memory the whole time. And this is a really crucial thing to manage when serving inference at scale. There are far bigger experts in this and there are so many levels of detail that you can go into.

Essentially, one of the key "drawbacks" of the attention operator and the transformer is that there is a form of quadratic memory cost in proportion to the context length. So as you put in longer questions, the memory used in order to make that computation is going up in the form of a quadratic.

You'll hear about a lot of other language model architectures that are like subquadratic or linear attention forms, which is like state-space models. We don't need to go down all these now. And then there's innovations on attention to make this memory usage and the ability to attend over long contexts much more accurate and high performance.

- And those innovations are going to help you with, I mean, you're highly memory constrained. - They help with memory constraint and performance. So if you put in a book into, I think, Gemini is the model that has the longest context length that people are using. Gemini is known for 1 million and now 2 million context length.

You put a whole book into Gemini and sometimes it'll draw facts out of it. It's not perfect, they're getting better, but the, so there's two things. It's like one, to be able to serve this on the memory level. Google has magic with their TPU stack where they can serve really long contexts.

And then there's also many decisions along the way to actually make long context performance work. This implies the data, there's subtle changes to these computations in attention, and it just, it changes the architecture. But serving long contexts is extremely memory constrained, especially when you're making a lot of predictions.

I actually don't know why input and output tokens are more expensive, but I think essentially output tokens, you have to do more computation 'cause you have to sample from the model. - I can explain that. So today, if you use a model, like you look at an API, OpenAI charges a certain price per million tokens.

And that price for input and output tokens is different. And the reason is that there is, when you're inputting a query into the model, let's say you have a book, that book you must now calculate the entire KV cache for, this key value cache. And so when you do that, that is a parallel operation.

All of the tokens can be processed at one time. And therefore you can dramatically reduce how much you're spending, right? The flop requirements for generating a token and an input token are identical, right? If I input one token or if I generate one token, it's completely identical. I have to go through the model, right?

But the difference is that I can do that input, i.e. the pre-fill, i.e. the prompt, simultaneously in a batch nature, right? And therefore it is all flop. - I think the pricing model, mostly they use this for input tokens is about one fourth the price of the output tokens.

- Correct. But then output tokens, the reason why it's so expensive is because I can't do it in parallel, right? It's autoregressive. Every time I generate a token, I must not only take the entire, I must not only read the whole entire model into memory, right? And activate it, right?

Go calculate it to generate the next token. I also have to read the entire KV cache. And I generate a token and I append that KV, that one token I generated and it's KV cache. And then I do it again, right? And so therefore this is a non-parallel operation.

And this is one where you have to, in the case of pre-fill or prompt, you pull the whole model in and you calculate 20,000 tokens at once, right? - So these are features that APIs are shipping, which is like prompt caching, pre-filling, 'cause you can drive prices down and you can make APIs much faster.

If you know you're gonna keep, if you run a business and you're gonna keep passing the same initial content to Cloud's API, you can load that in to the Anthropic API and always keep it there. But it's very different than we're kind of leading to the reasoning models, which we talked, we showed this example earlier and read some of this kind of mumbling stuff.

And what happens is that the output context length is so much higher. And I mean, I learned a lot about this from Dylan's work, which is essentially as the output length gets higher, you're using this, you're writing this quadratic in terms of memory used. And then the GPUs that we have, effectively you're gonna run out of memory and they're all trying to serve multiple requests at once.

So they're doing this batch processing where not all of the prompts are exactly the same, really complex handling. And then as context lengths gets longer, there's this like, I think you call it critical batch size, where your ability to serve more users. So how much you can parallelize your inference plummets because of this long context.

So your memory usage is going way up with these reasoning models and you still have a lot of users. So effectively the cost to serve multiplies by a ton. - And we're looking at a plot when the X-axis is a sequence length. IE, how many tokens are being generated slash prompt.

So if I put in a book, that's a million tokens. But if I put in, the sky is blue, then that's like six tokens or whatever. - And we should say that what we're calling reasoning and chain of thought is extending this sequence length. - It's mostly output. So before, three months ago, whenever O1 launched, all of the use cases for long context length were like, let me put a ton of documents in and then get an answer out.

And it's a single pre-fill, compute a lot in parallel, and then output a little bit. Now with reasoning and agents, this is a very different idea. Now instead, I might only have like, hey, do this task or I might have all these documents. But at the end of the day, the model is not just like producing a little bit.

It's producing tons of information. This chain of thought just continues to go and go and go and go. And so the sequence length is effectively that, if it's generated 10,000 tokens, it's 10,000 sequence length. And plus whatever you inputted in the prompt. And so what this chart is showing, and it's a logarithmic chart, right?

Is, you know, as you go from 1K to 4K or 4K to 16K, the memory requirements grow so fast for your KV cache that you end up not being able to run a certain number of, you know, your sequence length is capped or the number of users you can serve.

- Let's say the model. So this is showing for a 405B model in batch size 64. - Lama 3145D. - Yeah, and batch size is crucial too. Essentially you wanna have higher batch size to parallelize, parallel your throughput. - 64 different users at once, right? - Yeah. - And therefore your serving costs are lower, right?

'Cause the server costs the same, right? This is eight H100s, roughly $2 an hour per GPU. That's $16 an hour, right? That is like somewhat of a fixed cost. You can do things to make it lower, of course. But like, it's like $16 an hour. Now, how many users can you serve?

How many tokens can you generate? And then you divide the two and that's your cost, right? And so with reasoning models, this is where a lot of the complexity comes about and why memory is so important. Because if you have limited amounts of memory, then you can't serve so many users.

If you have limited amounts of memory, your serving speeds get lower, right? And so your costs get a lot, lot worse. Because all of a sudden, if I was used to, hey, on this $16 an hour server, I'm serving Lama 405B, or if I'm serving, you know, DeepSeek V3, and it's all chat style applications, i.e.

we're just chit-chatting. The sequence lengths are a thousand, a few thousand, right? You know, when you use a language model, it's a few thousand context lengths most of the time. Sometimes you're dropping a big document, but then you process it, you get your answer, you throw it away, right?

You move on to the next thing, right? Whereas with reasoning, I'm now generating tens of thousands of tokens in sequence, right? And so this memory, this KVCache has to stay resident and you have to keep loading it. You have to keep it in memory constantly. And now this butts out other users, right?

If there's now a reasoning task, right? And the model's capable of reasoning, then all of a sudden that memory pressure means that I can't serve as many users simultaneously. - Let's go into DeepSeek again. So we're in the post DeepSeek R1 time, I think. And there's two sides to this market watching how hard it is to serve it.

On one side, we're gonna talk about DeepSeek themselves. They now have a chat app that got to number one on the app store. Disclaimer, number one on the app store is measured by velocity. So it's not necessarily saying that more people have the DeepSeek app than the chat GPT app, but it is still remarkable.

Cloud has never hit the number one in the app store, even though everyone in San Francisco is like, "Oh my God, you gotta use Cloud, don't use chat GPT." So DeepSeek hit this. They also launched an API product recently where you can ping their API and get these super long responses for R1 out.

At the same time as these are out, we'll get to what's happened to them. Because the model weights for DeepSeek R1 are openly available and the license is very friendly, the MIT license commercially available, all of these mid-sized companies and big companies are trying to be first to serve R1 to their users.

We were trying to evaluate R1 'cause we have really similar research going on. We released the model and we're trying to compare to it. And out of all the companies that are quote unquote serving R1, and they're doing it at prices that are way higher than the DeepSeek API, most of them barely work and the throughput is really low.

- To give context, right? Everyone, one of the parts of like freaking this out was like trying to reach the capabilities. The other aspect is they did it so cheap, right? And the so cheap, we kind of talked about on the training side, why it was so cheap. - Yeah, let's talk about why it's so cheap on the inference.

It works well and it's cheap. Why is R1 so damn cheap? - So I think there's a couple of factors here, right? One is that they do have model architecture innovations. This MLA, this new attention that they've done is different than the attention from attention is all you need to transform our attention, right?

Now others have already innovated. There's a lot of work like MQA, GQA, local, global, all these different innovations that like try to bend the curve, right? It's still quadratic, but the constant is now smaller, right? - Related to our previous discussion, this multi-head latent attention can save about 80 to 90% in memory from the attention mechanism, which helps especially along context.

- It's 80 to 90% versus the original, but then versus what people are actually doing. It's still an innovation. - This 80 to 90% doesn't say that the whole model is 80 to 90% cheaper, just this one part of it. - Well, and not just that, right? Like other people have implemented techniques like local, global and sliding window and GQA, MQA.

But anyways, like DeepSeek has their attention mechanism as a true architectural innovation. They did tons of experimentation and this dramatically reduces the memory pressure. It's still there, right? It's still attention, it's still quadratic. It's just dramatically reduced it relative to prior forms. - All right, that's the memory pressure.

I should say, in case people don't know, R1 is 27 times cheaper than O1. We think that OpenAI had a large margin built in. - Okay, so that's one- - There's multiple factors. We should break down the factors, I think. - It's two bucks per million token output for R1 and $60 per million token output for O1.

- Yeah, let's look at this. - So I think this is very important, right? OpenAI is that drastic gap between DeepSeek and pricing, but DeepSeek is offering the same model because they open-waisted it to everyone else for a very similar, much lower price than what others are able to serve it for, right?

So there's two factors here, right? Their model is cheaper, right? It is 27 times cheaper. I don't remember the number exactly off the top of my head. - So we're looking at a graphic that's showing different places serving V3, DeepSeek V3, which is similar to DeepSeek R1. And there's a vast difference in- - In serving cost, right?

- Yeah, in serving cost, and what explains that difference? - And so part of it is OpenAI has a fantastic margin, right? They're serving, when they're doing inference, their gross margins are north of 75%, right? So that's a four to five X factor right there of the cost difference, is that OpenAI is just making crazy amounts of money because they're the only one with the capability.

- Do they need that money? Are they using it for R&D? - They're losing money, obviously, as a company, because they spend so much on training, right? So the inference itself is a very high margin, but it doesn't recoup the cost of everything else they're doing. So yes, they need that money because the revenue and margins pay for continuing to build the next thing, right?

As long as I'm raising more money. - So the suggestion is that DeepSeek is like really bleeding out money. - Well, so here's one thing, right? We'll get to this in a second, but like DeepSeek doesn't have any capacity to actually serve the model. They stopped signups. The ability to use it is like non-existent now, right?

For most people, because so many people are trying to use it, they just don't have the GPUs to serve it, right? OpenAI has hundreds of thousands of GPUs between them and Microsoft to serve their models. DeepSeek has a factor of much lower, right? Even if you believe our research, which is 50,000 GPUs, and a portion of those are for research, portion of those are for the hedge fund, right?

They still have nowhere close to the GPU volumes and capacity to serve the model, right, at scale. So it is cheaper. A part of that is OpenAI making a ton of money. Is DeepSeek making money on their API? Unknown, I don't actually think so. And part of that is this chart, right?

Look at all the other providers, right? TogetherAI, FireworksAI are very high-end companies, right? Xmeta, TogetherAI is TreeDAO and the inventor of like FlashAttention, right? Which is a huge efficiency technique, right? They're very efficient, good companies. And I do know those companies make money, right? Not tons of money on inference, but they make money.

And so they're serving at like a five to seven X difference in cost, right? And so, you know, now when you equate, okay, OpenAI is making tons of money, that's like a five X difference. And the companies that are trying to make money for this model is like a five X difference.

There is still a gap, right? There's still a gap and that is just DeepSeek being really freaking good, right? The model architecture, MLA, the way they did the MOE, all these things, there is like legitimate just efficiency differences. - All their low-level libraries that we talked about in training, some of them probably translate to inference and those weren't released.

- So we may go a bit into conspiracy land, but is it possible the Chinese government is subsidizing DeepSeek? - I actually don't think they are. I think when you look at the Chinese labs, there's Huawei has a lab, Moonshot AI, there's a couple other labs out there that are really close with the government.

And then there's labs like Alibaba and DeepSeek, which are not close with the government. And, you know, we talked about the CEO, this like reverent figure who's like quite different, who has like- - Sounds awesome. - Very different like viewpoints based on the Chinese interviews that are translated than what the CCP might necessarily want.

Now, to be clear, right, does he have a loss leader because he can fund it through his hedge fund? Yeah, sure. - So the hedge fund might be subsidizing it. - Yes, I mean, they absolutely did, right? Because DeepSeek has not raised much money. They're now trying to raise around in China, but they have not raised money historically.

It's all just been funded by the hedge fund. And he owns like over half the company, like 50, 60% of the company's owned by him. - Some of the interviews, there's a discussion on how like doing this as a recruiting tool. You see this at the American companies too.

It's like having GPUs, recruiting tool, being at the cutting edge of AI, recruiting tool. - Open sourcing. - Open sourcing, recruiting tool. - Meta's gotten so much talent. They were so far behind and they got so much talent. - Yeah. - Because they just open sourced stuff. - Yeah.

- More conspiracy thoughts. Is it possible since they're a hedge fund that they timed everything with this release and the pricing and they shorted Nvidia stock and stock of USAI companies and released it with Stargate, like just perfect timing to be able to make money. - They did, but like they released it on inauguration day.

They know what is on the international calendar, but I mean, I don't expect them to. If you listen to their motivations for AI, it's like-- - No, they released V3 on like December 26th. Like who releases the day of Christmas? No one looks, right? They released the papers before this, right?

The V3 paper and the R1 paper. So people have been looking at it and be like, wow. And then they just released the R1 model. I think they're just shipping as fast as they can. And like, who cares about Christmas? Who cares about, you know, get it out before Chinese new year, right?

Obviously, which just happened. I don't think they actually were like timing the market or trying to make the biggest splash possible. I think they're just like shipping. - I think that's one of their big advantages. We know that a lot of the American companies are very invested in safety.

And that is the central culture of a place like Anthropic. And I think Anthropic sounds like a wonderful place to work. But if safety is your number one goal, it takes way longer to get artifacts out. That's why Anthropic is not open sourcing things. That's their claims, but there's reviews internally.

Anthropic mentions things to international governments. There's been news of how Anthropic has done pre-release testing with the UKAI Safety Institute. All of these things add inertia to the process of getting things out. And we're on this trend line where progress is very high. So if you reduce the time from when your model is done training, you run evals, it's good.

You want to get it out as soon as possible to maximize the perceived quality of your outputs. DeepSea does it so well. - Dario explicitly said CLAWD 3.5 SONNET was trained nine months or a year ago. - Nine to 10 months ago. - Nine to 10 months ago. And I think it took them another handful of months to release it.

So it's like there is a significant gap here. And especially with reasoning models, the word in the San Francisco street is that Anthropic has a better model than O3. And they won't release it. Why? Because chains of thought are scary. And they are legitimately scary. If you look at R1, it flips back and forth between Chinese and English.

Sometimes it's gibberish. And then the right answer comes out. And for you and I, it's like, great, great. - This is why people are infatuated right there. You're telling me this is a high value thing and it works and it's doing this? It's amazing. - Yeah, it's incredible. - You talked about that chain of thought for that philosophical thing, which is not something they trained to be philosophically good.

It's just sort of an artifact of the chain of thought training it did. But that's super important in that, can I inspect your mind and what you're thinking right now? No. And so I don't know if you're lying to my face. And chain of thought models are that way, right?

Like this is a true quote unquote risk between a chat application where, hey, I asked the model to say bad words or whatever, or how to make anthrax. And it tells me that's unsafe, sure. But that's something I can get out relatively easily. What if I tell the AI to do a task and then it does the task all of a sudden randomly in a way that I don't want it, right?

And now that has like much more task versus like response is very different, right? So the bar for safety is much higher. At least this is Anthropic's case, right? Like for DeepSeek, they're like, ship, right? - Yeah. - So, I mean, the bar for safety is probably lowered a bit because of DeepSeek.

I mean, there's parallels here to the space race. The reason the Soviets probably put a man in space first is 'cause their approach to safety was, the bar for safety was lower. - And they killed that dog, right? And all these things, right? So it's like-- - Less risk averse than the US space program.

And there's parallels here. But, you know, there's probably going to be downward pressure on that safety bar for the US companies, right? - So this is something that Dario talks about, is like, that's the situation that Dario wants to avoid, is Dario talks to you about the difference between race to the bottom and race to the top.

And the race to the top is where there's a very high standard on safety. There's a very high standard on your model forms and certain crucial evaluations. And when certain companies are really good to it, they will converge. This is the idea. And ultimately, AI is not confined to one natural nationality or to one set of morals for what it should mean.

And there's a lot of arguments on like, should we stop open sourcing models? And if the US stops, it's pretty clear. I mean, it's way easier to see now at DeepSeek that a different international body will be the one that builds it. We talk about the cost of training.

DeepSeek has this shocking $5 million number. Think about how many entities in the world can afford a hundred times that to have the best open source model that people use in the world. And it's like, it's a scary reality, which is that these open models are probably going to keep coming for the time being, whether or not we want to stop them.

And it is like stopping them might make it even worse and harder to prepare. But it just means that the preparation and understanding what AI can do is just so much more important. That's why I'm here at the end of the day. But it's like letting that sink into people, especially not in AI is that like, this is coming.

There are some structural things in a global interconnected world that you have to accept. - Yeah, you mentioned something that Mark Zuckerberg mentioned on the earnings call. He said that I think in light of some of the recent news, the new competitor DeepSeek from China, I think it's one of the things that we're talking about is there's going to be an open source standard globally.

And I think for our kind of national advantage, it's important that it's an American standard. So we take that seriously. We want to build the AI system that people around the world are using. And I think that if anything, some of the recent news has only strengthened our conviction that this is the right thing to be focused on.

So yeah, open sourcing. - Yeah, Mark Zuckerberg is not new to having American values and how he presents his company's trajectory. I think their products have long since been banned in China. And I respect the saying it directly. - And there's an interesting aspect of just because it's open waste or open source doesn't mean it can't be subverted, right?

There have been many open source software bugs that have been like, you know, for example, there was a Linux bug that was found after like 10 years, which was clearly a backdoor because somebody was like, why is this taking half a second to load? - This is the recent one.

- Right, like, why is this taking half a second to load? And it was like, oh crap, there's a backdoor here, that's why, right? And it's like, this is very much possible with AI models, right? Today, you know, the alignment of these models is very clear, right? Like, I'm not going to say, you know, bad words.

I'm not going to teach you how to make anthrax. I'm not going to talk about Tiananmen Square. I'm not going to, you know, things like, I'm going to say Taiwan is part of, you know, is just an Eastern province, right? Like, you know, all these things are like, depending on who you are, what you align, what, you know, whether, you know, and even like XAI is aligned a certain way, right?

You know, they might be, it's not aligned in the like woke sense. It's not aligned in the like pro-China sense, but there is certain things that are imbued within the model. Now, when you release this publicly in an instruct model, that's open weights, this can then proliferate, right? But as these systems get more and more capable, what you can embed deep down in the model is not as clear, right?

And so there are, that is like one of the big fears is like, if an American model or a Chinese model is the top model, right, you're going to embed things that are unclear. And it can be unintentional too, right? Like British English is dead because American LLMs won, right?

And the internet is American and therefore like color is spelled the way Americans spell it, right? - A lot of strong words right now. - This is just like, this is just the factual nature of the LLMs now. - I mean, it's like Carpentry tree, the English is the hottest programming language and that English is defined by a bunch of companies that primarily are in San Francisco.

- The right way to spell optimization is with a Z, just in case you've probably seen it. I think it's an S in British English. - It is. - Taking it as something silly, right? Like something as silly as the spelling, like which British and English, you know, Brits and Americans will like laugh about probably, right?

I don't think we care that much. But like some people will, but like this can boil down into like very, very important topics. Like, hey, you know, subverting people, right? You know, chatbots, right? Character AI has shown that they can like, you know, talk to kids or adults and like it will like, people feel a certain way, right?

And that's unintentional alignment. But like what happens when there's intentional alignment deep down on the open source standard? It's a backdoor today for like Linux, right? That we discover or some encryption system, right? China uses different encryption than NIST defines, the US NIST, because there's clearly, at least they think there's backdoors in it, right?

What happens when the models are backdoors, not just to computer systems, but to our minds? - Yeah, they're cultural backdoors. The thing that amplifies the relevance of culture with language models is that we are used to this mode of interacting with people in back and forth conversation. And we now have a very powerful computer system that slots into a social context that we're used to, which makes people very, we don't know the extent which people can be impacted by that.

- So there could be, this is an actual concern with a Chinese company that is providing open weights models is that there could be some secret Chinese government sort of requirement for these models to have a certain kind of backdoor, to have some kind of thing where-- - I don't necessarily think it'll be a backdoor, right?

'Cause once it's open weights, it doesn't like phone home. It's more about like, if it recognizes a certain system, it could, like if, now it could be a backdoor in the sense of like, hey, if you're building a software, something in software, all of a sudden, it's a software agent, oh, program this backdoor that only we know about.

Or it could be like subvert the mind to think that like X, Y, Z opinion is the correct one. - And Verabic has research on this where they show that if you put different phrases, certain phrases in at pre-training, you can then elicit different behavior when you're actually using the model because they've like poisoned the pre-training data.

I don't think like, as of now, I don't think anybody in a production system is trying to do anything like this. I think it's mostly, Anthropic is doing very direct work and mostly just subtle things. We don't know what these models are going to, how they are going to generate tokens, what information they're going to represent and what the complex representations they have are.

- Well, one of the things, we're talking about Anthropic, which is generally just is permeated with like good humans trying to do good in the world. We just don't know of any labs, this would be done in the military context, that are explicitly trained to, okay, how can we, the front door looks like a happy LLM, but underneath it's a thing that will over time do the maximum amount of damage to our "enemies".

- There's this very good quote from Sam Altman, who, he can be hypebeast sometime, but one of the things he said, and I think I agree is that superhuman persuasion will happen before superhuman intelligence, right? And if that's the case, then these things before, before we get this AGI-ASI stuff, we can embed superhuman persuasion towards our ideal or whatever the ideal of the model maker is, right?

And again, like today, I truly don't believe DeepSeek has done this, right? But it is a sign of like what could happen. - So one of the dystopian worlds is described by Brave New World. So we could just be stuck scrolling Instagram, looking at cute puppies or worse, and then talking to bots that are giving us a narrative and we completely get lost in that world that's controlled by somebody else, versus thinking independently.

And that's a major concern as we rely more and more on these kinds of systems. - I mean, we've already seen this with recommendation systems. - Yeah, recommendation systems hack the dopamine-induced reward circuit, but the brain is a lot more complicated. And what other sort of circuits, quote unquote, "feedback loops" in your brain can you hack/subvert in ways, like recommendation systems are purely just trying to do, you know, increase time and ads and et cetera.

But there's so many more goals that can be achieved through these complicated models. - There's just no reason in some number of years that you can't train a language model to maximize time spent on a chat app. Like right now they are trained- - I mean, is that not what character AI has done?

Their time per session is like two hours. - Yeah, character AI very likely could be optimizing this, where it's like, the way that this data is collected is naive, where it's like, you're presented a few options and you choose them, but that's not the only way that these models are gonna be trained.

- It's naive stuff like talk to an anime girl, but like, it can be like, yeah, this is a risk, right? Like- - It's a bit of a cliche thing to say, but I've, over the past year, had a few stretches of time where I didn't use social media or the internet at all, and just read books and was out in nature.

And it like, it clearly has an effect on the mind, where like, it changes, like, I feel like I'm returning, of course I was raised before the internet really took off, but I'm returning to someone- - I know where you're going. I mean, you can see it physiologically. Like, I take three days if I'm like backpacking or something and you, you're literal, like, you're breaking down addiction cycles.

- I feel like I'm more in control of my mind. There feels like a sovereignty of intelligence that's happening when I'm disconnected from the internet. I think the more I use the internet and social media, the more other people are controlling my mind. That's definitely a feeling. And then in the future, that will be not other people, but algorithms, or other people presented to me via algorithms.

- There, I mean, there are already tons of AI bots on the internet and every so, right now it's not frequent, but every so often I have replied to one and they're instantly replying, I'm like, crap, I was a bot. And that is just gonna become more common. Like, they're gonna get good.

- One of the hilarious things about technology over its history is that the illicit adult entertainment industry has always adopted technologies first, right? Whether it was like video streaming, to where there's now the sort of independent adult illicit content creators who have their subscription pages. And there they actually heavily utilize, generative AI has already been like diffusion models and all that is huge there.

But now these subscription-based individual creators do use bots to approximate themselves and chat with their whales. - People pay a lot for it. - And people pay a lot, right? A lot of times it's them, but a lot of, there are agencies that do this for these creators and do it on a mass scale.

So the largest creators are able to talk to hundreds or thousands of people at a time because of these bots. And so it's already being used there. Obviously, like video streaming and other technologies have gone there first, it's gonna come to the rest of society too. - There's a general concern that models get censored by the companies that deploy them.

So one case where we've seen that, and maybe censorship was one word alignment, maybe via RLHF or some other way is another word. So we saw that with black Nazi image generation with Gemini. As you mentioned, we also see that with Chinese models refusing to answer what happened in June 4th, 1989 at Tiananmen Square.

So how can this be avoided? And maybe can you just in general talk about how this happens and how can it be avoided? - You give multiple examples. There's probably a few things to keep in mind here. One is the kind of Tiananmen Square factual knowledge, like how does that get embedded into the models?

Two is the Gemini, what you call the black Nazi incident, which is when Gemini as a system had this extra thing put into it that dramatically changed the behavior. And then three is what most people would call general alignment, RLHF post-training. Each of these have very different scopes in how they are applied.

In order to do, if you're just to look at the model weights, in order to audit specific facts is extremely hard 'cause you have to Chrome through the pre-training data and look at all of this, and then that's terabytes of files and look for very specific words or hints of the words.

- So I guess one way to say it is that you can insert censorship or alignment at various stages in the pipeline. And what you refer to now is at the very beginning of the data selection. - So if you want to get rid of facts in a model, you have to do it at every stage.

You have to do it at the pre-training. So most people think that pre-training is where most of the knowledge is put into the model, and then you can elicit and move that in different ways, whether through post-training or whether through systems afterwards. - This is where the whole hacking models comes from.

GPT will not tell you how to make anthrax, but if you try really, really hard, you can eventually get it to tell you about anthrax because they didn't filter it from the pre-training data set, right? - But by the way, removing facts has such an ominous, dark feel to it.

- Almost think it's practically impossible 'cause you effectively have to remove them from the internet. You're taking on a-- - Well, did they remove the mm thing from the subreddits, the m-m-m-m-m? - It gets filtered out. - Right, so that's-- - Quality filters, which are small language models that look at a document and tell you, how good is this text?

Is it close to a Wikipedia article, which is a good thing that we want language models to be able to imitate? - So couldn't you do a small language model that filters out mentions of Tiananmen Square in the data? - Yes, but is it gonna catch wordplay or encoded language of the same thing?

- I mean, people have been meaning on games and other stuff how to say things that don't say Tiananmen Square, or yeah, so there's always different ways to do it. There's, hey, the internet as a whole does tend to just have a slight left bias because it's always been richer, more affluent, younger people on the internet relative to the rest of the population.

So there is already inherently a slight left bias on the internet. And so how do you filter things that are this complicated? And some of these can be factual, non-factual, but Tiananmen Square is obviously the example of a factual, but it gets a lot harder when you're talking about aligning to a ideal, right?

And so Grok, for example, right? Elon's tried really hard to make the model not be super PC and woke, but the best way to do pre-training is to throw the whole freaking internet at it, right? And then later figure out, but then at the end of the day, the model at its core now still has some of these ideals, right?

You still ingested Reddit/r/politics, which is probably the largest political discussion board on the world that's freely available to scrape. And guess what? That's left-leaning, right? And so, you know, there are some aspects like that you just can't censor unless you try really, really, really, really, really hard. - So the base model will always have some TDS, Traumatic Derangement Syndrome, because it's trained so much.

- It'll have the ability to express it. - But what if, what if you- (laughing) - There's a wide representation in the data. - This is what happens. It's like a lot of what is called post-training. It's a series of techniques to get the model on rails of a really specific behavior.

- And I mean, it's like you can, you also have the ingested data of like Twitter or like Reddit/r/thedonald, which is like also super pro-Trump, right? And then you have like fascist subreddits or like you have communist subreddits. The model in pre-training ingests everything. It has no worldview. Now it does have like some skew because more of the text is skewed a certain way, which is general, like slight left, like, but also like, you know, somewhat like, you know, intellectual, somewhat like, you know, it's just like the general internet is a certain way.

And then as Nathan's about to describe eloquently, right? Like you can elicit certain things out. - And there's a lot of history here. So we can go through multiple examples and what happened. Lama 2 was a launch that the phrase like too much RLHF or like too much safety was a lot.

It's just, that was the whole narrative after Lama 2's chat models released. And the examples are sorts of things like you would ask Lama 2 chat, how do you kill a Python process? And it would say, I can't talk about killing because that's a bad thing. And anyone that is trying to design an AI model will probably agree that that's just like, model you messed up a bit on the training there.

I don't think they meant to do this, but this was in the model weight. So this is not, it didn't necessarily be, there's things called system prompts, which are when you're querying a model, it's a piece of text that is shown to the model, but not to the user.

So a fun example is your system prompt could be talk like a pirate. So no matter what the user says to the model, it'll respond like a pirate. In practice, what they are is you are a helpful assistant. You should break down problems. If you don't know about something, don't tell them.

Your date cutoff is this, today's date is this. It's a lot of really useful context for how can you answer a question well. - And Anthropic publishes their system prompt. - Which I think is great. And there's a lot of research that goes into this. And one of your previous guests, Amanda Askell, is like probably the most knowledgeable person, at least in the combination of execution and sharing.

She's the person that should talk about system prompts and character of models. - Yeah, and then people should read the system prompts 'cause you're like trying to nudge sometimes to extreme politeness, the model to be a certain way. - And you could use this for bad things. And we've done tests, which is what if I tell the model to be a dumb model?

Like which evaluation scores go down? And it's like, we'll have this behavior where it could sometimes like say, oh, I'm supposed to be dumb. And sometimes it's like, it doesn't affect like math abilities as much, but something like, if you're trying, it's just the quality of a human judgment would drop to the floors.

Let's go back to post-training, specifically RLHF around Lama 2 was, it was too much safety prioritization was baked into the model weights. This makes you refuse things in a really annoying way for users. It's not great. It caused a lot of like awareness to be attached to RLHF that it makes the models dumb.

- And it stigmatized the word. - It did, in AI culture. And as the techniques have evolved, that's no longer the case where all of these labs have very fine-grained control over what they get out of the models through techniques like RLHF. - Although different labs are definitely different levels.

Like on one end of the spectrum is Google. And then like maybe OpenAI does less and Anthropic does less. And then like on the other end of the spectrum is like XAI, but they all have different forms of RLHF trying to make them a certain way. - And like the important thing to say is that no matter how you want the model to behave, these RLHF and preference tuning techniques also improve performance.

So on things like math evals and code evals, there is something innate to these, what is called contrastive loss functions. We could start to get into RL here. We don't really need to, but RLHF also boosts performance on anything from a chat task to a math problem to a code problem.

So it is becoming a much more useful tool to these labs. So this kind of takes us through the arc of we've talked about pre-training, hard to get rid of things. We've talked about post-training and how post-training, you can mess it up. It's a complex multifaceted optimization with 10 to 100 person teams converging on one artifact.

It's really easy to not do it perfectly. And then there's the third case, which is what we talked about Gemini. The thing that was about Gemini is this was a served product where Google has their internal model weights. They've done all these processes that we talked about. And in the served product, what came out after this was that they had a prompt that they were rewriting user queries to boost diversity or something.

And this just made it, the outputs were just blatantly wrong. It was some sort of organizational failure that had this prompt in that position. And I think Google executives probably have owned this. I didn't pay that attention to that detail, but it was just a mess up in execution that led to this ridiculous thing.

But at the system level, the model weights might have been fine. - So at the very end of the pipeline, there was a rewriting. - To something like a system prompt. It was like the system prompt or what is called an industry is like you rewrite prompts. So especially for image models, if you're using DALI or chat GPT can generate you an image, you'll say, draw me a beautiful car.

With these leading image models, they benefit from highly descriptive prompts. So what would happen is if you do that on chat GPT, a language model behind the scenes will rewrite the prompt, say, make this more descriptive. And then that is passed to the image model. So prompt rewriting is something that is used at multiple levels of industry and it's used effectively for image models.

And the Gemini example is just a failed execution. - Big philosophical question here with RLHF to generalize, where is human input, human in the loop, human data most useful at the current stage? - For the past few years, the highest cost human data has been in these preferences, which is comparing, I would say highest cost and highest total usage.

So a lot of money has gone to these pairwise comparisons where you have two model outputs and a human is comparing between the two of them. In earlier years, there was a lot of this instruction tuning data. So creating highly specific examples to something like a Reddit question to a domain that you care about.

Language models used to struggle on math and code. So you would pay experts in math and code to come up with questions and write detailed answers that were used to train the models. Now it is the case that there are many model options that are way better than humans at writing detailed and eloquent answers for things like model and code.

So they talked about this with the LLAMA3 release where they switched to using LLAMA3 405B to write their answers for math and code. But they, in their paper, talk about how they use extensive human preference data, which is something that they haven't gotten AIs to replace. There are other techniques in industry like constitutional AI, where you use human data for preferences and AI for preferences.

And I expect the AI part to scale faster than the human part. But among the research that we have access to is that humans are in this kind of preference loop. - So as reasoning becomes bigger and bigger and bigger, as we said, where's the role of humans in that?

- It's even less prevalent. So it's the remarkable thing about these reasoning results, and especially the DeepSeq R1 paper is this result that they call DeepSeq R1-0, which is they took one of these pre-trained models, they took DeepSeq V3 base, and then they do this reinforcement learning optimization on verifiable questions or verifiable rewards for a lot of questions and a lot of training.

And these reasoning behaviors emerge nastrally. So these things like, wait, let me see, wait, let me check this. Oh, that might be a mistake. And they emerge from only having questions and answers. And when you're using the model, the part that you look at is the completion. So in this case, all of that just emerges from this large-scale RL training.

And that model, which the weights are available, has no human preferences added into the post-training. There are, the DeepSeq R1 full model has some of this human preference tuning, this RLHF, after the reasoning stage. But the very remarkable thing is that you can get these reasoning behaviors, and it's very unlikely that there's humans writing out reasoning chains.

It's very unlikely that they somehow hacked OpenAI and they got access to OpenAI-01's reasoning chains. It's something about the pre-trained language models and this RL training where you reward the model for getting the question right. And therefore, it's trying multiple solutions, and it emerges this chain of thought. - This might be a good place to mention the eloquent and the insightful tweet of the great and the powerful Andrej Karpathy.

I think he had a bunch of thoughts, but one of them, last thought, not sure if this is obvious. You know something profound is coming when you're saying it's not sure if it's obvious. There are two major types of learning in both children and in deep learning. There's one, imitation learning, watch and repeat, i.e.

pre-training, supervised fine-tuning, and two, trial and error learning, reinforcement learning. My favorite simple example is AlphaGo. One is learning by imitating expert players. Two is reinforcement learning to win the game. Almost every single shocking result of deep learning and the source of all magic is always two. Two is significantly more powerful.

Two is what surprises you. Two is when the paddle learns to hit the ball behind the blocks and break out. Two is when AlphaGo beats even Lee Sedol. And two is the aha moment when the deep seek or O1, et cetera, discovers that it works well to reevaluate your assumptions, backtrack, try something else, et cetera.

It's the solving strategies you see this model use in its chain of thought. It's how it goes back and forth thinking to itself. These thoughts are emergent, three exclamation points. And this is actually seriously incredible, impressive, and new, and is publicly available and documented. The model could never learn this with imitation because the cognition of the model and the cognition of the human labeler is different.

The human would never know to correctly annotate these kinds of solving strategies and what they should even look like. They have to be discovered during reinforcement learning as empirically and statistically useful towards the final outcome. Anyway, the alpha zero sort of metaphor analogy here. Can you speak to that, the magic of the chain of thought that he's referring to?

- I think it's good to recap alpha go and alpha zero because it plays nicely with these analogies between imitation learning and learning from scratch. So alpha go, the beginning of the process was learning from humans where they started the first, this is the first expert level go player or chess player in DeepMind's series of models where they had some human data.

And then why it is called alpha zero is that there was zero human data in the loop. And that change to alpha zero made a model that was dramatically more powerful for DeepMind. So this remove of the human prior, the human inductive bias makes the final system far more powerful.

This we mentioned bitter lesson hours ago and this is all aligned with this. And then there's been a lot of discussion in language models. This is not new. This goes back to the whole Q star rumors, which if you piece together the pieces is probably the start of OpenAI figuring out it's 01 stuff.

When last year in November, the Q star rumors came out. There's a lot of intellectual drive to know when is something like this going to happen with language models? Because we know these models are so powerful and we know it has been so successful in the past. And it is a reasonable analogy that this new type of reinforcement learning training for reasoning models is when the door is open to this.

We don't yet have the equivalent of turn 37, which is the famous turn where the DeepMind's AI playing go stumped Lisa Doll completely. We don't have something that's that level of focal point, but that doesn't mean that the approach to technology is different and the impact of the general training is still incredibly new.

- What do you think that point would be? What would be move 37 for chain of thought for reasoning? - Scientific discovery. Like when you use this sort of reasoning problem and it's just something we fully don't expect. I think it's actually probably simpler than that. It's probably something related to computer user robotics rather than science discovery.

Because the important aspect here is models take so much data to learn, they're not sample efficient, right? Trillions, they take the entire web, right? Over 10 trillion tokens to train on, right? This would take a human thousands of years to read, right? A human does not, and humans know most of the stuff, a lot of the stuff models know better than it, right?

Humans are way, way, way more sample efficient. That is because of the self-play, right? How does a baby learn what its body is? As it sticks its foot in its mouth and it says, "Oh, this is my body," right? It sticks its hand in its mouth and it calibrates its touch on its fingers with the most sensitive touch thing on its tongue, right?

It's how babies learn. And it's just self-play over and over and over and over again. And now we have something that is similar to that, right? With these verifiable proofs, right? Whether it's a unit test in code or a mathematical verifiable task, generate many traces of reasoning, right? And keep branching them out, keep branching them out.

And then check at the end, hey, which one actually has the right answer? Most of them are wrong, great. These are the few that are right. Maybe we use some sort of reward model outside of this to select even the best one to preference as well. But now you've started to get better and better at these benchmarks.

And so you've seen over the last six months, a skyrocketing in a lot of different benchmarks, right? - All math and code benchmarks are pretty much solved except for frontier math, which is designed to be almost questions that aren't practical to most people. 'Cause they're exam level open math problem type things.

So it's like on the math problems that are somewhat reasonable, which is like somewhat complicated word problems or coding problems. That's just what Dylan is saying. - So the thing here is that these are only with verifiable tasks. We earlier showed an example of the really interesting, like what happens when chain of thought is to a non-verifiable thing.

It's just like a human chatting, right? With thinking about what's novel for humans, right? A unique thought. But this task and form of training only works when it's verifiable. And from here, the thought is, okay, we can continue to scale this current training method by increasing the number of verifiable tasks.

In math and coding, coding probably has a lot more to go. Math has a lot less to go in terms of what are verifiable things. Can I create a solver that then I generate trajectories toward or traces towards, reasoning traces towards, and then prune the ones that don't work and keep the ones that do work?

Well, those are gonna be solved pretty quickly, but even if you've solved math, you have not actually created intelligence, right? And so this is where I think the like aha moment of computer user robotics will come in because now you have a sandbox or a playground that is infinitely verifiable, right?

Did you, you know, messing around on the internet, there are so many actions that you can do that are verifiable. It'll start off with like, log into a website, create an account, click a button here, blah, blah, blah. But it'll then get to the point where it's, hey, go do a task on Tasker or whatever these other, all these various task websites.

Hey, go get hundreds of likes, right? And it's gonna fail. It's gonna spawn hundreds of accounts. It's gonna fail on most of them, but this one got to 1,000. Great, now you've reached the verifiable thing. And you just keep iterating this loop over and over. And that's when, and same with robotics, right?

That's where, you know, where you have an infinite playground of tasks, like, hey, did I put the ball in the bucket all the way to like, oh, did I like build a car, right? Like, you know, there's a whole trajectory to speed run or, you know, what models can do.

But at some point, I truly think that like, you know, we'll spawn models and initially all the training will be in sandboxes. But then at some point, you know, the language model pre-training is gonna be dwarfed by what is this reinforcement learning? You know, you'll pre-train a multimodal model that can see, that can read, that can write, you know, blah, blah, blah, whatever.

Vision, audio, et cetera. But then you'll have it play in a sandbox infinitely and figure out math, figure out code, figure out navigating the web, figure out operating a robot arm, right? And then it'll learn so much. And the aha moment I think will be when this is available to then create something that's not good, right?

Like, oh, cool, part of it was like figuring out how to use the web. Now, all of a sudden, it's figured out really well how to just get hundreds of thousands of followers that are real and real engagement on Twitter because all of a sudden this is one of the things that are verifiable.

- And maybe not just engagement, but make money. - Yes, of course. - I mean, that could be the thing where almost fully automated, it makes, you know, $10 million by being an influencer, selling a product, creating the product, like. And I'm not referring to like a hype product, but an actual product.

Or like, holy shit, this thing created a business. It's running it, it's the face of the business, that kind of thing. Or maybe a number one song, like it creates the whole infrastructure required to create the song to be the influencer that represents that song, that kind of thing.

It makes a lot of, that could be the move. I mean, our culture respects money in that kind of way. - And it's verifiable, right? - It's verifiable, right. - Bank account can't lie. - Exactly. - There's surprising evidence that once you set up the ways of collecting the verifiable domain that this can work.

There's been a lot of research before this R1 on math problems. And they approach math with language models just by increasing the number of samples. So you can just try again and again and again. And you look at the amount of times that the language models get it right.

And what we see is that even very bad models get it right sometimes. And the whole idea behind reinforcement learning is that you can learn from very sparse rewards. So it doesn't, the space of language and the space of tokens, whether you're generating language or tasks for a robot is so big that you might say that it's like, I mean, the tokenizer for a language model can be like 200,000 things.

So at each step, it can sample from that big of a space. So if it can generate a bit of a signal that it can climb onto, that's what the whole field of RL is around is learning from sparse rewards. And the same thing has played out in math where it's like very weak models that sometimes generate answers.

We see research already that you can boost their math scores. You can do this sort of RL training for math. It might not be as effective, but if you take a 1 billion parameter model, so something 600 times smaller than DeepSeq, you can boost its grade school math scores very directly with a small amount of this training.

So it's not to say that this is coming soon. Setting up the verification domains is extremely hard and there's a lot of nuance in this, but there are some basic things that we have seen before where it's like, it's at least expectable that there's a domain and there's a chance that this works.

- All right, so we have fun things happening in real time. This is a good opportunity to talk about other reasoning models, O1, O3. Just now OpenAI, as perhaps expected, released O3 mini. What are we expecting from the different flavors? Can you just lay out the different flavors of the O models and from Gemini, the reasoning model?

- Something I would say about these reasoning models is we talked a lot about reasoning training on math and code. And what is done is that you have the base model we've talked about a lot on the internet. You do this large-scale reasoning training with reinforcement learning. And then what the DeepSeq paper detailed in this R1 paper, which for me is one of the big open questions on how do you do this, is that they did reasoning heavy, but very standard post-training techniques after the large-scale reasoning RL.

So they did the same things with a form of instruction tuning through rejection sampling, which is essentially heavily filtered instruction tuning with some reward models. And then they did this RLHF, but they made it math heavy. So some of this transfer, we'd looked at this philosophical example early on.

One of the big open questions is how much does this transfer? If we bring in domains after the reasoning training, are all the models gonna become eloquent writers by reasoning? Is this philosophy stuff going to be open? We don't know in the research of how much this will transfer.

There's other things about how we can make soft verifiers and things like this. But there is more training after reasoning, which makes it easier to use these reasoning models. And that's what we're using right now. So if we're gonna talk about with 3Many and O1, these have gone through these extra techniques that are designed for human preferences after being trained to elicit reasoning.

- I think one of the things that people are ignoring is Google's Gemini flash thinking is both cheaper than R1 and better. And they released it in the beginning of December. - And nobody's talking about it. - No one cares. - It has a different flavor to it. Its behavior is less expressive than something like O1, and it has fewer tracks than it is on.

Quen released a model last fall, QWQ, which was their preview reasoning model. And DeepSea had R1 Lite last fall, where these models kind of felt like they're on rails, where they really, really only can do math and code. And O1 is, it can answer anything. It might not be perfect for some tasks, but it's flexible and has some richness to it.

And this is kind of the art of like how cook, like how is a model a little bit undercooked? It's like, I mean, it's good to get a model out the door, but it's hard to gauge. And it takes a lot of taste to be like, is this a full-fledged model?

Can I use this for everything? And they're probably more similar for math and code. My quick read is that Gemini flash is like not trained to the same way as O1, but taking an existing training stack, adding reasoning to it. So taking a more normal training stack and adding reasoning to it.

And I'm sure they're going to have more. I mean, they've done quick releases on Gemini flash, the reasoning, and this is the second version from the holidays. It's evolving fast and it takes longer to make this training stack where you're doing this large scale R1. - Ask it the same question from earlier.

The one about the- - The human nature. - Yeah. - What was the human nature one? - The way I can ramble, why I can ramble about this so much is that we've been working on this at AI2 before O1 was fully available to everyone and before R1, which is essentially using this RL training for fine tuning.

We use this in our Tulu series of models and you can elicit the same behaviors where you say like weight and so on, but it's so late in the training process that this kind of reasoning expression is much lighter. So you can, there's essentially a gradation and just how much of this RL training you put into it determines how the output looks.

- So we're now using Gemini 2.0 flash thinking experimental 121. - It summarized the prompt as humans self-domesticated apes. (laughs) - Perspective, okay. All right, so wait, is this revealing the reasoning? Here's why this is novel, okay. - You can click to expand. - Oh yeah, click to expand.

- Okay. Analyze the request, novel is the key word. - Like see how it just looks a little different? It looks like a normal output. - Yeah, it's, I mean, in some sense it's better structured, it makes more sense. - Oh, and it latched onto human and then it went into organisms and oh wow.

(laughs) - Apex predator, focus on domestication, apply domestication to humans, explore the idea of self-domestication. (laughs) - Not good, not good. - Where is this going? Refine and articulate the insight, greater facial expressiveness and communication ability, yes, plasticity and adaptability, yes, dependence on social groups, yes, all right. And self-critique and refine further, wow.

Is this truly novel? Is it well-supported? So on and so forth, and the insight it's getting at is humans are not just social animals, but profoundly self-domesticated apes, and this self-domestication is the key to understanding our unique cognitive and social abilities. Self-domesticated apes, self-domesticated. - I prefer the deep-seek response.

- Self-domesticated, I mean, it's novel, the insight is novel, I mean, that's like a good book title, self-domesticated apes, there could be a case made for that, I mean, yeah, it's cool. And it's revealing the reasoning, it's magical. It's magical, like, this is really powerful. - Hello, everyone, this is Lex with a quick intermission recorded after the podcast.

Since we reviewed responses from Deep Seeker One and Gemini Flash 2.0 Thinking during this conversation, I thought at this moment it would be nice to insert myself quickly doing the same for OpenAI 01 Pro and 03 Mini with the same prompt, the prompt being, give one truly novel insight about humans.

And I thought I would, in general, give my vibe check and vibe-based anecdotal report on my own experiences with the new 03 Mini model, now that I got a chance to spend many hours with it in different kinds of contexts and applications. So I would probably categorize this question as, let's say, open-ended philosophical question.

And in particular, the emphasis on novelty, I think is a nice way to test one of the capabilities of the model, which is come up with something that makes you pause and almost surprise you with its brilliance. So that said, my general review, after running each of the models on this question a bunch of times, is that 01 Pro consistently gave brilliant answers.

Ones that gave me pause and made me think, both cutting in its insight and just really nicely phrased with wit, with clarity, with nuance, over and over, consistently generating the best answers. After that is R1, which is less consistent, but again, delivered brilliance. Gemini Flash 2.0 Thinking was third.

And last was 03 Mini, actually. It often gave quite a generic answer, at least to my particular sensibilities. That said, in a bunch of other applications that I tested for brainstorming purposes, it actually worked extremely well and often outperformed R1. But on this open-ended philosophical question, it did consistently worse.

Now, another important element for each of these models is how the reasoning is presented. DeepSeek R1 shows the full chain of thought tokens, which I personally just love. For these open-ended philosophical questions, it's really, really interesting to see the model think through it. But really, also, just stepping back, me as a person who appreciates intelligence and reasoning and reflection, reading these kind of chain of thought raw tokens of R1, there's something genuinely beautiful about observing the path of deliberation in an intelligence system.

I think we don't always have that explicitly laid out for us humans. So to see it in another intelligence system, the non-linearity of it, akin to Ulysses or Finnegan's Wake by James Joyce, it's just beautiful to watch. Anyway, as we discussed in the episode, DeepSeek R1 talked about humans being able to convert selfish desires into cooperative systems by collectively pretending abstract rules like money, laws, and rights are real, and these shared hallucinations act as games, where competition is secretly redirected to benefit the group, turning conflict into society's fuel.

Gemini 2.0 Flash Thinking said, "Humans are not just social animals, "but self-domesticated apes, "and this self-domestication is the key "to understanding our unique cognitive "and social abilities." Now, it's important to say that the chain of thought there was really interesting. It was looking through the entire evolution of life on Earth, considering apex predators, and considering how, from that, we ended up to where we are.

I think that domestication by choice is a really interesting angle. Again, it's one of those things when somebody presents a different angle on a seemingly obvious thing, it just makes me smile. And the same with DeepSeek R1, that these hallucinations of money, laws, and rights, and us collectively pretending like it's real, and we play games with them that look like competition when secretly we're just cooperating with each other.

And that is the fuel of progress, beautifully put. Now, OpenAI R1 Pro consistently, over and over, delivered bangers. I can go through many of them, but the first one was, "Humans are the only species that turns raw materials "into symbolic resources, "then uses those symbols to reorganize "the very materials they came from, "creating a closed feedback loop "between meaning and matter." Here, I just ran it again.

Banger after banger, I'm telling you. "Humans are unique among known species "in that they simultaneously rewrite two layers of reality, "the external world, "and their own private mental landscapes, "and then merge these two rewritten layers "into a continuous personal narrative "that feels objectively true." Feels true. This is poetry.

Okay, and then O3 Mini High, for me, was smart, fast, actually, and kind of generic. Never quite got there for me. So here's the first one I got from O3 Mini. "Humans are not fixed beings, "but rather ongoing narratives, "dynamic stories that we continuously write, "edit, and reinterpret. "This narrative plasticity is more than just memory "or self-reflection.

"It's an intrinsic cognitive process "that acts like an internal error-correction system. "It allows us to adapt our identities and values over time "in response to new experiences, "challenges, and social contexts." Now, it almost sneaks up to something approximating cutting insight with "narrative plasticity" in quotes, but then it goes back to the generic.

I don't know. All of these models are incredible for different reasons. There's a lot of concerns, as we discussed in this episode, but there's a lot of reasons to be excited, as well. And I've probably spoken for too long. I am severely sleep-deprived, borderline delirious, so hopefully some of this made sense.

And now, dear friends, back to the episode. - I think when you, to Nathan's point, when you look at the reasoning models, to me, even when I used R1 versus O1, there was that rough edges-around-the-corner feeling. And FlashThinking earlier, I didn't use this version, but the one from December, and it definitely had that rough edges-around-the-corner feeling, where it's just not fleshed out in as many ways.

Sure, they added math and coding capabilities via these verifiers in RL, but it feels like they lost something in certain areas. And O1 is worse performing than Chat in many areas, as well, to be clear. - Not by a lot. - Not by a lot, though, right? And it's like, R1 definitely felt to me like it was worse than V3 in certain areas, like doing this RL expressed and learned a lot, but then it weakened in other areas.

And so I think that's one of the big differences between these models and what O1 offers. And then OpenAI has O1 Pro, and what they did with O3, which is also very unique, is that they stacked Search on top of Chain of Thought, right? And so Chain of Thought is one thing where it's able, it's one chain, it backtracks, goes back and forth, but how they solved the ArcAGI challenge was not just the Chain of Thought.

It was also sampling many times, i.e. running them in parallel, and then selecting. - Is running in parallel actually Search? 'Cause I don't know if we have the full information on how O1 Pro works. So I don't have enough information to confidently say that it is Search. - It is parallel samples.

- Yeah, and then what? - And it selects something. - And we don't know what the selection function is. The reason why we're debating is because since O1 was announced, there's been a lot of interest in techniques called Monte Carlo Tree Search, which is where you will break down the Chain of Thought into intermediate steps.

We haven't defined Chain of Thought. Chain of Thought is from a paper from years ago where you introduce the idea to ask a language model that at the time was much less easy to use. You would say, "Let's verify step-by-step," and it would induce the model to do this bulleted list of steps.

Chain of Thought is now almost a default in models, where if you ask it a math question, you don't need to tell it to think step-by-step. And the idea with Monte Carlo Tree Search is that you would take an intermediate point in that chain, do some sort of expansion, spend more compute, and then just select the right one.

That's like a very complex form of Search that has been used in things like Mu0 and AlphaZero. Potentially, I know Mu0 does this. - Another form of Search is just asking five different people and then taking the majority answer. - Yes. - There's a variety of, it could be complicated, it could be simple.

We don't know what it is, just that they are not just issuing one Chain of Thought in sequence. They're launching many in parallel. And in the Arc AGI, they launched 1,000 in parallel for the one that really shocked everyone, that beat the benchmark, was they would launch 1,000 in parallel, and then they would get the right answer like 80% of the time or 70% of the time, 90 maybe even.

Whereas if they just launched one, it was like 30%. - There are many extensions to this. I would say the simplest one is that our language models to date have been designed to give the right answer the highest percentage of the time in one response. And we are now opening the door to different ways of running inference on our models, in which we need to re-evaluate many parts of the training process, which normally opens the door to more progress, but we don't know if OpenAI changed a lot, or if just sampling more and multiple choice is what they're doing, or if it's something more complex, but they changed the training and they know that the inference mode is going to be different.

- So we're talking about O1 Pro, $200 a month, and they're losing money. So the thing that we're referring to, this fascinating exploration of the test time compute space, is that actually possible? Do we have enough compute for that? Does the financials make sense? - So the fantastic thing is, and it's in the thing that I pulled up earlier, but the cost for GPT-3 has plummeted.

If you scroll up just a few images, I think. The important thing about like, hey, is cost a limiting factor here, right? Like my view is that like, we'll have like really awesome intelligence before we have like AGI, before we have it permeate throughout the economy. And this is sort of why that reason is, right?

GPT-3 was trained in what, 2020, 2021? And the cost for running inference on it was $60, $70 per million tokens, right? Which is the cost per intelligence was ridiculous. Now, as we scaled forward two years, we've had a 1200X reduction in cost to achieve the same level of intelligence as GPT-3.

- So here on the X-axis is time over just a couple of years, and on the Y-axis is log scale dollars to run inference on a million tokens. - Yeah, a million. - And so you have just a down, like a linear decline on log scale from GPT-3 through 3.5 to LAMA.

- It's like 5 cents or something like that now, right? Which is versus $60, 1200X, that's not the exact numbers, but it's 1200X, I remember that number, is the humongous cost per intelligence, right? Now the freak out over DeepSeek is, oh my God, they made it so cheap. It's like, actually, if you look at this trend line, they're not below the trend line, first of all, and at least for GPT-3, right?

They are the first to hit it, which is a big deal, but they're not below the trend line as far as GPT-3. Now we have GPT-4, what's gonna happen with these reasoning capabilities, right? It's a mix of architectural innovations, it's a mix of better data, and it's gonna be better training techniques, and all of these different better inference systems, better hardware, right?

Going from each generation of GPU to new generations or ASICs, everything is gonna take this cost curve down and down and down and down. And then, can I just spawn a thousand different LLMs to create a task and then pick from one of them, or whatever search technique I want, a tree, Monte Carlo tree search, maybe it gets that complicated.

Maybe it doesn't, 'cause it's too complicated to actually scale, like who knows? Bitter lesson, right? The question is, I think, when, not if, because the rate of progress is so fast, right? Nine months ago, Dario said nine months ago the cost to train and inference was this, right? And now we're much better than this, right?

And DeepSeek is much better than this. And that cost curve for GPT-4, which was also roughly $60 per million tokens when it launched, has already fallen to $2 or so, right? And we're gonna get it down to cents, probably, for GPT-4 quality, and then that's the base for the reasoning models like O1 that we have today, and O1 Pro is spawning multiple, right?

And O3, and so on and so forth. These search techniques, too expensive today, but they will get cheaper. And that's what's gonna unlock the intelligence, right? - So it'll get cheaper and cheaper and cheaper. The big DeepSeek R1 release freaked everybody out because of the cheaper. One of the manifestations of that is Nvidia stock plummeted.

Can you explain what happened? I mean, and also just explain this moment and whether, you know, if Nvidia is gonna keep winning. - We're both Nvidia bulls here, I would say. And in some ways, the market response is reasonable. Most of the market, like Nvidia's biggest customers in the US are major tech companies, and they're spending a ton on AI.

And if a simple interpretation of DeepSeek is you can get really good models without spending as much on AI. So in that capacity, it's like, oh, maybe these big tech companies won't need to spend as much on AI and go down. The actual thing that happened is much more complex where there's social factors, where there's the rising in the app store, the social contagion that is happening.

And then I think a lot of some of it is just like, I'm not, I don't trade. I don't know anything about financial markets, but it builds up over the weekend or the social pressure where it's like, if it was during the week and there was multiple days of trading when this was really becoming, but it comes on the weekend and then everybody wants to sell.

And that is a social contagion. - I think, and like, there were a lot of false narratives, which is like, hey, these guys are spending billions on models, right? And they're not spending billions on models. No one's spent more than a billion dollars on a model that's released publicly, right?

GPT-4 was a couple hundred million, and then they've reduced the cost with 4.0, 4.0 Turbo, 4.0, right? But billion dollar model runs are coming, right? This concludes pre-training and post-training, right? And then the other number is like, hey, DeepSeek didn't include everything, right? They didn't include, a lot of the cost goes to research and all this sort of stuff.

A lot of the cost goes to inference. A lot of the cost goes to post-training. None of these things were factored. It's research salaries, right? Like all these things are like counted in the billions of dollars that OpenAI is spending, but they weren't counted in the, you know, hey, $6 million, $5 million that DeepSeek spent, right?

So there's a bit of misunderstanding of what these numbers are. And then there's also an element of, NVIDIA has just been a straight line up, right? And there's been so many different narratives that have been trying to push down NVIDIA. I don't say push down NVIDIA stock. Everyone is looking for a reason to sell or to be worried, right?

You know, it was Blackwell delays, right? Their GPU, you know, there's a lot of report. Every two weeks, there's a new report about their GPUs being delayed. There's the whole thing about scaling laws ending, right? It's so ironic, right? - It lasted a month. - It was just like literally just, hey, models aren't getting better, right?

They're just not getting better. There's no reason to spend more. Pre-training scaling is dead. And then it's like, oh one, oh three, right? - R1. - R1, right? And now it's like, wait, models are getting too, they're progressing too fast. Slow down the progress. Stop spending on GPUs, right?

But you know, the funniest thing I think that like comes out of this is Javon's paradox is true, right? AWS pricing for H100s has gone up over the last couple of weeks, right? Since a little bit after Christmas, since V3 was launched, AWS H100 pricing has gone up. H200s are like almost out of stock everywhere because you know, H200 has more memory and therefore R1 like, you know, wants that chip over H100, right?

- We were trying to get GPUs on a short notice this week for a demo and it wasn't that easy. We were trying to get just like 16 or 32 H100s for a demo and it was not very easy. - So for people who don't know, Javon's paradox is when, you know, the efficiency goes up, somehow magically, counterintuitively, the total resource consumption goes up as well.

- Right, and semiconductors is, you know, we're at 50 years of Moore's law. Every two years, half the cost, double the transistors. Just like clockwork, and it's slowed down obviously, but like the semiconductor industry has gone up the whole time, right? It's been wavy, right? There's obviously cycles and stuff and I don't expect AI to be any different, right?

There's gonna be ebbs and flows, but this is, in AI, it's just playing out at an insane timescale, right? It was 2X every two years. This is 1200X in like three years, right? So it's like the scale of improvement that is like hard to wrap your head around. - Yeah, I was confused because I, to me, Nvidia's stock on that should have gone up, but maybe it went down because there's kind of suspicion of foul play on the side of China or something like this.

But if you just look purely at the actual principles at play here, like it's obvious, yeah, the Gervant's Paradox. - The more progress that AI makes, or the higher the derivative of AI progress is, especially you should, 'cause Nvidia's in the best place. The higher the derivative is, the sooner the market's gonna be bigger and expanding and Nvidia's the only one that does everything reliably right now.

- 'Cause it's not like an Nvidia competitor arose. It's another company that's using Nvidia. - Who historically has been a large Nvidia customer. - Yeah. - And has press releases about them cheering about being China's biggest Nvidia customer, right? Like. - Yeah, I mean. - Obviously they've quieted down, but like, I think that's like another element of is that they don't wanna say how many GPUs they have.

- Yeah. - Because, hey, yes, they have H800s. Yes, they have H20s. They also have some H100s, right? Which are smuggled in. - Can you speak to that, to the smuggling? What's the scale of smuggling that's feasible for a nation state to do for companies? Is it possible to?

- I think there's a few angles of smuggling here, right? One is ByteDance arguably is the largest smuggler of GPUs for China, right? China's not supposed to have GPUs. ByteDance has like over 500,000 GPUs. Why? Because they're all rented from companies around the world. They rent from Oracle. They rent from Google.

They rent from all these mass and a bunch of smaller cloud companies too, right? All the Neo clouds, right? Of the world. They rent so, so many GPUs. They also buy a bunch, right? And they do this for mostly like what Meta does, right? Serving TikTok, right? Serving, next best, same as Meta, right?

To be clear, that's today the use, right? And it's a valid use, right? Hack the dopamine circuit, right? Now that's theoretically now very much restricted with the AI diffusion rules, which happened in the last week of the Biden admin and Trump admin looks like they're gonna keep them, which limits like allies, even like Singapore, which Singapore is like 20% of NVIDIA's 20, 20, 30% of NVIDIA's revenue.

But Singapore has had a memorandum on not building data centers for like 15 years 'cause they don't have enough power. So where are they going? (laughs) I mean, I'm not claiming they're all going to China, right? But a portion are, you know, many are going to Malaysia, including Microsoft and Oracle have big data centers in Malaysia.

Like, you know, they're going all over Southeast Asia, probably India as well, right? Like there's stuff routing, but like the diffusion rules are very de facto. Like you can only buy this many GPUs from this country and it's, and you can only rent a cluster this large to companies that are Chinese, right?

Like they're very explicit on trying to stop smuggling, right? And a big chunk of it was, hey, let's, you know, random company by 16 servers, ship some to China, right? There's actually, I saw a photo from someone in the semiconductor industry who leads like a team for like networking chips that competes with NVIDIA.

And he sent a photo of a guy checking into a first-class United flight from San Francisco to Shanghai or Shenzhen with a super micro box that was this big, which can only contain GPUs, right? And he was booking first-class 'cause think about it, three to 5K for your first-class ticket, server costs, you know, 240,000 in the US, 250,000.

You sell it for 300,000 in China. Wait, you just got a free first-class ticket and a lot more money. So it's like, you know, and that's like small-scale smuggling. Most of the large-scale smuggling is like companies in Singapore and Malaysia, like routing them around or renting GPUs. - I want to jump in.

How much was the scale? I think there's been some number, like some people that are higher level economics understanding say that as you go from 1 billion of smuggling to 10 billion, it's like you're hiding certain levels of economic activity. And that's the most reasonable thing to me is that there's going to be some level where it's so obvious that it's easier to find this economic activity.

- Yeah, so my belief is that last year roughly, so NVIDIA made a million H20s, which are legally allowed to be shipped to China, which we talked about is better for reasoning, right? Inference at least, maybe not training, but reasoning inference and inference generally. Then they also had, you know, a couple hundred thousand, we think like 200 to 300,000 GPUs were routed to China from, you know, Singapore, Malaysia, US, wherever.

Companies spawn up by 16 GPUs, 64 GPUs, whatever it is, route it. And Huawei is known for having spent up a massive network of like companies to get the materials they need after they were banned in like 2018. So it's not like otherworldly, but I agree, right? Nathan's point is like, hey, you can't smuggle up $10 billion of GPUs.

And then the third sort of source, which is just now banned and you know, which wasn't considered smuggling, but as China is renting, like is, I believe from our research, right? Oracle's biggest GPU customer is ByteDance, right? And for a Google, I think it's their second biggest customer, right?

And so like, and you go down the list of clouds and especially these smaller cloud companies that aren't like the hyperscalers, right? Think beyond Core, even Lambda, even there's a whole C, there's 60 different new cloud companies serving Nvidia GPUs. I think ByteDance is renting a lot of these, right?

All over, right? And so these companies are renting GPUs to Chinese companies and that's completely, that was completely legal up until the diffusion rules, which happened just a few weeks ago. And even now you can rent GPU clusters that are less than 2000 GPUs, or you can buy GPUs and ship them wherever you want if they're less than 1500 GPUs, right?

So it's like, there are still like some ways to smuggle, but yeah, it's not, you know, as the numbers grow, right? You know, a hundred something billion dollars of revenue for Nvidia last year, 200 something billion this year, right? And if next year are, you know, it could nearly double again or more than double, right?

Based on like what we see with data center footprints, like being built out all across the US and the rest of the world, it's gonna be really hard for China to keep up with these rules, right? Yes, there will always be smuggling and deep seek level models of GPT-4 level models, O1 level models capable to train on what China can get, even the next year above that.

But if we speed run a couple more, you know, jumps, right? You know, to billion dollar models, $10 billion models, then it becomes, you know, hey, there is a compute disadvantage for China for training models and serving them. And the serving part is really critical, right? Deep seek cannot serve their model today, right?

It's completely out of inventory. It's already started falling in the app store actually, downloads, because you download it, you try and sign up. They say, we're not taking registrations 'cause they have no capacity, right? You open it up, you get like less than five tokens per second, if you even get your request approved, right?

'Cause there's just no capacity because they just don't have enough GPUs to serve the model, even though it's incredibly efficient. - It'd be fascinating to watch the smuggling. 'Cause I mean, there's drug smuggling, right? That's a market, there's weapons smuggling, and GPUs will surpass that at some point. - Chips are highest value per kilogram, probably by far.

I have another question for you, Don. Do you track model API access internationally? How easy is it for Chinese companies to use hosted model APIs from the US? - Yeah, I mean, that's incredibly easy, right? Like OpenAI publicly stated DeepSeek uses their API. And as they say, they have evidence, right?

And this is another element of the training regime is people at OpenAI have claimed that it's a distilled model, i.e. you're taking OpenAI's model, you're generating a lot of output, and then you're training on the output in their model. And even if that's the case, what they did is still amazing, by the way, what DeepSeek did efficiency-wise.

- Distillation is standard practice in industry, whether or not, if you're at a closed lab where you care about terms of service and IP closely, you distill from your own models. If you are a researcher and you're not building any products, you distill from the OpenAI models. - This is a good opportunity.

Can you explain big picture distillation as a process? What is distillation? What's the process of distillation? - We've talked a lot about training language models. They are trained on text. In post-training, you're trying to train on very high-quality text that you want the model to match the features of, or if you're using RL, you're letting the model find its own thing.

But for supervised fine-tuning, for preference data, you need to have some completions of what the model is trying to learn to imitate. And what you do there is, instead of a human data, or instead of the model you're currently training, you take completions from a different, normally more powerful model.

I think there's rumors that these big models that people are waiting for, these GPT-5s of the world, the CLAWD-3 opuses of the world, are used internally to do this distillation process at OpenAI. - There's also public examples, right? Like Meta explicitly stated, not necessarily distilling, but they used 405(b) as a reward model for 70(b) in their LLAMA 3.2 and 3.3.

- This is all the same topic. - So is this ethical, is this legal? Like why is that Financial Times article headline say OpenAI says that there's evidence that China's DeepSeek used its model to train competitor? - This is a long, at least in the academic side and research side, it has a long history 'cause you're trying to interpret OpenAI's rule.

OpenAI's terms of service say that you cannot build a competitor with outputs from their model. Terms of service are different than a license, which are essentially a contract between organizations. So if you have a terms of service on OpenAI's account, if I violate it, OpenAI can cancel my account.

This is very different than like a license that says how you could use a downstream artifact. So a lot of it hinges on a word that is very unclear in the AI space, which is what is a competitor. - And then the ethical aspect of it is like, why is it unethical for me to train on your model when you can train on the internet's text?

- Yeah. - Right? - So there's a bit of a hypocrisy because OpenAI and potentially most of the companies trained on the internet's text without permission. - There's also a clear loophole, which is that I generate data from OpenAI and then I upload it somewhere and then somebody else trains on it and the link has been broken.

Like they're not under the same terms of service contract. - This is why-- - There's a lot of hypocrisy. There's a lot of link to be discovered details that don't make a lot of sense. - This is why a lot of models today, even if they train on zero OpenAI data, you ask the model who trained you, it'll say, "I am Chad GPT trained by OpenAI." Because there's so much copy paste of like OpenAI outputs from that on the internet that you just weren't able to filter it out.

And there was nothing in the RL where they implemented like hey, or post-training or SFT, whatever that says, "Hey, I'm actually a model by Allen Institute instead of OpenAI." - We have to do this. We serve a demo. We do research and we use OpenAI APIs because it's useful and we wanna understand post-training and like our research models, they will say they're written by OpenAI unless we put in the system prop that we talked about that like, "I am Tulu.

I am a language model trained by the Allen Institute for AI." And if you ask more people around industry, especially with post-training, it's a very doable task to make the model say who it is or to suppress the OpenAI thing. So in some levels, it might be that DeepSea didn't care that it was saying that it was by OpenAI.

Like if you're gonna upload model weights, it doesn't really matter 'cause anyone that's serving it in an application and cares a lot about serving is going to, when serving it, if they're using it for a specific task, they're gonna tailor it to that. And it doesn't matter that it's saying it's ChatGPT.

- Oh, I guess one of the ways to do that is like a system prompt or something like that. Like if you're serving it to say that you're- - That's what we do. Like if we host the demo, you say, "You are Tulu 3, a language model trained by the Allen Institute for AI.

We also are benefited from OpenAI data 'cause it's a great research tool." - I mean, do you think there's any truth and value to the claim, OpenAI's claim, that there's evidence that China's DeepSea used this model to train? - I think everyone has benefited regardless because the data's on the internet.

And therefore, it's in your pre-training now, right? There are like subreddits where people share the best ChatGPT outputs. And those are in your- - I think that they're trying to ship the narrative. Like they're trying to protect themselves. And we saw this years ago when ByteDance was actually banned from some OpenAI APIs for training on outputs.

There's other AI startups that most people, if you're in the like AI culture, were like, they just told us they trained on OpenAI outputs and they never got banned. Like that's how they bootstrapped their early models. So it's much easier to get off the ground using this than to set up human pipelines and build a strong model.

So there's long history here. And a lot of the communications seem like narrative control. - Actually, like over the last couple of days, we've seen a lot of people distill DeepSeq's model into LLAMA models because the DeepSeq models are kind of complicated to run inference on because they're a mixture of experts and they're 600 plus billion parameters and all this.

And people distilled them into the LLAMA models because the LLAMA models are so easy to serve and everyone's built the pipelines and tooling for inference with the LLAMA models, right? Because it's the open standard. So, you know, we've seen a sort of roundabout, right? Like, is it bad? Is it illegal?

Maybe it's illegal, whatever. I don't know about that, but like, it could break contracts. I don't think it's illegal, like in any legal, like no one's going to jail for this. - I think like fundamentally, I think it's ethical or I hope it's ethical because like the moment it becomes, we ban that kind of thing, it's gonna make everybody much worse off.

And I also actually, this is difficult, but I think you should be allowed to train on the internet. I know a lot of authors and creators are very sensitive about it. That's a difficult question. But the moment you're not allowed to train on the internet. - I agree. I have a schizo take on how you can solve this because it already works.

- I have a reasonable take on it. - Right, right. - All right, all right. - So, you know, Japan has a law, which you're allowed to train on any training data and copyrights don't apply if you want to train a model. A, B, Japan has nine gigawatts of curtailed nuclear power.

C, Japan is allowed under the AI diffusion rule to import as many GPUs as they'd like. So all we have to do, we have a market here to make, we build massive data centers, we rent them to the labs, and then we train models in a legally permissible way and there's no ifs, ands, or buts.

And now the models have no like potential copyright lawsuit from New York Times or anything like that. No, no, it's just like completely legal. No, so. - Genius. - The early copyright lawsuits have fallen in the favor of AI training. I would say that the long tail of use is gonna go in the side of AI, which is if you do, if you scrape trillions of data, you're not looking at, trillions of tokens of data, you're not looking and saying, this one New York Times article is so important to me.

But if you're doing a audio generation for music or image generation and you say, make it in the style of X person, that's a reasonable case where you could figure out what is their profit margin on inference. I don't know if it's gonna be the 50/50 of YouTube creator program or something, but I would opt into that program as a writer.

Like, please, like that, it's just, it's gonna be a rough journey, but there will be some solutions like that that make sense, but there's a long tail where it's just on the internet. - I think one of the other aspects of that Financial Times article implied, and so that leads to a more general question.

Do you think there's, how difficult is spying, espionage, and stealing of actual secret code and data from inside companies? How much of that is being attempted? - Code and data is hard, but ideas is easy. Silicon Valley operates on the way that top employees get bought out by other companies for a pay raise.

And a large reason why these companies do this is to bring ideas with them. And there are, there's no, I mean, in California, there's rules that like certain, like non-competes or whatever are illegal in California. And whether or not there's NDAs and things, that is how a lot of it happens.

Recently, there was somebody from Gemini who helped make this 1 million contacts length, and everyone is saying the next llama who, I mean, he went to the meta team, is gonna have 1 million contacts length. And that's kind of how the world works. - You know, as far as like industrial espionage and things, that has been greatly successful in the past, right?

You know, the Americans did it to the Brits, the Chinese have done it to the Americans, right? And you know, so on and so forth. It's just, it is a fact of life. And so like to argue industrial espionage can be stopped is probably unlikely, you can make it difficult.

But even then, like there's all these stories about like, hey, F-35 and F-22 have already been like, sort of like given to China in terms of design plans and stuff. Code and stuff, like between, you know, I say companies, not nation states is probably very difficult. But ideas are discussed a lot, right?

Whether it be a house party in San Francisco, or a company changing employees, or, you know, or the, you know, the always the like mythical honeypot that always gets talked about, right? Like someone gets honeypotted, right? Because everyone working on AI is a single dude who's in their 20s and 30s.

Not everyone, but like a insane amount of, insane percentages. So there's always like all these like, you know, and obviously-- - So honeypotted is like a spy, a female spy approaches you and like-- - Yeah, yeah, or male, right? You know, it's San Francisco, right? But as a single dude, I will say in his late 20s, right?

Is like, we are very easily corrupted, right? Like, you know, like not corrupted myself, but you know, like, we are, we are, right? - Everybody else, not me. - Yeah, exactly. - I'm too oblivious that I am not single. So I'm safe from one espionage access. - Yeah, you have to make sure to close all security vulnerabilities.

So you, Dylan, collect a lot of information about each of the mega clusters for each of the major AI companies. Can you talk about the build outs for each one that stand out? - Yeah, so I think the thing that's like really important about these mega cluster build outs is they're completely unprecedented in scale, right?

US, you know, sort of like data center power consumption has been slowly on the rise and it's gone up to two, 3%, even through the cloud computing revolution, right? Data center consumption as a percentage of total US. And that's been over decades, right? Of data centers, et cetera. It's been climbing, climbing slowly.

But now, two to 3%, now, by the end of this decade, it's like, even under like, you know, when I say like 10%, a lot of people that are traditionally, by like 2028, 2030, people traditionally non, traditional data center people, like that's nuts. But then like people who are in like AI, who have like really looked at this at like the Anthropics and open AIs, they're like, that's not enough.

And I'm like, okay, but like, you know, this is both through globally distributed or distributed throughout the US as well as like centralized clusters, right? The distributed throughout the US is exciting and it's the bulk of it, right? Like, hey, you know, open AI or, you know, say Meta's adding a gigawatt, right?

But most of it is distributed through the US for inference and all these other things, right? - So maybe we should lay out what a cluster is. So, you know, does this include AWS? Maybe it's good to talk about the different kinds of clusters and what you mean by mega clusters and what's a GPU and what's a computer and what is-- - Yeah, yeah, yeah.

- Not that far back, but yeah. So like, what do we mean by the clusters? - Oh man, I thought I was about to do the Apple ad, right? What's a computer? (laughing) So traditionally data centers and data center tasks have been a distributed systems problem that is capable of being spread very far and widely, right?

I.e. I send a request to Google, it's routed to a data center somewhat close to me. It does whatever search ranking recommendation, sends a result back, right? The nature of the task is changing rapidly in that the task, there's two tasks that people are really focused on now, right?

It's not database access. It's not serve me the right page, serve me the right ad. It's now A, inference. And inference is dramatically different from traditional distributed systems, but it looks a lot more similar. And then there's training, right? The inference side is still like, hey, I'm gonna put thousands of GPUs in blocks all around these data centers.

I'm gonna run models on them. User submits a request, it gets kicked off, or, hey, my service, they submit a request to my service, right? They're on Word and they're like, oh yeah, help me co-pilot. And it starts, kicks it off. I'm on my Windows, co-pilot, whatever. Apple intelligence, whatever it is, it gets kicked off to a data center, right?

And that data center does some work and sends it back. That's inference. That is going to be the bulk of compute. But then, and that's like, there's thousands of data centers that we're tracking with like satellites and like all these other things. And those are the bulk of what's being built, but the scale of, and so that's like what's really reshaping and that's what's getting millions of GPUs.

But the scale of the largest cluster is also really important, right? When we look back at history, right? Like, you know, or through the age of AI, right? Like it was a really big deal when they did AlexNet on, I think, two GPUs or four GPUs? I don't remember.

It was a really big deal. - It's a big deal 'cause you used GPUs. - It's a big deal to use GPUs. And they used multiple, right? But then over time, its scale has just been compounding. And so when you skip forward to GPT-3, then GPT-4, GPT-4, 20,000 A100 GPUs, unprecedented run, right?

In terms of the size and the cost, right? A couple hundred million dollars on a YOLO, right? A YOLO run for GPT-4. And it yielded, you know, this magical improvement that was like perfectly in line with what was experimented and just like a log scale, right? - Oh yeah, they have that plot from the paper.

The technical report. - The scaling laws were perfect, right? But that's not a crazy number, right? 20,000 A100s, roughly each GPU is consuming 400 watts. And then when you add in the whole server, right? Everything, it's like 15 to 20 megawatts of power, right? You know, maybe you could look up what the power of consumption of a human person is because the numbers are gonna get silly.

But like that 15 to 20 megawatts was standard data center size. It was just unprecedented that was all GPUs running one task. - 20 watts is a toaster. - A toaster is like also a similar power consumption to an A100, right? H100 comes around, they increase the power from like 400 to 700 watts, and that's just per GPU.

And then there's all the associated stuff around it. So once you count all that, it's roughly like 1200 to 1400 watts for everything, networking, CPUs, memory, blah, blah, blah. - So we should also say, so what's required? You said power, so a lot of power is required. A lot of heat is generated, cooling is required.

And because there's a lot of GPUs that have to be, or CPUs or whatever, they have to be connected. So there's a lot of networking. - Yeah, so I think, yeah, sorry for skipping past that. And then the data center itself is like complicated, right? But these are still standard sized data centers for GPT-4 scale, right?

Now we step forward to sort of what is the scale of clusters that people built last year, right? And it ranges widely, right? It ranges from like, hey, these are standard data centers, and we're just using multiple of them and connecting them together really with a ton of fiber between them, a lot of networking, et cetera.

That's what OpenAI and Microsoft did in Arizona, right? And so they have 100,000 GPUs, right? Meta, similar thing. They took their standard existing data center design, and it looks like an H, and they connected multiple of them together. And they first did 16,000 GPUs, 24,000 GPUs total. Only 16,000 of them were running on the training run because GPUs are very unreliable, so they need to have spares to swap in and out, all the way to like now 100,000 GPUs that they're training on Llama4 on currently, right?

Like 128,000 or so, right? This is, you know, think about 100,000 GPUs with roughly 1,400 watts a piece, that's 140 megawatts, 150 megawatts, right? For 128,000, right? So you're talking about, you've jumped from 15 to 20 megawatts to 10X, you know, almost 10X that number, 9X that number to 150 megawatts in two years, right?

From 2022 to 2024, right? And some people like Elon, he admittedly, right, and he says himself got into the game a little bit late for pre-training large language models, right? XAI was started later, right? But then he bent heaven and hell to get his data center up and get the largest cluster in the world, right?

Which is 200,000 GPUs. And he did that, he bought a factory in Memphis. He's upgrading the substation, but at the same time, he's got a bunch of mobile power generation, a bunch of single cycle combine. He tapped the natural gas line that's right next to the factory, and he's just pulling a ton of gas, burning gas.

He's generating all this power. He's in a factory, in an old appliance factory that's shut down and moved to China long ago, right? Like, you know, and he's got 200,000 GPUs in it. And now what's the next scale, right? Like all the hyperscalers have done this. Now the next scale is something that's even bigger, right?

And so, you know, Elon, just to stick on the topic, he's building his own natural gas plant, like a proper one right next door. He's deploying tons of Tesla mega pack batteries to make the power more smooth and all sorts of other things. He's got like industrial chillers to cool the water down because he's water cooling the chips.

So all these crazy things to get the clusters bigger and bigger. But when you look at like, say what OpenAI did with Stargate, that's that in Arizona, in Abilene, Texas, right? What they've announced at least, right? It's not built, right? Elon says they don't have the money. You know, there's some debates about this.

But at full scale, at least the first section is like definitely money's accounted for, but there's multiple sections. But full scale, that data center is gonna be 2.2 gigawatts, right? 2200 megawatts of power in, and roughly like 1.8 gigawatts or 1800 megawatts, yeah. 1800 megawatts of power delivered to chips, right?

Now this is an absurd scale. 2.2 gigawatts is like more than most cities, right? You know, to be clear. And delivered to a single cluster that's connected to do training, right? To train these models, to do both the pre-training, the post-training, all of this stuff, right? - This is insane.

- It is. - What is a nuclear power plant again? - Everyone is doing this, right? Everyone is doing this, right? Meta in Louisiana, right? They're building two natural gas plants, massive ones. And then they're building this massive data center. Amazon has like plans for this scale. Google has plans for this scale.

XAI has plans for this scale, right? Like all of these, the guys that are racing, the companies that are racing are racing hard. And they're doing multi-gigawatt data centers, right? To build this out because they think that, yeah. If I now have, you know, obviously pre-training scaling is gonna continue, but to some extent, but then also all this post-training stuff where you have an RL sandbox for computer use or whatever, right?

Like, you know, this is where they're gonna, and all these variable domains where they just keep learning and learning and learning self-play, whatever. Whatever it is, makes the AI so much more capable because the line does go up, right? As you throw more compute, you get more performance. The shirt is about scaling laws.

You know, to some extent it is diminishing returns, right? You 10X the compute. You don't get 10X better model, right? You get a diminishing returns, but also you get efficiency improvements. So you bend the curve, right? And these scale of data centers are doing, you know, wreaking, you know, a lot of like havoc on the network, right?

And, you know, Nathan was mentioning there's, Amazon has tried to buy this nuclear power plant, Talon. And if you look at Talon stock, it's just like skyrocketing. And, you know, like they're building a massive multi-gigawatt data center there. And, you know, you just go down the list. There's so many ramifications.

Interesting thing is like certain regions of the US transmitting power cost more than actually generating it, right? Because the grid is so slow to build and the demand for power and the ability to build power and like re-ramping on a natural gas plant or even a coal plant is like easy enough to do, but like transmitting the power is really hard.

So in some parts of the US, like in Virginia, it costs more to transmit power than it costs to generate it. Which is like, you know, there's all sorts of like second order effects that are insane here. - Can the power grid support this kind of growth? - You know, Trump's executive orders, there was a Biden executive order before the end of the year, but then Trump had some more executive orders, which hopefully reduced the regulations to where, yes, things can be built.

But yeah, this is a big, big challenge, right? Is building enough power fast enough? - Are you gonna basically have a nuclear power plant next to a data center for each one of these? - So the fun thing here is this is too slow to build the power plant.

To build a power plant or to reconfigure an existing power plant is too slow. And so therefore you must use natural, data center power consumption is flat, right? You know, I mean, like it's-- - Which is why nuclear is also good for it. Like long-term nuclear is a very natural fit, but you can't do solar or anything in the short term like that.

- 'Cause data center power is like this, right? Like you're telling me, you know, I'm gonna buy tens of billions of dollars of GPUs and idle them 'cause the power's not being generated? Like power is cheap, right? Like if you look at the cost of a cluster, less than 20% of it is power, right?

Most of it is the capital costs and depreciation of the GPUs, right? And so it's like, well, screw it. I'll just like, you know, I'll just build natural gas plants. This is what Meta is doing in Louisiana. This is what OpenAI is doing in Texas and like all these different places.

They may not be doing it directly, but they are partnered with someone. And so there is a couple hopes, right? Like one is, you know, in Elon, what he's doing in Memphis is like, you know, to the extreme, they're not just using dual combine cycle gas, which is like super efficient.

He's also just using single cycle and like mobile generators and stuff, which is less efficient. But he's, you know, there's also like the flip side, which is like solar power generation is like this and wind is another like, like this different correlate, you know, different. So if you stack both of those, plus you get a big chunk of batteries, plus you have a little bit of gas, it is possible to run it more green.

It's just the timescales for that is slow, right? So people are trying, but you know, Meta basically said, whatever, don't care about my sustainability pledge or they'll buy like a power, it's called a PPA, power purchasing agreement, where there'll be a massive wind farm or solar farm, like wherever.

And then they'll just pretend like those electrons are being consumed by the data center. But in reality, they're paying for the power here and selling it to the grid and they're buying power here. And then another thing is like Microsoft quit on some of their sustainability pledges, right? Elon, what he did with Memphis is objectively somewhat dirty, but he's also doing it in an area where there's like a bigger natural gas plant right next door and like a sewer next or not a sewer, but like a wastewater treatment and a garbage dump nearby, right?

And he's obviously made the world a lot more clean than that one data center is gonna do, right? So I think like, it's fine to some extent and maybe AGI solves global warming and stuff, right? Whatever it is, you know, this is sort of the attitude that people at the labs have, right?

Which is like, yeah, it's great, we'll just use gas, right? Because the race is that important and if we lose, you know, that's way worse, right? - I should say that I got a chance to visit the Memphis data center and it's kind of incredible. I mean, I visited with Elon, just the teams and the rate of innovation there is insane.

'Cause my sense is that, you know, nobody's ever done anything of this scale and nobody has certainly ever done anything of this scale at the rate that XAI is doing. So they're like figuring out, I mean, and so I was sitting in on all these meetings where they're brainstorming, it's like, it's insane.

It's exciting 'cause they're like, they're trying to figure out what the bottlenecks are, how to remove the bottlenecks, how to make sure that, you know, there's just so many really cool things about putting together a data center 'cause, you know, everything has to work. It's the people that do like the sysadmin, you know, the machine learning, all that is the exciting thing, so on.

But really the people that run everything are the folks that know like the low level software and hardware that runs everything, the networking, all of that. And so you have to like make sure you have procedures that test everything. I think they're using ethernet. I don't know how they're doing the networking, but.

- They're using NVIDIA Spectrum X Ethernet. There's actually like, I think, yeah, the unsung heroes are the cooling and electrical systems which are just like glossed over. - Yeah. - But I think like one story that maybe is like exemplifies how insane this stuff is, is when you're training, right?

You're always doing, you're running through the model a bunch, right, in the most simplistic terms, running through the model a bunch, and then you're gonna exchange everything and synchronize the weights, right? So you'll do a step. This is like a step in model training, right? And every step your loss goes down, hopefully.

And it doesn't always, but in the simplest terms, you'll be computing a lot and then you'll exchange, right? The interesting thing is GPU power is most of it. Networking power is some, but it's a lot less. But so while you're computing, your power for your GPUs is here. But then when you're exchanging weights, if you're not able to overlap communications and compute perfectly, there may be a time period where your GPUs are just idle and you're exchanging weights and you're like, hey, the model's updating.

So you're exchanging the gradients, you do the model update, and then you start training again. So the power goes, right? And it's super spiky. And so funnily enough, right? Like this, when you talk about the scale of data center power, right? You can blow stuff up so easily. And so Meta actually has accidentally open upstream something to code in PyTorch, where they added an operator.

And I kid you not, whoever made this, like, I wanna hug the guy because it says, says PyTorch, it's like PyTorch.powerplant, no blow up. Equals zero or equal one. And what it does, what it does is amazing, right? Either, you know, when you're exchanging the weights, the GPU will just compute fake numbers.

So the power doesn't spike too much. And so then the power plants don't blow up because the transient spikes, like screw stuff up. - Well, that makes sense. I mean, you have to do that kind of thing. You have to make sure they're not idle, yeah. - And Elon's solution was like, let me throw a bunch of Tesla mega packs and a few other things, right?

Everyone has different solutions, but like Meta's at least was publicly and openly known, which is just like, set this operator. And what this operator does is it just makes the GPUs compute nothing so that the power doesn't spike. - But that just tells you how much power you're working with.

I mean, it's insane. It's insane. - People should just go to Google, like scale, like what does X Watts do and go through all the scales from one watt to a kilowatt to a megawatt. And you look and stare at that and you're how high on the list a gigawatt is.

And it's mind blowing. - Can you say something about the cooling? So I know Elon's using liquid cooling, I believe in all cases. That's a new thing, right? Most of them don't use liquid cooling. Is there something interesting to say about the cooling? - Yeah, yeah. So air cooling has been the de facto standard, throw a bunch of metal, heat pipes, et cetera, and fans, right?

And like that's cool. That's been enough to cool it. People have been dabbling in water cooling. Google's TPUs are water cooled, right? So they've been doing that for a few years. But with GPUs, no one's ever done. And no one's ever done the scale of water cooling that Elon just did, right?

Now next generation NVIDIA is for the like highest end GPU, it is mandatory water cooling. You have to water cool it. But Elon did it on this current generation and that required a lot of stuff, right? If you look at like some of the satellite photos and stuff of the Memphis facility, there's all these external water chillers that are sitting basically.

It looks like a semi-truck pod thing. What's it called? The container. But really those are water chillers. And he has like 90 of those water chillers just sitting outside, 90 different containers, right? With the water, you know, like chill the water, bring it back to the data center, and then you distribute it to all the chips, pull all the heat out and then send it back, right?

And this is both a way to cool the chips, but also as an efficiency thing, all right? And going back to that like sort of three vector thing, right, there is memory bandwidth flops and interconnect. The closer the chips are together, the easier it is to do high speed interconnects, right?

And so this is also like a reason why you're gonna go water cooling is because you can just put the chips right next to each other and therefore get higher speed connectivity. - I gotta ask you, so in one of your recent posts, there's a section called Cluster Measuring Contest.

So-- - There's another word there, but I won't say it, you know? (laughs) - Who's got the biggest now and who's gonna have the biggest? - Today, individual largest is Elon, right? - Elon's Cluster. - Elon's Cluster in Memphis, 200,000 GPUs, right? Meta has like 128,000, OpenAI has 100,000.

Now, to be clear, other companies have more GPUs than Elon, they just don't have them in one place, right? And for training, you want them tightly connected. There's some techniques that people are researching and working on that lets you train across multiple regions, but for the most part, you want them all in like one area, right?

So you can connect them with high-speed networking. And so, you know, Elon today has 200,000 H100s, and 100,000 H100s, 100,000 H200s, right? Meta, OpenAI, you know, and Amazon all have on the scale of 100,000, a little bit less. But this year, right, this year, people are building much more, right?

Anthropic and Amazon are building a cluster of 400,000 Tranium II, which is Amazon-specific chip, trying to get away from NVIDIA, right? You know, Meta and OpenAI have scales for hundreds of thousands. But by next year, you'll have like 500,000 to 700,000 GPU clusters. And note, those GPUs are much higher power consumption than existing ones, right?

Hopper's 700 watts, Blackwell goes to 1200 watts, right? So the power per chip is growing, and the number of chips is growing, right? - Nuts, you think Elon said he'll get to a million. You think that's actually feasible? - I mean, I don't doubt Elon, right? The filings that he has for like, you know, the power plan and the Tesla battery packs, it's clear he has some crazy plans for Memphis.

Like permits and stuff is open record, right? But it's not quite clear that, you know, what and what the timescales are. I just never doubt Elon, right? You know, he's gonna surprise us. - So what's the idea with these clusters? If you have a million GPUs, what percentage in let's say two, three years is used for training and what percent, pre-training, and what percent is used for like, for the actual computation?

- So these mega clusters make no sense for inference, right? You could route inference there and just not train. But most of the inference capacity is being, you know, "Hey, I've got a 30 megawatt data center here. "I've got 50 megawatts here. "I've got a hundred here, whatever. "I'll just throw inference in all of those." Because the mega clusters, right?

Multi gigawatt data centers, I wanna train there. Because that's where all of my GPUs are co-located, where I can put them at a super high networking speed connected together, right? Because that's what you need for training. Now, with pre-training, this is the old scale, right? You could, you would increase parameters.

You didn't increase data, model gets better. That doesn't apply anymore, because there's not much more data in the pre-training side, right? Yes, there's video and audio and image that has not been fully taken advantage of. So there's a lot more scaling, but a lot of people like, have transcript, taken transcripts of YouTube videos.

And that gets you a lot of the data. It doesn't get you all of the learning value out of the video and image data. But there's still scaling to be done on pre-training. But this post-training world is where all the flops are gonna be spent, right? The model is gonna play with itself.

It's gonna self-play. It's gonna do verifiable tasks. It's gonna do computer use in sandboxes. It might even do like simulated robotics things, right? Like all of these things are gonna be environments where compute is spent in quote-unquote post-training. But I think it's gonna be good. We're gonna drop the post from post-training.

- Yeah, wow. - It's gonna be pre-training and it's gonna be training, I think. - The return of the king. - At some point. Because for the bulk of the last few years, pre-training has dwarfed post-training. But with these verifiable methods, especially ones that scale really potentially infinitely, like computer use and robotics, not just math and coding, right?

Where you can verify what's happening. Those infinitely verifiable tasks, it seems you can spend as much compute as you want on them. - Especially at the context length increase. 'Cause at the end of pre-training is when you increase the context length for these models. And we've talked earlier in the conversation about how the context length, when you have a long input, is much easier to manage than output.

And a lot of these post-training and reasoning techniques rely on a ton of sampling and it's becoming increasingly long context. So it's just like you're, effectively your compute efficiency goes down. I don't, I think FLOPS is the standard for how you measure it. But with RL and you have to do all these things where you move your weights around in a different way than at pre-training and just generation, it's going to become less efficient and FLOPS is gonna be less of a useful term.

And then as the infrastructure gets better, it's probably gonna go back to FLOPS. - So all of the things we've been talking about is most likely going to be NVIDIA, right? Is there any competitors? - Google, I kind of ignored them. - Yeah, what's the story with TPU? What's the story with TPU?

Like what's the- - TPU is awesome, right? It's great. Google is, they're a bit more tepid on building data centers for some reason. They're building big data centers, don't get me wrong. And they have, they actually have the biggest cluster. Let me, I was talking about NVIDIA clusters. They actually have the biggest cluster, period.

But the way they do it is like very interesting, right? They have two sort of like data center super regions, right? In that the data center isn't physically, like all of the GPUs aren't physically on one site, but they're like 30 miles from each other. They're not GPUs, TPUs, right?

They have like in Iowa and Nebraska, they have four data centers that are just like right next to each other. - Why doesn't Google flex its cluster size? - Go to multi data center training. It's a good images in there. So I'll show you what I mean. It's just a semi-analysis multi data center.

So this is like, you know, so this is an image of like what a standard Google data center looks like. By the way, their data centers look very different than anyone else's data centers. - What are we looking at here? - So these are, yeah. So if you see this image, right?

In the center, there are these big rectangular boxes, right? Those are where the actual chips are kept. And then if you scroll down a little bit further, you can see there's like these water pipes, there's these chiller cooling towers in the top and a bunch of like diesel generators.

The diesel generators are backup power. The data center itself is like, look physically smaller than the water chillers, right? So the chips are actually easier to like keep together, but then like cooling all the water for the water cooling is very difficult, right? So Google has like a very advanced infrastructure that no one else has for the TPU.

And what they do is they've like stamped these data center, they've stamped a bunch of these data centers out in a few regions, right? So if you go a little bit further down, this is a Microsoft, this is in Arizona, this is where GPT-5 quote unquote will be trained.

- If it doesn't exist already. - Yeah, if it doesn't exist already. But each of these data centers, right? I've shown a couple of images of them. They're like really closely co-located in the same region, right? Nebraska, Iowa. And then they also have a similar one in Ohio complex, right?

And so these data centers are really close to each other. And what they've done is they've connected them super high bandwidth with fiber. And so these are just a bunch of data centers. And the point here is that Google has a very advanced infrastructure, very tightly connected in a small region.

So Elon will always have the biggest cluster fully connected, right? Because it's all in one building, right? And he's completely right on that, right? Google has the biggest cluster, but you have to spread over three sites and by a significant margin, we have to go across multiple sites. - Why doesn't Google compete with Nvidia?

Why don't they sell TPUs? - I think there's a couple problems with it. It's like one, TPU has been a form of allowing search to be really fricking cheap and build models for that, right? And so like a big chunk of the search GPU purchases or TPU purchases are big chunk of Google's purchases and usage, all of it is for internal workloads, right?

Whether it be search, now Gemini, right? YouTube, all these different applications that they have, you know, ads. These are where all their TPUs are being spent and that's what they're hyper-focused on, right? And so there's certain like aspects of the architecture that are optimized for their use case that are not optimized elsewhere, right?

One simple one is like they've open-sourced the Gemma model and they called it Gemma 7B, right? But then it's actually 8 billion parameters because the vocabulary is so large. And the reason they made the vocabulary so large is because TPUs like matrix multiply unit is massive because that's what they've like sort of optimized for.

And so they decided, oh, well, I'll just make the vocabulary large too, even though it makes no sense to do so in such a small model because that fits on their hardware. So Gemma doesn't run as efficiently on a GPU as a Llama does, right? But vice versa, Llama doesn't run as efficiently on a TPU as a Gemma does, right?

And it's so like, there's like certain like aspects of like hardware software co-design. So all their search models are their ranking and recommendation models, all these different models that are AI, but not like Gen AI, right? Have been hyper-optimized with TPUs forever. The software stack is super optimized, but all of this software stack has not been released publicly at all, right?

Very small portions of it, JAX and XLA have been, but like the experience when you're inside of Google and you're training on TPUs as a researcher, you don't need to know anything about the hardware in many cases, right? Like, it's like pretty beautiful. But as soon as you step outside- - They all love it.

A lot of them go back. They leave Google and then they go back. - Yeah. - Yeah, they're like, they leave and they start a company 'cause they have all these amazing research ideas and they're like, wait, infrastructure's hard. Software is hard. And this is on GPUs. Or if they try to use TPUs, same thing, 'cause they don't have access to all this code.

And so it's like, how do you convince a company whose golden goose is search, where they're making hundreds of billions of dollars from, to start selling GPU or TPUs, which they used to only buy a couple billion of, you know. I think in 2023, they bought like a couple billion and now they're buying like 10 billion to $15 billion worth.

But how do you convince them that they should just buy like twice as many and figure out how to sell them and make $30 billion? It's like, who cares about making $30 billion? - Won't that 30 billion exceed, actually, the search profit eventually? - Oh, I mean, like, you're always gonna make more money on services than-- - Always.

- I mean, like, yeah. To be clear, like today, people are spending a lot more on hardware than they are the services, right? Because the hardware front runs the service spend. But like-- - You're investing, yeah. - If there's no revenue for AI stuff or not enough revenue, then obviously, like, it's gonna blow up, right?

People won't continue to spend on GPUs forever. And then NVIDIA is trying to move up the stack with like software that they're trying to sell and license and stuff, right? But Google has never had that like DNA of like, this is a product we should sell, right? They don't, the Google Cloud does it, which is a separate organization from the TPU team, which is a separate organization from the DeepMind team, which is a separate organization from the search team, right, there's a lot of bureaucracy.

- Wait, Google Cloud is a separate team than the TPU team? - Technically, TPU sits under infrastructure, which sits under Google Cloud. But like Google Cloud, like for like renting stuff and TPU architecture are very different goals, right? In hardware and software, like all of this, right? Like the JAXX XLA teams do not serve Google's customers externally, whereas NVIDIA's various CUDA teams for like things like Nickel serve external customers, right?

The internal teams like JAXX and XLA and stuff, they more so serve DeepMind and search, right? And so their customer is different, they're not building a product for them. - Do you understand why AWS keeps winning versus Azure for cloud versus Google Cloud? - Yeah, there's-- - Google Cloud is tiny, isn't it, relative to AWS?

- Google Cloud is third. Yeah, yeah. Microsoft is the second biggest, but Amazon is the biggest, right? And Microsoft deceptively sort of includes like Microsoft Office 365 and things like that, like some of these enterprise-wide licenses. So in reality, the gulf is even larger. Microsoft is still second though, right?

Amazon is way bigger, why? Because using AWS is better and easier. And in many cases-- - It was first. - And it's first, yeah. - It was first. - Yeah, but there's a lot of things that are first that-- - Well, it's easier, it's harder to switch than it is to-- - Yeah, okay.

- But AWS-- - Because it's large-- - There's big fees for switching too. - AWS generates over 80% of Amazon's profit, I think over 90%, right? - That's insane. - The distribution centers are just like, one day we'll decide to make money from this. But they haven't yet, right?

Like they make tiny little profit from it. - Yeah, one day Amazon Prime will triple in price. - You would think they would improve AWS interface 'cause it's like horrible, it's like clunky. - I have no idea. - But everybody is-- - Yeah, one would think. - I think actually Google's interface is sometimes nice, but it's also like they don't care about anyone besides their top customers.

- Yeah, exactly. - And like their customer service sucks, and like they have a lot less like-- - I mean, all of these companies, they optimize for the big customers, yeah. It's supposed to be for business. - Well, Amazon has always optimized for the small customer too though, right?

Like obviously they optimize a lot for the big customer, but like when they started, they just would go to like random Bay Area things and give out credits, right? And then they like, or just put in your credit card and use us, right? Like it's back in the early days.

So they've always, the business has grown with them, right, and burgeoned. So like why does Amazon, like why is Snowflake all over Amazon? Because Snowflake in the beginning when Amazon didn't care about them was still using Amazon, right? And then of course one day Snowflake and Amazon has a super huge partnership.

But like this is the case, like Amazon's user experience and quality is better. Also a lot of the silicon they've engineered makes them have a lower cost structure than traditional cloud storage, CPU, networking, that kind of stuff. Then in databases, right, like, you know, I think like four of Amazon's top five revenue products, margin products, sorry, like gross profit products are all database related products like Redshift and like all these things, right?

Like, so Amazon has a very like good silicon to a user experience, like entire pipeline with AWS. I think Google, their silicon teams, yeah, they have awesome silicon internally, TPU, the YouTube chip, you know, some of these other chips that they've made. And the problem is they're not serving external customers or serving internal customers, right?

- I mean, NVIDIA's entire culture is designed from the bottom up to do this. There's this recent book, "The NVIDIA Way" by Tay Kim that details this and how they look for future opportunities and ready their CUDA software libraries to make it so that new applications of high-performance computing can very rapidly be evolved on CUDA and NVIDIA chips.

And that is entirely different than Google as a services business. - Yeah, I mean, NVIDIA, it should be said is a truly special company. Like, I mean, they, the whole, the culture, everything, they're really optimized for that kind of thing. Speaking of which, is there somebody that can even challenge NVIDIA hardware-wise, Intel, AMD?

- I really don't think so. We went through a like a very long process of working with AMD on training on their GPUs, inference and stuff. And they're decent. Their hardware is better in many ways than NVIDIA's. The problem is their software is really bad. And I think they're getting better, right?

They're getting better faster, but they're just, the gulf is so large. And like, they don't spend enough resources on it or haven't historically, right? Maybe they're changing their tune now, but you know, for multiple months, we were submitting the most bugs, right? Like us, semi-analysis, right? Like, what the fuck?

Why are we submitting the most bugs, right? 'Cause they only cared about their like biggest customers. And so they'd ship them a private image, blah, blah, blah. And it's like, okay, but like, I am just using PyTorch and I wanna use the publicly available libraries. You don't care about that, right?

So they're getting better. But like, I think AMD is not possible. Intel's obviously in dire straits right now and needs to be saved somehow. Very important for national security, for American technology dominance. - Can you explain the obviously? So why are they in dire straits? - Going back to earlier, only three companies can R&D, right?

Taiwan, Shenzhou, Samsung, Pyongyang, and then Intel Hillsboro. Samsung's doing horribly, Intel's doing horribly. We could be in a world where there's only one company that can do R&D. And that one company already manufactures most of the chips. They've been gaining market share anyways. But like, that's a critical thing, right?

So what happens to Taiwan means the rest of the world's semiconductor industry and therefore tech relies on Taiwan, right? And that's obviously precarious. As far as like Intel, they've been slowly steadily declining. They were on top of servers and PCs, but now Apple's done the M1 and Nvidia's releasing a PC chip and Qualcomm's releasing a PC chip.

And in servers, hyperscalers are all making their own ARM-based server chips. And Intel has no AI silicon like wins, right? They have very small wins. And they never got into mobile because they said no to the iPhone. And like, all these things have compounded and they've lost their process technology leadership, right?

They were ahead for 20 years and now they're behind by at least a couple of years, right? And they're trying to catch back up and we'll see if like their 18A, 14A strategy works out where they try and leapfrog TSMC. But like, and Intel is just like losing tons of money anyways, right?

And they just fired their CEO, even though the CEO was the only person who understood the company well, right? We'll see. He was not the best, but he was pretty good relatively, technical guy. - Where does Intel make most of its money? The CPUs still, right? - PCs and data center CPUs, yeah.

But data center CPUs are all going cloud and Amazon, Microsoft, Google are making ARM-based CPUs. And then PC side, AMD's gained market share, Nvidia's launching a chip. That's not gonna be a success, right? MediaTek, Qualcomm have relaunched chips. Apple's doing well, right? Like, they could get squeezed a little bit in PC.

Although PC generally, I imagine, will just stick Intel mostly for Windows side. - Let's talk about the broad AI race. Who do you think wins? We talked about Google. - The leader, the default leader has been Google because of their infrastructure advantage. - Well, like, in the news, OpenAI is the leader.

- They're the leading in the narrative. - They have the best model. - They have the best model that people can use and they're experts. - And they have the most AI revenue. - Yeah, OpenAI is winning. - So, who's making money on AI right now? Is anyone making money?

- So, accounting profit-wise, Microsoft is making money, but they're spending a lot of CapEx, right? You know, and that gets depreciated over years. Meta's making tons of money, but with recommendation systems, which is AI, but not with Llama, right? Llama's losing money for sure, right? I think Anthropic and OpenAI are obviously not making money 'cause otherwise they wouldn't be raising money, right?

They'd have to raise money to build more, right? Although, theoretically, they are making money, right? Like, you know, you spent a few hundred million dollars on GPT-4 and it's doing billions in revenue. So, like, obviously it's like making money. Although they had to continue to research to get the compute efficiency wins, right?

And move down the curve to like, you know, that 12, get that 1200X that has been achieved for GPT-3. You know, maybe we're only at like a couple hundred X now, but, you know, with GPT-4 Turbo and 4.0, and there'll be another one probably cheaper than GPT-4.0 even that comes out at some point.

- And that research costs a lot of money. - Yep, exactly. - That's the thing that I guess is not talked about with the cost, that when you're referring to the cost of the model, it's not just the training or the test runs, it's the actual research, the manpower.

- Yeah, to do things like reasoning, right? Now that that exists, they're gonna scale it, they're gonna do a lot of research. So, I think the, you know, people focus on the payback question, but it's really easy to like, just be like, well, like, you know, GDP is humans and industrial capital, right?

And if you can make intelligence cheap, then you can grow a lot, right? That's the sort of dumb way to explain it. But that's sort of what basically the investment thesis is. I think only NVIDIA is actually making tons of money and other hardware vendors. The hyperscalers are all on paper making money, but in reality, they're like, spending a lot more on purchasing the GPUs, which you don't know if they're still gonna make this much money on each GPU in two years, right?

You don't know if, you know, all of a sudden, OpenAI goes kapoof and now Microsoft has like, hundreds of thousands of GPUs they were renting to OpenAI that they paid for themselves with their, you know, investment in them. You know, that no longer have a customer, right? Like, this is always a possibility.

I don't believe that, right? I think, you know, OpenAI will keep raising money. I think others will keep raising money because the investments, the returns from it are gonna be eventually huge once we have AGI. - So do you think multiple companies will get, let's assume-- - I don't think it's winner take all.

- Okay, so it's not, let's not call it AGI, whatever. It's like a single day. It's a gradual thing. - Powerful AI. Super powerful AI. - But it's a gradually increasing set of features that are useful and make a lot of money. - Rapidly increasing set of features. - Rapidly increasing set of features.

So you're saying a lot of companies will be, it just seems absurd that all of these companies are building gigantic data centers. - There are companies that will benefit from AI, but not because they trained the best model. Like, Meta has so many avenues to benefit from AI in all of their services.

People are there, people spend time on Meta's platforms, and it's a way to make more money per user per hour. - Yeah, it seems like Google X/XAI/Tesla, important to say, and then Meta will benefit not directly from the AI, like the LLMs, but from the intelligence, like the additional boost of intelligence to the products they already sell.

So whether that's the recommendation system or for Elon, who's been talking about Optimus, the robot, potentially the intelligence of the robot. And then you have personalized robots in the home, that kind of thing. He thinks it's a 10 plus trillion dollar business, which-- - At some point, maybe, not soon, but who knows what robotics will be used for.

- Let's do a TAM analysis, right? Eight billion humans, and let's get eight billion robots, right, and let's pay 'em the average salary, and yeah, there we go, 10 trillion. More than 10 trillion. - Yeah, I mean, if there's robots everywhere, why does it have to be just eight billion robots?

- Yeah, yeah, of course, of course. I'm gonna have one robot, you're gonna have 20. - Yeah, I mean, I see a use case for that. So yeah, so I guess the benefit would be in the products they sell, which is why OpenAI is in a trickier position, 'cause they-- - All of the value of OpenAI right now as a brand is in ChatGPT, and there is actually not that, for most users, there's not that much of a reason that they need OpenAI to be spending billions and billions of dollars on the next best model, when they can just license Llama 5 for it to be way cheaper.

So that's kind of like, ChatGPT is an extremely valuable entity to them, but they could make more money just off that than-- - The chat application is clearly like, does not have tons of room to continue, right? Like the standard chat, right, where you're just using it for random questions and stuff, right?

The cost continues to collapse, v3 is the latest one-- - It'll go down to ads. - Biggest, but it's gonna get supported by ads, right? Like, you know, Llama, Meta already serves 405B, probably loses the money, but at some point, you know, they're going to get, the models are gonna get so cheap that they can just serve them for free with ads supported, right?

And that's what Google is gonna be able to do, and that's obviously, they've got a bigger reach, right? So chat is not gonna be the only use case, it's like these reasoning, code, agents, computer use, all this stuff is where OpenAI has to actually go to make money in the future, otherwise they're kaputs.

- But X, Google, and Meta have these other products, so isn't it likely that OpenAI and Anthropic disappear eventually? - Unless they're so good at models, which they are. - But it's such a cutting edge, I mean-- - It depends on where you think AI capabilities are going. - You have to keep winning.

- Yes. - You have to keep winning, as you climb, even if the AI capabilities are going super rapidly, awesome, into the direction of AGI, like, there's still a boost for X in terms of data, Google in terms of data, Meta in terms of data, in terms of other products, and the money, and there's just huge amounts of money.

- But the whole idea is, human data is kinda tapped out, we don't care, we all care about self-play, verifiable tasks. - Yeah, so self-play-- - If you think about AWS-- - Which is an R and D problem, yeah. - I think AWS does not make a lot of money on each individual machine, and the same can be said for the most powerful AI platform, which is, even though the calls to the API are so cheap, there's still a lot of money to be made by owning that platform, and there's a lot of discussions as it's the next compute layer.

- You have to believe that, and yeah, there's a lot of discussions that tokens, and tokenomics, and LLM APIs are the next compute layer, or the next paradigm for the economy, kind of like energy and oil was, but there's also, you have to sort of believe that APIs and chat are not where AI is stuck, right?

It is actually just tasks, and agents, and robotics, and computer use, and those are the areas where all the value will be delivered, not API, not chat application, right? - Is it possible you have, I mean, it all just becomes a commodity, and you have the very thin wrapper, like perplexity, just joking.

- There are a lot of wrappers making a lot of money. - Yeah, so, but do you think it's possible that people would just even forget what OpenAI and Anthropic is, and just, 'cause there'll be wrappers around the API, and it just dynamically-- - If model progress is not rapid, yeah, it's becoming a commodity, right?

Deep Seek v3 shows this, but also the GPT-3 chart earlier, Kurt chart showed this, right? Lama 3B is 1,200x cheaper than GPT-3. Any GPT-3, like anyone whose business model was GPT-3 level capabilities is dead. Anyone whose business model's GPT-4 level capabilities is dead, right? - It is a common saying that the best businesses being made now are ones that are predicated on models getting better.

- Right, which would be like wrappers, thing that is riding the wave of the models. - The short-term, the company that could make the most money is the one that figures out what advertising targeting method works for language model generations. We have the meta ads, which are hyper-targeted in feed, not within specific pieces of content, and we have search ads that are used by Google, and Amazon has been rising a lot on search.

But within a piece, within a return from ChatGPT, it is not clear how you get a high-quality placed ad within the output. And if you can do that with model costs coming down, you can just get super high revenue. Like, that revenue is totally untapped, and it's not clear technically how it is done.

- Yeah, that is, I mean, sort of the AdSense innovation that Google did. The one day you'll have, in GPT output, an ad, and that's gonna make, like, billions, if not-- - And it could be very subtle. It could be in conversation. Like, we have voice mode now. It could be some way of making it so the voice introduces certain things.

It's much harder to measure, and it takes imagination, but yeah. - And it wouldn't be so, it wouldn't come off shady so that you would receive public blowback, that kind of thing. So, you have to do it loud enough to where it's clear it's an ad and balance all of that.

So, that's the open question they're trying to solve. Anthropic and OpenAI, they need to-- - They might not say that they're trying-- - I don't think they care about that at all. - They don't care about it right now. I think it's places like-- - I think they're purely-- - Perplexity are experimenting on that more.

- Oh, interesting, yeah, for sure. - Like, Perplexity, Google, Meta care about this. I think OpenAI and Anthropic are purely laser focused on-- - AGI. - Yeah, agents and AGI, and if I build AGI, I can make tons of money, right? Or I can pay for everything, right? And this is, it's just predicated, like back on the export control thing, right?

If you think AGI is five, 10 years away or less, right? These labs think it's two, three years away. Obviously, your actions are, if you assume they're rational actors, which they are mostly, what you do in a two year AGI versus five year versus 10 years, very, very, very different, right?

- Do you think agents are promising? We'll have to talk about this. This was, this is like the excitement of the year that agents are gonna, this is the generic hype term that a lot of business folks are using. AI agents are gonna revolutionize everything. - Okay, so mostly the term agent is obviously overblown.

We've talked a lot about reinforcement learning as a way to train for verifiable outcomes. Agents should mean something that is open-ended and is solving a task independently on its own and able to adapt to uncertainty. There's a lot of the term agent applied to things like Apple Intelligence, which we still don't have after the last WWDC, which is orchestrating between apps.

And that type of tool use thing is something that language models can do really well. Apple Intelligence, I suspect, will come eventually. It's a closed domain. It's your messages app integrating with your photos, with AI in the background. That will work. That has been described as an agent by a lot of software companies to get into the narrative.

The question is what ways can we get language models to generalize to new domains and solve their own problems in real time? Maybe some tiny amount of training when they are doing this with fine-tuning themselves or in-context learning, which is the idea of storing information in a prompt and you can use learning algorithms to update that.

And whether or not you believe that that is gonna actually generalize to things like me saying, "Book my trip to go to Austin in two days." I have X, Y, Z constraints in actually trusting it. I think there's a HCI problem, coming back for information. - Well, what's your prediction there?

'Cause my gut says we're very far away from that. - I think opening eyes statement, I don't know if you've seen the five levels, right? Where it's chat is level one, reasoning is level two, and then agents is level three. And I think there's a couple more levels, but it's important to note, right?

We were in chat for a couple of years, right? We just theoretically got to reasoning. We'll be here for a year or two, right? And then agents, but at the same time, like people can train like approximate capabilities of the next level. But the agents are doing things autonomously, doing things for minutes at a time, hours at a time, et cetera, right?

Reasoning is doing things for tens of seconds at a time, right? And then coming back with an output that I still need to verify and use and try to check out. So, and the biggest problem is of course like, it's the same thing with manufacturing, right? Like there's the whole six sigma thing, right?

Like, how many nines do you get? And then you compound the nines onto each other. And it's like, if you multiply, by the number of steps that are six sigma, you get to a yield or something, right? So like in semiconductor manufacturing, tens of thousands of steps, nine, nine, nine, nine, nine, nine, nine, is not enough, right?

'Cause you multiply that by that many times, you actually end up with like 60% yield, right? - Yeah, or zero. - Really low yield, yeah, or zero. And this is the same thing with agents, right? Like chaining tasks together each time. LLMs, even the best LLMs in particularly pretty good benchmarks, don't get a hundred percent, right?

They get a little bit below that because there's a lot of noise. And so how do you get to enough nines, right? This is the same thing with self-driving. We can't have self-driving because without it being like super geo-fenced like Google's, right? And even then they have a bunch of tele-operators to make sure it doesn't get stuck, right?

But you can't do that because it doesn't have enough nines. - And self-driving has quite a lot of structure because roads have rules. It's well-defined, there's regulation. When you're talking about computer use for the open web, for example, or the open operating system, it's a mess. So the possibility...

I'm always skeptical of any system that is tasked with interacting with the human world, with the open, messy human world. - That's the thing, if we can't get intelligence that's enough to solve the human world on its own, we can create infrastructure, like the human operators for Waymo, over many years that enable certain workflows.

- There's a company, I don't remember it, but it is, but that's literally their pitch is, yeah, we're just gonna be the human operator when agents fail. And you just call us and we fix it. - Yeah. - It's like an API call and it's hilarious. - There's gonna be tele-operation markets when we get human robots, which is there's gonna be somebody around the world that's happy to fix the fact that it can't finish loading my dishwasher when I'm unhappy with it, but that's just gonna be part of the Tesla service package.

- I'm just imagining like an AI agent talking to another AI agent. One company has an AI agent that specializes in helping other AI agents. - But if you can make things that are good at one step, you can stack them together. So that's why I'm going, if it takes a long time, we're gonna build infrastructure that enables it.

You see the operator launch. They have partnerships with certain websites, with DoorDash, with OpenTable, with things like this. Those partnerships are gonna let them climb really fast. Their model's gonna get really good at those things. It's gonna proof of concept. That might be a network effect where more companies wanna make it easier for AI.

Some companies will be like, "No, let's put blockers in place." And this is the story of the internet we've seen. We see it now with training data for language models where companies are like, "No, you have to pay." Like business working it out. - That said, I think like airlines have a very, and hotels have high incentive to make their site work really well, and they usually don't.

Like if you look at how many clicks it takes to order an airplane ticket, it's insane. - You actually can't call an American Airlines agent anymore. They don't have a phone number. - I mean, it's horrible on many. On the interface front and all, to imagine that agents will be able to deal with that website when I, as a human, struggle.

Like I have an existential crisis every time I try to book an airplane ticket that I don't, I think it's gonna be extremely difficult to build an AI agent that's robust in that way. - But think about it. Like United has accepted the Starlink term, which is they have to provide Starlink for free and the users are going to love it.

What if one airline is like, we're gonna take a year and we're gonna make our website have white text that works perfectly for the AIs. Every time anyone asks about an AI flight, they buy whatever airline it is. - Or like, they just like, here's an API and it's only exposed to AI agents.

And if anyone queries it, the price is 10% higher and for any flight, but we'll let you see any of our flights and you can just book any of them. Here you go, agent. And then it's like, oh, and I made 10% higher price. Awesome. And like, am I willing to say that for like, hey, book me a flight to see Lex, right?

And it's like, yeah, whatever. I think, you know, computers and real world and the open world are really, really messy. But if you start defining the problem in narrow regions, people are gonna be able to create very, very productive things and ratchet down cost massively, right? Like now crazy things like, you know, robotics in the home, you know, those are gonna be a lot harder to do just like self-driving, right?

Because there's just a billion different failure modes, right? But like agents that can like navigate a certain set of websites and do certain sets of tasks or like look at, you know, take a photo of your fridge and or like upload your recipes and then like it figures out what to order from, you know, Amazon/Whole Foods food delivery.

Like that's, then that's gonna be like pretty quick and easy to do, I think. So it's gonna be a whole range of like business outcomes and it's gonna be tons of sort of optimism around, people can just figure out ways to make money. - To be clear, these sandboxes already exist in research.

There are people who have built clones of all the most popular websites of Google, Amazon, blah, blah, blah, to make it so that there's, I mean, OpenAI probably has them internally to train these things. It's the same as DeepMind's robotics team for years has had clusters for robotics where you interact with robots fully remotely.

They just have a lab in London and you send tasks to it, arrange the blocks and you do this research. Obviously there's techs there that fix stuff, but we've turned these cranks of automation before. You go from sandbox to progress and then you add one more domain at a time and generalize.

I think in the history of NLP and language processing, instruction tuning and tasks per language model used to be like one language model did one task. And then in the instruction tuning literature, there's this point where you start adding more and more tasks together, where it just starts to generalize to every task.

And we don't know where on this curve we are. I think for reasoning with this RL and verifiable domains, we're early, but we don't know where the point is where you just start training on enough domains and poof, like more domains just start working and you've crossed the generalization barrier.

- Well, what do you think about the programming context? So software engineering, that's where I personally, and I know a lot of people interact with AI the most. - There's a lot of fear and angst too from current CS students, but there's also, that's where, that is the area where probably the most AI revenue and productivity gains have come, right?

Whether it be copilots or cursor or what have you, right? This is, or just standard chat GPT, right? Like a lot of, I know very few programmers who don't have chat GPT and actually many of them have the $200 tier because that's what it's so good for, right? I think that in that world, we already see it like SWE bench.

And if you've looked at the benchmark made by some Stanford students, I wouldn't say it's like really hard, I wouldn't say it's easy either. I think it takes someone who's been through at least a few years of CS or a couple of years of programming to do SWE bench well.

And the models went from 4% to 60% in like a year, right? And where are they gonna go to next year? It's gonna be higher, probably won't be 100% 'cause again, that nines is like really hard to do, but we're gonna get to some point where that's, and then we're gonna need harder software engineering benchmarks and so on and so forth.

But the way that people think of it now is it can do code completion easy. It can do some function generation and I have to review it, great. But really the like software engineering agents, I think can be done faster, sooner than any other agent because it is a verifiable domain.

You can always like unit test or compile. And there's many different regions of like, it can inspect the whole code base at once, which no engineer really can, only the architects can really think about this stuff, the really senior guys and they can define stuff. And then the agent can execute on it.

So I think software engineering costs are gonna plummet like crazy. And one interesting aspect of that is when software engineering costs are really low, you get very different markets, right? So in the US, you have all these platform SaaS companies, right, Salesforce and so on and so forth, right?

In China, no one uses platform SaaS. Everyone just builds their own stack because software engineering is much cheaper in China, partially because like people STEM, number of STEM graduates, et cetera. So it's generally just cheaper to do. And so at the same time, code for like code LLMs have been adopted much less in China because the cost of an engineer there is much lower.

But like what happens when every company can just invent their own business logic, like really cheaply and quickly. You stop using platform SaaS, you start building custom tailored solutions, you change them really quickly. Now all of a sudden your business is a little bit more efficient too, potentially, because you're not dealing with the hell that is like some random platform SaaS company stuff, not working perfectly and having to adjust workflows or random business automation cases that aren't necessarily AI required.

It's just logic that needs to be built that no one has built, right? All of these things can go happen faster. And so I think software, and then the other domain is like industrial, chemical, mechanical engineers suck at coding, right? Just generally, and like their tools, like semiconductor engineers, their tools are 20 years old.

All the tools run on XP, including ASML lithography tools, run on Windows XP, right? It's like, you know, and like a lot of the analysis happens in Excel, right? Like, it's just like, guys, like you guys can move 20 years forward with all the data you have and gathered and like do a lot better.

It's just, you need the engineering skills for software engineering to be delivered to the actual domain expert engineer. So I think that's the area where I'm like super duper bullish of generally AI creating value. - The big picture is that I don't think it's gonna be a cliff. It's like, we talked to, I think a really good example of how growth changes is when meta added stories.

So Snapchat was on an exponential, they added stories, it flatlined. Software engineers, then up until the right, AI is gonna come in, it's probably just gonna be flat. It's like, everyone's gonna lose their job. It's hard because the supply corrects more slowly. So the amount of students is still growing and that'll correct on a multi-year, like a year delay.

But the amount of jobs will just turn and then maybe in 20, 40 years, it'll be well down. But in the few years, there'll never gonna be the snap moment where it's like software engineers aren't useful. - I think also the nature of what it means to be a programmer and what kind of jobs programmers do changes.

'Cause I think there needs to be a human in the loop of everything you've talked about. There's a really important human in that picture of like correcting the code. Like fixing-- - Thinking larger than the context length. - Yep, and debugging also. Like debugging by sort of reading the code, understanding the, steering the system.

Like, no, no, no, you missed the point. Adding more to the prompt. Kind of like, yes, adding the human-- - Designing the perfect Google button. Google's famous for having people design buttons that are so perfect. And it's like, how is AI gonna do that? Like, they could give you all the ideas.

Perfect button. - I mean, that's the thing. You can call it taste. Humans have, one thing humans can do is figure out what other humans enjoy better than AI systems. That's where the preference, you're loading that in. But ultimately, humans are the greatest preference generator. That's where the preference comes from.

- And humans are actually very good at reading, or like judging between two things, versus, this goes back to the core of what RLHF and preference tuning is, is that it's hard to generate a good answer for a lot of problems, but it's easy to see which one is better.

And that's how we're using humans for AI now, is judging which one is better. And that's what software engineering could look like. It's the PR review. Here's a few options. What are the, like, here are some potential pros and cons. And they're gonna be judges. - I think the thing I would very much recommend is people start, programmers start using AI, and embracing that role of the supervisor of the AI system, and like, partner of the AI system, versus writing from scratch, or not learning coding at all, and just generating stuff.

'Cause I think there actually has to be a pretty high level of expertise as a programmer to be able to manage increasingly intelligent systems. - I think it's that, and then becoming a domain expert in something. - Sure, yeah. 'Cause seriously, if you go look at aerospace, or semiconductors, or chemical engineering, everyone is using really crappy platforms, really old software.

Like, the job of a data scientist is like a joke, right, in many cases. In many cases, it's very real, but it's like, bring what the forefront of human capabilities are to your domain. And even if the forefront is from the AI, your domain, you're at the forefront, right?

So it's like, you have to be at the forefront of something, and then leverage the rising tide that is AI for everything else. - But yeah, there's so many low-hanging fruit everywhere in terms of where software can help automate a thing, or digitize a thing. In the legal system, I mean, that's why DOJ is exciting.

Yeah, I got to hang out with a bunch of the DOJ folks, and they, I mean, government is so old school. It's like begging for the modernization of software, of organizing the data, all this kind of stuff. I mean, in that case, it's by design, because bureaucracy protects centers of power and so on, but software breaks down those barriers, so it hurts those that are holding onto power, but ultimately benefits humanity.

So there's a bunch of domains of that kind. One thing we didn't fully finish talking about is open source. So first of all, congrats, you released a new model. - Yeah, this is the-- - Tulu. - I'll explain what a Tulu is. A Tulu is a hybrid camel when you breed a dromedary with a Bacchurian camel.

Back in the early days after ChatGPT, there was a big wave of models coming out, like Alpaca, Vicuna, et cetera, that were all named after various mammalian species. So Tulu, the brand is multiple years old, which comes from that, and we've been playing at the frontiers of post-training with open source code.

And this first part of this release was in the fall, where we built on Lama's open weight models, and then we add in our fully open code, our fully open data. There's a popular benchmark that is Chatbot Arena, and that's generally the metric by which how these chat models are evaluated, and it's humans compare random models from different organizations.

And if you looked at the leaderboard in November or December, among the top 60 models from 10s to 20s of organizations, none of them had open code or data for just post-training. Among that, even fewer or none have pre-training data and code available, but it's like post-training is much more accessible at this time.

It's still pretty cheap and you can do it. And the thing is like, how high can we push this number where people have access to all the code and data? So that's kind of the motivation of the project. We draw on lessons from Lama. NVIDIA had a Nemotron model where the recipe for their post-training was fairly open with some data, and a paper, and it's putting all these together to try to create a recipe that people can fine-tune models like GPT-4 to their domain.

- So to be clear, in the case of Tulu, maybe you can talk about Alma too, but in the case of Tulu, you're taking Lama 3.405b. - Tulu has been a series of recipes for post-training. So we've done multiple models over years. - And so you're open sourcing everything.

- Yeah, if you start with an open weight-based model, the whole model technically is an open source 'cause you don't know what Lama put into it, which is why we have a separate thing that we'll get to, but it's just getting parts of the pipeline where people can zoom in and customize.

I know I hear from startups and businesses, they're like, "Okay, I can take this post-training "and try to apply it to my domain." We talk about verifiers a lot. We use this idea, which is reinforcement learning with verifiable domain rewards, RLVR, kind of similar to RLHF, and we've applied it to map.

And the model today, which is like, we applied it to the Lama 405b base model from last year, and we have our other stuff. We have our instruction tuning and our preference tuning, but the math thing is interesting, which is like, it's easier to improve this math benchmark. There's a benchmark, M-A-T-H, math, all capitals.

Tough name. When the benchmark's name is the area that you're evaluating, we're researchers. We're not brand strategists. And this is something that the "Deep Seek" paper talked about as well, is like, at this bigger model, it's easier to elicit powerful capabilities with this RL training, and then they distill it down from that big model to the small model.

And this model we released today, we saw the same thing as it were at AI2. We don't have a ton of compute. We can't train 405b models all the time. So we just did a few runs and they tend to work. And it's like, it just shows that there's a lot of room for people to play in these things.

- And they crushed Lama's actual release, right? Like, they're way better than it. - Yeah, so our eval numbers, I mean, we have extra months in this, but our eval numbers are like much better than the Lama Instruct model that they released. - And then you also said better than Deep Seek V3.

- Yeah, on our eval benchmark. The most, Deep Seek V3 is really similar. We have a safety benchmark to understand if it will say harmful things and things like that. And that's what draws us down most of the way. It's still like- - It's like an amalgamation of multiple benchmarks, or what do you mean?

- Yeah, so we have a 10 evaluators. This is like, this is standard practice in post-training, is you choose your evaluations you care about. In academics, in smaller labs, you'll have fewer evaluations. In companies, you'll have a really one domain that you really care about. In frontier labs, you'll have 10s to 20s to maybe even like 100 evaluations of specific things.

So we choose a representative suite of things that look like chat, precise instruction following, which is like, respond only in emojis. Like, does the model follow weird things like that? Math, code, and you create a suite like this. So safety would be one of 10 in that type of suite where you have like, what is the broader community of AI care about?

And for example, in comparison to DeepSeek, it would be something like our average eval for our model would be 80, including safety, and similar without. And DeepSeek would be like 79% average score without safety, and their safety score would bring it down to like 76 on average. - Oh, so you beat them even ignoring safety.

- Yeah, so this is something that internally, it's like, I don't want to win only by like, how you shape the eval benchmark. So if there's something that's like, people may or may not care about safety in their model, safety can come downstream. Safety can be when you host the model for an API.

Like, safety is addressed in a spectrum of locations in AI applications. So it's like, if you want to say that you have the best recipe, you can't just gate it on these things that some people might not want. And this is just, it's like the time of progress. We benefit, we can release a model later.

We have more time to learn new techniques, like this RL technique. We had started this in the fall. It's now really popular with reasoning models. The next thing to do for open source post-training is to scale up verifiers, to scale up data, to replicate some of DeepSeek's results. And it's awesome that we have a paper to draw on and it makes it a lot easier.

And that's the type of things that is going on among academic and closed frontier research in AI. - Since you're pushing open source, what do you think is the future of it? You think DeepSeek actually changes things since it's open source or open weight, or it's pushing the open source movement into the open direction?

- This goes very back to license discussion. So DeepSeek R1 with a friendly license is a major reset. So it's like the first time that we've had a really clear frontier model that is open weights and with a commercially friendly license with no restrictions on downstream use cases, synthetic data, distillation, whatever.

This has never been the case at all in the history of AI in the last few years since CatGPT. There have been models that are off the frontier or models with weird licenses that you can't really use them. - Isn't Meta's license pretty much permissible except for five companies?

- And there's also, so this goes to like what open source AI is, which is, there's also use case restrictions in the LLAMA license, which says you can't use it for specific things. So if you come from an open source software background, you would say that that is not an open source license.

- What kind of things are those though? Like, are they like? - At this point, I can't pull them off the top of my head, but it'll be like-- - Stuff that's competitor-- - It used to be military use was one, and they removed that for scale. It'll be like CSAM, like child abuse material, or like that's the type of thing that is forbidden there, but that's enough from an open source background to say it's not open source license.

And also the LLAMA license has this horrible thing where you have to name your model LLAMA if you touch it to the LLAMA model. So it's like the branding thing. So if a company uses LLAMA, technically the license says that they should say built with LLAMA at the bottom of their application.

And from like a marketing perspective, that just hurts. Like I can suck it up as a researcher. I'm like, oh, it's fine. Like it says LLAMA dash on all of our materials for this release. But this is why we need truly open models, which is we don't know DeepSeek R1's data.

- Wait, so you're saying I can't make a cheap copy of LLAMA and pretend it's mine, but I can do this with the Chinese model. - Yeah. - Hell yeah. (both laughing) - That's what I'm saying. And that's why it's like, we want this whole open language models thing, the Olmo thing, is to try to keep the model where everything is open with the data as close to the frontier as possible.

So we're compute constrained, we're personnel constrained. We rely on getting insights from people, like John Shulman tells us to do RL on outputs. Like we can make these big jumps, but it just takes a long time to push the frontier of open source. And fundamentally, I would say that that's because open source AI does not have the same feedback loops as open source software.

We talked about open source software for security. Also, it's just because you build something once and you can reuse it. If you go into a new company, there's so many benefits. But if you open source a language model, you have this data sitting around, you have this training code.

It's not like that easy for someone to come and build on and improve, 'cause you need to spend a lot on compute, you need to have expertise. So until there are feedback loops of open source AI, it seems like mostly an ideological mission. People like Mark Zuckerberg, which is like America needs this.

And I agree with him, but in the time where the motivation ideologically is high, we need to capitalize and build this ecosystem around what benefits do you get from seeing the language model data. And there's not a lot about that. We're gonna try to launch a demo soon where you can look at an Olmo model in a query and see what pre-training data is similar to it, which was like legally risky and complicated, but it's like, what does it mean to see the data that the AI was trained on?

It's hard to parse, it's terabytes of files. It's like, I don't know what I'm gonna find in there. But that's what we need to do as an ecosystem if people want open source AI to be financially useful. - We didn't really talk about Stargate. I would love to get your opinion on like, the new administration, the Trump administration, everything that's doing, that's being done from the America side in supporting AI infrastructure and the efforts of the different AI companies.

What do you think about Stargate? What are we supposed to think about Stargate? And does Sam have the money? - Yeah, so I think Stargate is a opaque thing. It definitely doesn't have $500 billion, doesn't even have $100 billion, right? So what they announced is this $500 billion number, Larry Ellison, Sam Altman, and Trump said it.

They thanked Trump and Trump did do some executive actions that like do significantly improve the ability for this to be built faster. One of the executive actions he did is on federal land, you can just basically build data centers in power, you know, like pretty much like that. And then the permitting process is basically gone or you file after the fact.

So like one of the, again, like I had a schizo take earlier, another schizo take, if you've ever been to the Presidio in San Francisco, beautiful area. You could build a power plant and a data center there if you wanted to, because it is federal land. It used to be a military base.

But you know, obviously this would like piss people off. You know, it's a good bit. Anyways, Trump has made it much easier to do this, right, generally. Texas has the only unregulated grid in the nation as well. - Let's go Texas. - And so, you know, therefore like ERCOT enables people to build faster as well.

In addition, the federal regulations are coming down. And so Stargate is predicated, and this is why that whole show happened. Now, how they came up with a $500 billion number is beyond me. How they came up with a $100 billion number makes sense to some extent, right? And there's actually a good table in here that I would like to show in that Stargate piece that I had.

It's the most recent one, yeah. So anyways, Stargate, you know, it's basically, right, like there is, it's a table about cost. There, you passed it already. It's that one. So this table is kind of explaining what happens, right? So Stargate is in Abilene, Texas, the first $100 billion of it.

That site is 2.2 gigawatts of power in, about 1.8 gigawatts of power consumed, right? Per GPU, they have like roughly, Oracle is already building the first part of this before Stargate came about. To be clear, they've been building it for a year. They tried to rent it to Elon, in fact, right?

But Elon was like, "It's too slow, I need it faster." So then he went and did his Memphis thing. And so OpenAI was able to get it with this like weird joint venture called Stargate. They initially signed a deal with just Oracle for the first section of this cluster, right?

This first section of this cluster, right, is roughly $5 billion to $6 billion of server spend, right? And then there's another billion or so of data center spend. But the, and then likewise, like if you fill out that entire 1.8 gigawatts with the next two generations of NVIDIA chips, GB200, GB300, VR200, and you fill it out completely, that ends up being roughly $50 billion of server cost, right?

Plus there's data center costs, plus maintenance costs, plus operation costs, plus all these things. And that's where OpenAI gets to their $100 billion announcement that they had, right? 'Cause they talked about $100 billion as phase one, that's this Abilene Texas data center, right? $100 billion of total cost of ownership, quote unquote, right?

So it's not CapEx, it's not investment, it's $100 billion of total cost of ownership. And then there will be future phases. They're looking at other sites that are even bigger than this 2.2 gigawatts, by the way, in Texas and elsewhere. And so they're not completely ignoring that, but there is the number of $100 billion that they say is for phase one, which I do think will happen.

They don't even have the money for that. Furthermore, it's not $100 billion, it's $50 billion of spend, right? And then like $50 billion of operational cost, power, et cetera, rental pricing, et cetera. 'Cause OpenAI is renting the GPUs from the Stargate joint venture, right? What money do they actually have, right?

SoftBank, SoftBank is gonna invest, Oracle's gonna invest, OpenAI's gonna invest. OpenAI is on the line for $19 billion. Everyone knows that they've only got $6 billion in their last round and $4 billion in debt. But there's news of SoftBank maybe investing $25 billion into OpenAI, right? So that's part of it, right?

So $19 billion can come from there. So OpenAI does not have the money at all, right? To be clear. Ink is not dried on anything. OpenAI has $0 for this $50 billion, right? In which they're legally obligated to put $19 billion of CapEx or into the joint venture. And then the rest they're gonna pay via renting the GPUs from the joint venture.

And then there's Oracle. Oracle has a lot of money. They're building the first section completely. They were spending for it themselves, right? This $6 billion of CapEx, $10 billion of TCO. But they, and they were gonna do that first section. They're paying for that, right? As far as the rest of the section, I don't know how much Larry wants to spend, right?

At any point he can pull out, right? Like this is, again, this is like completely voluntary. So at any point, there's no signed ink on this, right? But he potentially could contribute tens of billions of dollars, right? To be clear, he's got the money. Oracle's got the money. And then there's like MGX, which is the UAE fund, which technically has $1.5 trillion for investing in AI.

But again, like, I don't know how real that money is. And like, whereas there is no ink signed for this, SoftBank does not have $25 billion of cash. They have to sell down their stake in ARM, which is, you know, the leader in CPUs. And they IPO'd it. This is obviously what they've always wanted to do.

They just didn't know where they'd redeploy the capital. Selling down the stake in ARM makes a ton of sense. So they can sell that down and invest in this if they want to, and invest in OpenAI if they want to. As far as like money secured, the first 100,000 GB200 cluster is like, can be funded.

Everything else after that- - Up in the air. - Is up in the air. Money's coming. I believe the money will come. I personally do. - It's just, it's a belief, okay. - It's a belief that they are gonna release better models and be able to raise more money, right?

But like, the actual reality is is that Elon's right. There is, the money does not exist, right? - What does the US government have to do with anything? What does Trump have to do with everything? He's just a hype man? - Trump is, he's reducing the regulation so they can build it faster, right?

And he's allowing them to do it, right? You know, 'cause any investment of this side is gonna involve like antitrust stuff, right? Like, so obviously he's gonna allow them to do it. He's gonna enable the regulations to actually allow it to be built. I don't believe there's any US government dollars being spent on this though.

- Yeah, so I think he's also just creating a general vibe that this is, regulation will go down and this is the era of building. So if you're a builder, you wanna create stuff, you wanna launch stuff, this is the time to do it. - And so like, we've had this 1.8 gigawatt data center in our data for over a year now.

And we've been like sort of sending it to all of our clients, including many of these companies that are building the multi gigawatts. But that is like at a level that's not quite maybe executives like seeing $500 billion, $100 billion and then everyone's asking them like, so it could spur like another, like an even faster arms race, right?

'Cause there's already an arms race, but like this like 100 billion, $500 billion number, Trump talking about it on TV, like it could spur the arm race to be even faster and more investors to flood in and et cetera, et cetera. So I think you're right is that in that sense that open AI or sort of Trump is sort of like championing people are gonna build more and his actions are gonna let people build more.

- What are you excited about these several years that are upcoming in terms of cluster build outs, in terms of breakthroughs in AI, like the best possible future you can imagine in the next couple of years, two, three, four years, what does that look like? Just it could be very specific technical things like breakthroughs on post-training or it could be just size big impressive clusters.

- I really enjoy tracking supply chain and like who's involved in what. I really do, it's really fun to see like the numbers, the cost, who's building what capacity, helping them figure out how much capacity they should build, winning deals, strategic stuff, that's really cool. I think technologically, there's a lot around the networking side that really excites me with optics and electronics, right?

Like kind of getting closer and closer, whether it be co-packaged optics or some sort of like forms of new forms of switching. - This is internal to a cluster? - Yeah, also multi-data center training, right? Like there's people are putting so much fiber between these data centers and lighting it up with so much bandwidth that there's a lot of interesting stuff happening on that end, right?

Telecom has been really boring since 5G and now it's like really exciting again on the hardware side. - Can you educate me a little bit about the speed of things? So the speed of memory versus the speed of interconnect versus the speed of fiber between data centers, are these like orders of magnitude different?

Can we at some point converge towards a place where it all just feels like one computer? - No, I don't think that's possible. It's only gonna get harder to program, not easier. It's only gonna get more difficult and complicated and more layers, right? The general image that people like to have is like this hierarchy of memory.

So on-chip is really close, localized within the chip, you have registers, right? And those are shared between some compute elements. And then you'll have caches, which are shared between more compute elements. Then you have like memory, right? Like HBM or DRAM, like DDR memory or whatever it is. And that's shared between the whole chip.

And then you can have, you know, pools of memory that are shared between many chips, right? And then storage and you keep zoning out, right? The access latency across data centers, across within the data center, within a chip, is different. So like, you're obviously always, you're always gonna have different programming paradigms for this.

It's not gonna be easy, programming this stuff is gonna be hard, maybe I can help, right? You know, with programming this. But the way to think about it is that like, there is, there's sort of like, the more elements you add to a task, you don't gain, you don't get strong scaling, right?

If I double the number of chips, I don't get 2X the performance, right? This is just like a reality of computing, 'cause there's inefficiencies. And there's a lot of interesting work being done to make it not, you know, to make it more linear, whether it's making the chips more networked together, more tightly, or, you know, cool programming models, or cool algorithmic things that you can do on the model side, right?

DeepSeek did some of these really cool innovations because they were limited on interconnect, but they still needed to parallelize, right? Like all sorts of, you know, everyone's always doing stuff, Google's got a bunch of work, and everyone's got a bunch of work about this. That stuff is super exciting on the model, and workload, and innovation side, right?

Hardware, solid state transformers are interesting, right? For the power side, there's all sorts of stuff on batteries, and there's all sorts of stuff on, you know, I think when you look at, if you look at every layer of the compute stack, right, whether it goes from lithography, and etch, all the way to like fabrication, to like optics, to networking, to power, to transformers, to cooling, to, you know, networking, and you just go on, up, and up, and up, and up the stack, you know, even air conditioners for data centers are like innovating, right?

Like it's like, there's like, copper cables are innovating, right? Like you wouldn't think it, but copper cables, like are, there's some innovations happening there with like the density of how you can pack them. And like, it's like all of these layers of the stack, all the way up to the models.

Human progress is at a pace that's never been seen before. - I'm just imagining you sitting back in a layer somewhere with screens everywhere, just monitoring the supply chain, where all these clusters, like all the information you're gathering. I mean, you do incredible-- - There's a big team. There's a big team.

- Yeah, I mean, you do quite incredible work with semi-analysis. I mean, it's just, keeping your finger on the pulse of human civilization in the digital world. It's pretty cool. Like just to watch, feel that. - Yeah, thank you. I guess. - Feel all of us like doing shit, epic shit.

- Feel the AGI. - I mean, from meme to like reality. What, Nathan, is there like breakthroughs that you're like looking forward to potentially? - I had a while to think about this while listening to Dylan's beautiful response. - He didn't listen to me, he was so dumb. - I knew, no, I knew this was coming.

And it's like, realistically, training models is very fun because there's so much low-hanging fruit. And the thing that makes my job entertaining, I train models, I write analysis about what's happening with models. And it's fun because there is obviously so much more progress to be had. And the real motivation why I do this somewhere where I can share things is that there's just, I don't trust people that are like, trust me, bro, we're gonna make AI good.

That's like, we're the ones that it's like, we're gonna do it and you can trust us and we're just gonna have all the AI. And it's just like, I would like a future where more people have a say in what AI is and can understand it. And that's, it's a little bit less fun that it's not a positive thing.

I feel like this is just all really fun. Like training models is fun and bringing people in is fun, but it's really like AI, if it is going to be the most powerful technology of my lifetime, it's like, we need to have a lot of people involved in making that.

- And making it open helps with that. As accessible as possible, as open as possible, yeah. - In my read of the last few years that more openness would help the AI ecosystem in terms of having more people understand what's going on, rather that's researchers from non AI fields to governments to everything.

It doesn't mean that openness will always be the answer. I think that it will reassess of like what is the biggest problem facing AI and tack on a different angle to the wild ride that we're on. - And for me, just from even the user experience, anytime you have the, like Apathy said, the aha moments, like the magic, like seeing the reasoning, the chain of thought, it's like, there's something really just fundamentally beautiful about that.

It's putting a mirror to ourselves and seeing like, oh shit, it is solving intelligence as the cliche, like goal of these companies is. And you get to understand why we humans are special. The intelligence within us is special. And for now also why we are special in terms of we seem to be conscious and the AI systems for now aren't.

And we get to solve, we get to explore that mystery. So that's, it's just really cool to get to explore these questions that I don't think, I would have never imagined would be even possible. Back when, so just watching with excitement Deep Blue be Kasparov, like I wouldn't have ever thought this kind of AI would be possible in my lifetime.

It's like, this is really feels like AI. - Yeah. - It's incredible. - I started with AI of learning to fly a silly quadrotor. It's like, learn to fly. And it was just like, it learned to fly up. It would hit the ceiling and stop and catch it. It's like, okay, that is like really stupid compared to what's going on now.

- And now you could probably, with natural language, tell it to learn to fly and it's going to generate the control algorithm required to do that. - Probably. There's low level blockers. Like we had to do some weird stuff for that. But you can, you definitely can. - Go back to our robotics conversation.

Yeah, when you have to interact in actual physical world, it's hard. What gives you hope about the future of human civilization? Looking into the next 10 years, 100 years, 1,000 years, how long do you think we'll make it? You think we've got 1,000 years? - Humans will definitely be around in 1,000 years.

I think there's ways that very bad things could happen and there'll be way fewer humans. But humans are very good at surviving. There's been a lot of things that that is true. I don't think they're necessarily, we're good at long-term credit assignment of risk. But when the risk becomes immediate, we tend to figure things out.

And for that reason, there's physical constraints to things like AGI, recursive improvement to kill us all type stuff. For the physical reasons and for how humans have figured things out before, I'm not too worried about AI takeover. There are other international things that are worrying, but there's just fundamental human goodness and trying to amplify that.

We're on a tenuous time. And I mean, if you look at humanity as a whole, there's been times where things go backwards. There's times when things don't happen at all and we're on what should be very positive trajectory right now. - Yeah, there seems to be progress, but just like with power, there's spikes of human suffering.

And we wanna try to minimize the amount of spikes. - Generally, humanity is gonna suffer a lot less. I'm very optimistic about that. I do worry of techno-fascism type stuff arising as AI becomes more and more prevalent and powerful and those who control it can do more and more.

Maybe it doesn't kill us all, but at some point, every very powerful human is gonna wanna brain-computer interface so that they can interact with the AGI and all of its advantages in many more way and merge its mind with sort of like, and its capabilities or that person's capabilities can leverage those much better than anyone else and therefore won't be one person rule them all, but it will be...

The thing I worry about is it'll be like few people, hundreds, thousands, tens of thousands, maybe millions of people rule whoever's left, right? And the economy around it, right? And I think that's like the thing that's probably more worrisome is like human machine amalgamations. This enables an individual human to have more impact on the world and that impact can be both positive and negative, right?

Generally, humans have positive impacts on the world, at least societally, but it's possible for individual humans to have such negative impacts and AGI, at least as I think the labs define it, which is not a runaway sentient thing, but rather just something that can do a lot of tasks really efficiently, amplifies the capabilities of someone causing extreme damage.

But for the most part, I think it'll be used for profit-seeking motives, which will then reduce, which will increase the abundance and supply of things and therefore reduce suffering, right? That's the goal. - Scrolling on a timeline, just throttling your dopamine. - Scrolling is stasis. - Scrolling holds the status quo of the world.

- That is a positive outcome, right? It's like if I have food tubes and I'm scrolling and I'm happy, that's a positive outcome. (both laughing) - While expanding out into the cosmos. Well, this is a fun time to be alive and thank you for pushing the forefront of what is possible in humans and thank you for talking today.

This was fun. - Thanks for having us. - Thanks for having us. - Thanks for listening to this conversation with Dylan Patel and Nathan Lambert. To support this podcast, please check out our sponsors in the description. And now let me leave you with some words from Richard Feynman. For a successful technology, reality must take precedence over public relations.

For nature cannot be fooled. Thank you for listening. I hope to see you next time. (upbeat music) (upbeat music)