State of AI 2023: Highlights of 163 Page Report + Eureka Self-Improvement, MEG, Suno AI and GPT F

00:00:00.080 | The annual State of AI report is one of the most cited documents in the field of artificial

00:00:06.240 | intelligence endorsed by the likes of Andrej Karpathy. Led by Nathan Benet this year,

00:00:11.360 | the original co-author Ian Hogarth is now one of the key players at the Bletchley Global AI

00:00:17.440 | Summit in November. Yes, I read all 163 pages of this year's edition and thought that giving

00:00:23.600 | you just the highlights interspersed with developments from just the last few days

00:00:28.560 | like Eureka would be a good way of keeping you guys up to speed. But let's start out with a

00:00:34.720 | modality that the report didn't cover. Notice on page 7 we have text, image, video, music, robot

00:00:41.120 | states and we'll cover all of those but what we don't have is thoughts. And why am I bringing that

00:00:46.240 | up you might ask? Well because a few days ago we had this from Meta: "Towards a real-time decoding

00:00:52.480 | of images from brain activity". Now I have showcased thought to image and thought to text before on the

00:00:58.080 | channel but never real-time thought to image. It's not perfect but progress is astonishing and

00:01:04.160 | I think it points to the explosion of modalities that we are seeing at the moment in AI. Now if

00:01:09.680 | you're willing to sacrifice a real-time flow of images from the brain this is the state of the

00:01:15.520 | art using fMRI. Much more accurate but not real-time. Now many of you might wonder if it works if

00:01:21.760 | someone imagines an image and apparently it doesn't work nearly as well. And also at the moment it tends not to be a real-time decoding.

00:01:26.640 | Also at the moment it tends not to work if the participants are distracted by other things.

00:01:31.680 | As the paper says: "In other words the subject's consent is not only a legal but also a technical

00:01:37.920 | requirement for brain decoding". Back to the report and many pages are dedicated to covering GPT-4

00:01:44.480 | but to be honest anyone who's followed this channel for any length of time will know a fair bit more

00:01:49.760 | than what is covered in these few slides. So I'm going to skip to page 14 where I have a slight

00:01:55.600 | difference with one of the conclusions. They cite a paper which describes the false promise of

00:02:00.880 | imitating proprietary LLMs, large language models. Basically saying that you can imitate the style

00:02:06.880 | but not the content of a smarter model. But I wish that they had cited Orca which despite being more

00:02:12.560 | than 10 times smaller than the original ChachiPT reaches according to the technical paper parity

00:02:18.800 | with ChachiPT on various benchmarks showing only a four-point gap in professional and academic

00:02:25.200 | examinations like the SAT, LSAT, GRE and GMAT. That's maybe why the information reported on

00:02:31.840 | Microsoft trying to substitute for GPT-4 with Orca. Claiming to mimic the quality of OpenAI

00:02:38.560 | software at a fraction of the cost. On page 38 of the report they describe

00:02:43.440 | open-ended learning with large language models. Describing how they can explore

00:02:47.840 | and gain skills in a game like Minecraft. And they say that the best example of this kind of

00:02:54.080 | self-improvement and iterative prompting is Voyager. Which is again GPT-4 getting better

00:02:59.680 | and better at Minecraft. But just after the report came out we got Eureka. And yes it deserves its

00:03:05.680 | own video but as Jim Phan, one of the key authors said, it's like Voyager in the space of physics

00:03:11.600 | simulations. And essentially what they did was feed the source code of the environment to

00:03:16.880 | GPT-4. And then they asked it to write the code for the reward function. And then using the sheer

00:03:22.960 | power of parallelization inside Nvidia's Isaac Jim. They could then test out those reward functions.

00:03:29.360 | See which ones in simulation work the best. Thanks to GPU acceleration you could speed

00:03:34.400 | up reality a thousand x. Each time the best grading systems would be used, refined upon

00:03:39.440 | and then tested again. All in simulation. Ultimately Eureka rewards outperformed expert human written

00:03:45.920 | reward functions on 83% of the tasks performed. The average margin was 52% and Eureka was able to

00:03:53.520 | learn even pin spinning tricks. Which are notoriously difficult even for CGI artists.

00:03:59.040 | This marrying up of large language models and iterative feedback from the environment is

00:04:04.400 | probably the future of AI. As Shetal Shah from Microsoft said, the proverbial positive feedback

00:04:10.880 | loop of self-improvement might be just around the corner. That allows us to go beyond human-made

00:04:14.960 | training data and capabilities. And to those wondering, yes they are planning to connect it to

00:04:21.120 | a real robotic hand soon. I was lucky enough to get early access to the report and one of the most

00:04:26.400 | convincing and amusing demonstrations was Eureka learning how to run efficiently. Back to the

00:04:33.040 | report though and if Eureka was all about simulated embodiment. Getting better with

00:04:38.480 | language feedback. How about language tasks being improved with robotics data? As this page

00:04:44.400 | demonstrates the positive transfer seems to go both ways. Gaining vision data and embodiment or

00:04:50.000 | robotics data seems to make a model like Palm-E better at pure language tasks. And while the

00:04:56.080 | report went on to mention RT2, watchers of my channel will know about RT2X. That showed how

00:05:02.960 | data from a multiplicity of different robotic setups could complement each other and improve

00:05:08.800 | the base model. The basic message across the board is the same. More data from more modalities in the

00:05:13.840 | same modality. And so when we get Eureka improving via feedback from GPT-4, it reminds me of a

00:05:21.440 | previous video where I talked about DALI-3 improving from feedback from GPT-4 vision.

00:05:27.280 | The same themes are reoccurring. Improvements in one modality ricocheting across two others.

00:05:33.280 | And remember the original Palm language model and GPT-4 weren't even designed for this. I

00:05:38.480 | am genuinely curious about the kind of recursive improvement that might occur with

00:05:43.280 | a model designed to be multimodal from the start like Gemini. Speaking of robotics though,

00:05:48.320 | the report did mention something I missed on the channel. That's the first time win for a robot in

00:05:53.840 | a competitive sport, which was first person view drone racing. Here is a glimpse of why the Swift

00:05:59.760 | system beat all the human champions. "The trajectories flown by the humans and Swift,

00:06:04.960 | we noticed that the autonomous system is more consistent across laps and is able to take tighter

00:06:10.160 | turns." "The trajectory is flown by the humans and swift, we noticed that the autonomous system is more consistent across laps and is able to take tighter turns."

00:06:13.280 | "This gives it a decisive advantage in a race."

00:06:15.760 | "Swift won multiple races against each of the human champions and achieved the fastest recorded race time."

00:06:23.600 | "This marks the first time that an autonomous mobile robot has achieved

00:06:33.200 | world champion level performance in a real world competitive sport."

00:06:36.960 | What might come next? Well, these have received approval in China for fully autonomous flights,

00:06:42.560 | and the new model is expected to be launched in the next few months.

00:06:45.120 | Let me know in the comments if you would fly in one of these e-hang drones. At the moment I'm

00:06:49.200 | undecided, probably not right now, I'd have to see in action for a bit longer.

00:06:54.240 | On page 58 the authors describe another year of progress in music generation. But rather than

00:07:00.480 | just describe this progress, I want you to hear it. I'm not sponsored by them, but I've been honestly

00:07:06.400 | very impressed by Suno AI. Check out a snippet of the following rap, all generated by Suno.

00:07:12.400 | "A.I. explain, breaking through the mold, diving deep discovering secrets untold.

00:07:22.240 | Smarter than human, it's a whole new game. Unleashing power, a technological flame."

00:07:32.080 | If you like that one, I'm going to end the video with another sample. Very quickly though, the way

00:07:36.560 | to test it out is to join their discord, click on one of the chirp beaters, and then just type

00:07:41.440 | forward slash chirpbeaters.

00:07:42.240 | Then you have to press enter again, for some reason, and up this will come.

00:07:48.080 | At the top you can put things like rap or classical, whatever you want. And then you

00:07:51.920 | can either enter 4-8 lines of your own lyrics or have ChatGPT do it for you.

00:07:56.560 | Honestly don't make more than around 8 lines because it degrades performance quite a lot.

00:08:00.880 | Page 63 briefly describes the progress in medical performance of language models.

00:08:06.720 | And I do honestly think that was worth highlighting. I just wonder if MedPalm 2, when it got 85%

00:08:12.080 | used a smart GPT like framework. Because if it didn't, things like self consistency,

00:08:17.520 | chain of thought prompting, and reflection could boost the score even higher than 85%.

00:08:22.880 | In other words, it is not out of the question that we could soon have in our pockets

00:08:26.880 | a model capable of outperforming all but the very best doctors at medical question answering.

00:08:32.800 | Speaking of medicine, apparently it's the fastest growing field in terms of mentions

00:08:37.280 | of AI within papers. And just look at the breadth of fields now applying

00:08:41.920 | AI to achieve research breakthroughs and the increase in volume.

00:08:46.320 | And notice at the bottom how much mathematics is benefiting from this progress.

00:08:50.800 | And I bet many of you don't know about the automated prover and proof assistant GPT-F.

00:08:56.400 | One of the co-authors of that in 2020 was Ilya Sutskova of OpenAI.

00:09:00.800 | It was about proving mathematical theorems using a GPT model.

00:09:05.040 | And again something many might not know is that that was superseded in 2022 by Meta. Their neural

00:09:11.760 | theorem prover solved 10 international math olympiad problems. That's 5 times more than

00:09:17.360 | any previous system. Their method was called "hypertree proof search" and is way too complicated

00:09:23.040 | for me to cover in this video. And that's not even touching Alpha Tensor from Google DeepMind

00:09:27.920 | which improved upon decades old methods for matrix multiplication. That has direct ramifications for

00:09:34.160 | the usage of GPUs in modern day AI. Very quickly what they did to automate algorithmic discovery was

00:09:41.600 | to convert the finding of efficient algorithms for matrix multiplication into a single player game.

00:09:47.440 | And I love this bit. They say this game is incredibly challenging. The number of possible

00:09:52.160 | algorithms to consider is much greater than the number of atoms in the universe. And compared to

00:09:56.720 | the game of Go, which remained a challenge for AI for decades, the number of possible moves at

00:10:01.840 | each step of our game is 30 orders of magnitude larger. And it's the synergies that I'm interested

00:10:07.200 | in. The breakthroughs in one field that will then be applied to another. The

00:10:11.440 | art results in one modality that push forward another modality. And the number of people trying

00:10:17.200 | to do this is multiplying as well. Honestly, I hadn't even heard of Imbue. Apparently they have

00:10:23.280 | 10,000 H100s which are the most cutting edge GPUs. I don't know how I missed an AI company with 5

00:10:30.160 | figure numbers of H100 GPUs but here's the CEO describing their approach. At Imbue we're training

00:10:36.480 | large foundation models optimized for reasoning. On top of those models we build agents that we use

00:10:41.280 | to accelerate our own research. And these do more than just output something. They also iterate and

00:10:46.720 | reflect and figure out what's the next step to do and then take that next step. We're starting with

00:10:50.880 | agents that code because it requires complex reasoning to be able to code well and because

00:10:55.600 | that's the work we do every day. They seem to be flying completely under the radar with only about

00:11:00.480 | 300 subscribers on their channel. So to be honest, if we see the kind of manual dexterity

00:11:05.840 | that Eureka promises and a breakthrough in reasoning like Imbue and frankly every

00:11:11.120 | other company is working on. And if AI can start to crack international math olympiad problems and

00:11:16.160 | compose rap battles, it does start to become a genuine question of what can't they do. I'm not

00:11:22.320 | saying that's true now, far from it, but in the not too distant future that may well be a very

00:11:27.920 | valid question. In the industry section on page 82 the authors describe the chip export bans that

00:11:34.400 | apply to H100s. Basically companies like Nvidia and Intel can't sell their cutting edge chips to China

00:11:40.960 | and select other countries. And those companies got around it by creating models like the A100

00:11:46.160 | which were just outside of the export restriction range. Well the report is already somewhat out of

00:11:52.720 | date in that respect. Just four days ago we hear that those chips are now also under export

00:11:58.880 | restrictions. Another development that's occurred since the publication of the report concerns

00:12:03.200 | copyright. The report detailed how Microsoft has moved to reassure the users of its tools that the

00:12:09.040 | corporation will assume any use of the chip.

00:12:10.800 | So the report is that the company is now under the jurisdiction of the company that is responsible

00:12:14.560 | for the copyright claim. And of course every day we seem to be reading reports of one company or

00:12:19.360 | another getting sued as it relates to AI data. But just a few days ago we had this, Google giving

00:12:24.800 | the same kind of indemnification. Basically if the use of one of its Gen AI products triggers

00:12:30.560 | any copyright claims, it, Google, will assume legal responsibility, it says.

00:12:35.600 | My only prediction is that this battle over data and copyright is not going anywhere.

00:12:40.640 | So the report is that the company is now under the jurisdiction of the corporation that is

00:12:44.400 | responsible for the use of the chip. And that's why it's so important that the company

00:12:47.840 | is not under the jurisdiction of the corporation that it's not going anywhere.

00:12:50.400 | So the report is that the company is now under the jurisdiction of the corporation that it's

00:12:53.280 | not going anywhere. So the report is that the company is now under the jurisdiction of the

00:12:56.160 | corporation that it's not going anywhere. So the report is that the company is now under the

00:12:59.600 | jurisdiction of the corporation that it's not going anywhere. So the company is now under the

00:13:03.360 | jurisdiction of the corporation that it's not going anywhere. So the company is now under the

00:13:06.160 | jurisdiction of the corporation that it's not going anywhere. So the company is now under the jurisdiction

00:13:10.480 | of the corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:12.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:14.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:16.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:18.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:20.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:22.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:24.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:26.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:28.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:30.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:32.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:34.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:36.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:38.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:40.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:42.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:44.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:46.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:48.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:50.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:52.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:54.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:56.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:13:58.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:00.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:02.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:04.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:06.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:08.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:10.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:12.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:14.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:16.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:18.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:20.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:22.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:24.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:26.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:28.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:30.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:32.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:34.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:36.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:38.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:40.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:42.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:44.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:46.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:48.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:50.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:52.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:54.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:56.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:14:58.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:00.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:02.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:04.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:06.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:08.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:10.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:12.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:14.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:16.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:18.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:20.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:22.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:24.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:26.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:28.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:30.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:32.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:34.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:36.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:38.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:40.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:42.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:44.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:46.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:48.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:50.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:52.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:54.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:56.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:15:58.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:00.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:02.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:04.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:06.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:08.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:10.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:12.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:14.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:16.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:18.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:20.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:22.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:24.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:26.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:28.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:30.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

00:16:32.480 | corporation that it's not going anywhere. So the company is now under the jurisdiction of the

State of AI 2023: Highlights of 163 Page Report + Eureka Self-Improvement, MEG, Suno AI and GPT F

Chapters