‘Everything is Going to Be Robotic’ Nvidia Promises, as AI Gets More Real
00:00:00.000 | The CEO of NVIDIA revealed that he wants his company to become ultimately one giant AI.
00:00:07.200 | Even if that feels a little ways away, he did showcase in the last couple of days
00:00:11.920 | a string of capabilities that are possible now with AI.
00:00:16.560 | Yes, we're going to hear three big promises about the future of AI,
00:00:20.160 | but we're going to see a host of demos of things that are possible right now.
00:00:24.320 | I'll bring in clips from some recent interviews I've conducted,
00:00:28.080 | and we'll hear from the chief of staff of one prominent AI company predicting
00:00:32.800 | the end of employment as we know it in three to five years, which I think is a tad overstated.
00:00:39.520 | Speaking of which, you'll also see some AI fails as a spam campaign flops hard.
00:00:45.520 | So what about those three promises I mentioned from the CEO of NVIDIA,
00:00:50.160 | which looks set to become the largest company in the world if current trends hold?
00:00:55.440 | Well, first we heard and saw that NVIDIA anticipates robots revolutionizing industry.
00:01:01.840 | That's still pretty general though, right? So how about the prediction
00:01:16.880 | that everything is going to be robotic?
00:01:19.840 | Let me talk about what's next.
00:01:23.280 | The next wave of AI is physical AI. AI that understands the laws of physics.
00:01:30.480 | AI that can work among us.
00:01:34.400 | Of course, when I say robotics, there's a humanoid robotics that's usually the representation of that.
00:01:41.200 | Everything is going to be robotic. All of the factories will be robotic.
00:01:45.600 | The factories will orchestrate robots and those robots will be building products that are robotic.
00:01:52.560 | Robots interacting with robots, building products that are robotic.
00:01:58.320 | And of course, we don't just have robots building robots.
00:02:01.520 | We have artificial intelligence improving artificial intelligence.
00:02:05.600 | Here is Jason Huang on a separate, less reported occasion,
00:02:09.360 | promising to turn NVIDIA into one giant AI.
00:02:14.000 | We can't design a chip anymore without AI.
00:02:17.440 | At night, our AIs are exploring design spaces vast and wide that we would never do ourselves
00:02:24.800 | because it costs too much money to explore it.
00:02:26.640 | We can't write software without without AI anymore.
00:02:29.120 | We have to explore all the, you know, the design space of optimizing compilers is too large.
00:02:34.560 | We use AIs to file bugs.
00:02:37.200 | So our bug, you know, our bugs database actually tells you what's wrong with the code,
00:02:41.920 | who's likely involved and activates that person to go fix it.
00:02:46.560 | You know, and so I think we, I want everybody,
00:02:51.040 | every organization or company to use AI very aggressively.
00:02:53.920 | I want to turn NVIDIA into one giant AI.
00:02:56.640 | But it's well past time that I become a bit more concrete about what models can do right now, today.
00:03:03.040 | Here is a 30 second clip from NVIDIA that actually undersold what AI is capable of.
00:03:11.040 | Multimodal LLMs are breakthroughs that enable robots to learn,
00:03:15.760 | perceive and understand the world around them and plan how they'll act.
00:03:20.320 | And from human demonstrations,
00:03:22.960 | robots can now learn the skills required to interact with the world using gross and fine motor skills.
00:03:29.840 | But how was that underselling the capabilities of AI?
00:03:34.000 | It looked pretty impressive, right?
00:03:35.840 | Well, they focused on AI learning from human demonstrations.
00:03:40.160 | But if you've watched my Dr. Eureka video recently,
00:03:42.800 | you'll know that it's not just about LLMs coming up with high level plans
00:03:47.360 | and then relying on human demonstrations to exercise fine grained robotic control,
00:03:52.960 | in this case of a robot dog.
00:03:54.720 | LLMs are actually really good at programming the robo dog to,
00:03:59.600 | in this case, stay balanced on a moving rolling yoga ball.
00:04:03.200 | And I spoke with Jason Ma, the lead author of the Dr. Eureka paper,
00:04:07.520 | which was made in collaboration with NVIDIA,
00:04:09.840 | about how that will only accelerate.
00:04:12.240 | Robot capabilities will be bootstrapped by large language models.
00:04:16.080 | And I think that's the most interesting thing of using LLM for robotics, honestly.
00:04:19.200 | Like there's a lot of work in using large language models for robotics
00:04:22.480 | in the high level planning category.
00:04:24.560 | I can plan the sequence of tasks the robot needs to do,
00:04:27.360 | but I think fundamentally the bottleneck for robotics
00:04:29.760 | is still like the low level of physical control, right?
00:04:32.560 | LLM can tell the robot to cook some food,
00:04:34.640 | but if the robot can't even pick up a knife properly,
00:04:37.040 | it's not going to work.
00:04:37.920 | But I think a lot of Eureka where my work is focused on
00:04:41.360 | how do we use this highly capable reasoning, coding,
00:04:44.320 | text models, multimodal models to supervise the low level learning.
00:04:48.240 | So the robots can do the very complex tasks in the first place.
00:04:51.360 | And I think that will only accelerate.
00:04:53.120 | The key edge that AI has is that it can iterate thousands and thousands of times
00:04:58.480 | in parallel in simulation until it's got a program it's happy with.
00:05:03.200 | And dipping back into the virtual world for a moment,
00:05:06.560 | how about the long awaited promise of being able to interact live with video game characters?
00:05:33.600 | And speaking of realism, before I get to the latest clips from NVIDIA,
00:05:38.320 | here's me speaking six weeks ago about how good lip syncing was getting.
00:05:43.440 | Using just a single photo of you, we can now get you to say anything.
00:06:01.280 | I have to remind myself that these aren't projections.
00:06:03.840 | This is what is currently possible.
00:06:05.920 | Imagine that accuracy of lip syncing on a digital human of this level of realism.
00:06:27.920 | I do wonder sometimes how many decades away we are
00:06:31.360 | from a time where you could be speaking to someone and not be
00:06:35.280 | entirely certain in the real world whether or not they are embodied AI.
00:06:40.160 | I might previously have said that's a hundred years away,
00:06:42.880 | but now I think it might be in my lifetime.
00:06:45.600 | But I'm off track because I promised more demos of things that are possible with AI today.
00:06:50.720 | So how about a weather report that's localized to your building, your pavement?
00:06:56.720 | But we are not stopping there.
00:06:58.240 | The next frontier is hyperlocal forecasting down to tens of meters
00:07:03.520 | where the effects of city infrastructure are taken into account.
00:07:06.560 | When combined with weather simulation windfields, it can model the airflow around buildings.
00:07:12.400 | We expect to predict phenomena such as downwash,
00:07:15.920 | where strong winds funnel down to street level, causing damage and affecting pedestrians.
00:07:23.440 | NVIDIA Earth 2, an excellent example of a digital twin that fuses AI,
00:07:29.120 | physics simulations, and observed data can help countries and companies
00:07:34.880 | see the future and respond to the impact of extreme weather.
00:07:39.040 | Or what about a coffee shop which is staffed by dozens of robots
00:07:42.880 | with just one or two humans to oversee things?
00:07:45.840 | Wait, that's happening right now.
00:07:47.840 | All of these things feel futuristic and far away until they actually happen.
00:07:53.120 | And how about a sound effect generator that can generate any sound?
00:07:58.240 | Well, that is now possible today with Eleven Labs.
00:08:01.680 | Actually, I'm going to test it with something like a robot being crushed.
00:08:05.920 | Let's see if it comes up with something interesting or not.
00:08:10.560 | So far, about five, six, seven seconds. Not too bad.
00:08:15.440 | And how is it?
00:08:18.960 | Whoa.
00:08:19.460 | Not perfect, obviously, but if you feel that all of this is in the future,
00:08:27.200 | let me bring you a video from a graphic designer who lost his job recently to AI.
00:08:33.680 | He just lost my job and I lost it to AI, which is very unfortunate.
00:08:39.040 | I think many people joke about the, you know, the fact that,
00:08:43.280 | oh, AI is going to take all our jobs and we're all going to get replaced.
00:08:46.400 | And especially within my industry, which is graphic design.
00:08:50.400 | And it turns out basically all of the material that I've provided over the past six years
00:08:55.040 | is now being fed to AI and templated.
00:08:59.360 | So a design that would take me 30 minutes now takes AI 30 seconds
00:09:05.600 | as it's been trained on all my templates.
00:09:08.560 | Essentially, I think it just literally reuses my templates
00:09:11.760 | and then they can input the hex codes they want the email or the website designed to be,
00:09:16.800 | drag and drop in the client's logo, upload the client's font and boom,
00:09:21.680 | it will generate my template by using their brand assets.
00:09:25.840 | It's a reminder that even though almost all AI needs human generated training data to get started,
00:09:32.160 | they don't necessarily need more of it to keep going.
00:09:35.360 | Or to put it another way, this is the worst that AI, embodied or not, will ever be.
00:09:40.640 | Which is probably why some people,
00:09:42.720 | including the chief of staff to the CEO of Anthropic, makers of the Claw chatbots,
00:09:48.960 | think that this will massively impact the short-term outlook on employment.
00:09:54.160 | This article by that chief of staff, Avital Balwit, came out just two weeks ago.
00:09:58.720 | While I think the outlook isn't quite this stark, here's what she had to say.
00:10:03.920 | She predicted these next three years might be the last few years that I work.
00:10:08.560 | I stand at the edge of a technological development that seems likely,
00:10:12.080 | should it arrive, to end employment as I know it.
00:10:15.040 | And she makes the point that would have been relevant
00:10:17.760 | to that graphic designer we just heard from.
00:10:20.080 | The economically and politically relevant comparison on most tasks
00:10:24.240 | is not whether the language model, or I would say the embodied AI,
00:10:27.680 | is better than the best human.
00:10:29.360 | It's whether they are better than the human who would otherwise do that task.
00:10:33.200 | Doesn't have to be perfect in other words, just has to be a bit cheaper.
00:10:36.800 | She makes the somewhat common prediction by now that things like copywriting,
00:10:41.440 | tax preparation and customer service will be heavily automated.
00:10:45.600 | But let me give you two examples how the future is a bit more
00:10:49.120 | unpredictable than it can sometimes seem.
00:10:52.080 | First, I remember the frenzied reporting on this report from the think tank,
00:10:56.960 | the IPPR, here in Britain.
00:10:58.960 | According to the headlines at least, they were warning of an AI jobs apocalypse.
00:11:04.240 | But the very next day I contacted the lead author, Carsten Jung,
00:11:08.560 | and we had a detailed discussion for AI Insiders.
00:11:11.920 | First, he said head on that he was disappointed by the media's coverage.
00:11:16.080 | No, I'm not fully happy with how this is being covered,
00:11:20.320 | both our report but in general, because it can sound very scary.
00:11:25.200 | And I think just scaring people doesn't necessarily lead to
00:11:29.840 | incremental, thoughtful policy progress.
00:11:33.600 | When people talk about jobs apocalypse, I think some people might just switch off
00:11:37.920 | and throw up their hands and say, oh God, we're all doomed.
00:11:40.240 | Whereas what we try to do in the report is actually to say there's a range of scenarios
00:11:45.600 | and it's not some kind of external event like a pandemic that's like happening to us
00:11:50.880 | and it's all doom and gloom, but it's actually a thing that
00:11:54.000 | totally depends on decisions by policymakers, but also by organisations that implement AI.
00:12:00.000 | Then we discussed how a more likely medium term outcome is wage inequality.
00:12:05.520 | In short, low wages for many, but not for those who utilise AI to boost their productivity.
00:12:11.840 | So those that remain in work, their productivity will be hugely aided by AI.
00:12:16.400 | So you have this wage inequality aspect.
00:12:19.040 | But then of course, and I think this is also Sam Altman's point,
00:12:22.400 | is that profits are likely going to go up.
00:12:24.800 | So we have lower labour costs, AI likely is able to do things more cheaply.
00:12:29.360 | So profits will go up.
00:12:30.720 | So those that own companies will have higher returns.
00:12:34.400 | And so wealth inequality will likely go up.
00:12:38.000 | And the second cautionary tale about how AI's impacts say on jobs can sometimes be overhyped
00:12:44.480 | actually comes from open AI itself, albeit unintentionally.
00:12:48.400 | When we're talking about good things,
00:12:49.840 | we talk about customer service being revolutionised and productivity accelerating.
00:12:54.880 | But when the focus is on people using AI for nefarious purposes,
00:12:59.280 | suddenly the AI is kind of useless.
00:13:01.600 | This was a report released by open AI a few days ago
00:13:04.720 | about how some bad actors were trying to generate disinformation campaigns en masse.
00:13:09.840 | Open AI terminated those accounts,
00:13:11.760 | but gave a summary of the impact of these campaigns using the GPT models.
00:13:16.480 | There was no significant audience increase due to our services.
00:13:20.560 | Later on in the report, they say this.
00:13:24.560 | So far, these operations from places like Russia, Israel and China
00:13:28.560 | do not appear to have benefited from meaningfully
00:13:31.440 | increased audience engagement or reach as a result of our services.
00:13:34.960 | They basically describe how these guys came up with a load of spam,
00:13:39.440 | but people weren't buying it.
00:13:41.040 | For the most part, it was because the spam just wasn't very good.
00:13:44.560 | I don't know, it might be me, but I just find it a little bit ironic
00:13:47.520 | that when we're talking about a negative use of the technology,
00:13:51.200 | the party line is that the models are kind of useless.
00:13:54.560 | Of course, what we really need are better benchmarks.
00:13:57.520 | And so I was pleased to see this initiative from Scale AI.
00:14:01.200 | They describe these benchmarks and leaderboards that can't be gamed,
00:14:05.600 | are uncontaminated and unbiased.
00:14:08.480 | According to these benchmarks, at least,
00:14:10.320 | GPT 4.0 is not a million miles ahead of other models.
00:14:14.160 | This initiative reminds me, at least,
00:14:16.000 | how we should always benchmark models on our own use cases,
00:14:20.320 | because leaderboards chop and change quite a lot.
00:14:23.600 | Notice how the table on the left is quite different
00:14:26.960 | to the one that OpenAI put out on release.
00:14:29.520 | That initiative, by the way,
00:14:30.800 | isn't the only reason to be optimistic about benchmarks,
00:14:33.760 | which I covered in this video on Patreon.
00:14:36.400 | In short, though, I think just about the only thing we can all agree on
00:14:40.720 | is that the future is about as unpredictable as it has ever been.
00:14:45.360 | In terms of at least referring to AI in academic papers,
00:14:49.120 | you can see the recent exponential increase across virtually every field.
00:14:54.160 | How this all actually plays out, though,
00:14:56.320 | in the real world, in society, with jobs, with embodied physical AI,
00:15:01.280 | we simply don't know.
00:15:02.560 | Thank you, though, for being here with me as we watch it all unfold.
00:15:06.480 | Have a wonderful day.