AI Safety, The China Problem, LLMs & Job Displacement - Dwarkesh Patel - #979

**Chris Williamson** (0:00)
What do you think that we've realized about human learning and human intelligence from architecting AI intelligence?

**Dwarkesh Patel** (0:09)
There's this really interesting thing we've seen, where these AI models are making progress first in the domains that we think of as the archetype of, where humans have their primacy. So if you look at Aristotle, what does he say, what makes humans unique? Well, it's reasoning. Humans can reason, other animals can't. These models, these AI models, they're just not that useful if you've tried to use them for your work. They're useful in certain domains, but broadly, they're just not widely deployable. What is the one thing that they can do? They can reason. But obviously, they can't carry a cup of water, robotics isn't solved. They can't even do a job. They can't even do a white collar job. So there's this interesting thing called Moravec's paradox. Hans Moravec came up with this idea in the 90s, where he noticed that the tasks which are easiest for humans are taking computers the longest to solve. So we still haven't solved robotics yet. It's so easy for us to move around. Whereas the tasks which are quite hard for humans, like adding numbers, adding long numbers. Computers could do that in the 60s. And the logic there is that evolution has only optimized us for, let's say, the last million years, to be good at reasoning, to be good at arithmetic, to be good at these kinds of high-level abstractions. Evolutionists have spent four billion years teaching us how to move around the world, how to pursue your goals in a long-term basis, so not just do this task over the next hour, but spend the next month planning how to kill this gazelle.
That has been, I think, a remarkably accurate predictor of the places we've seen the progress.
They're automating coding. Coding, we thought of, was this thing that 0.1% of the population could do really well. That's the first thing that went below the water line. Yeah, just basic manual work might genuinely be the last thing that goes away.

**Chris Williamson** (2:02)
Right. Yeah, there's a difficulty in getting a robot to crack an egg, a particular difficulty in being able to do that, the right amount of tension to hold. Is there a, this may be outside of your domain of competence, but that's why we do podcasting, to talk about things that are outside our domain of competence. Is there a potential to use some sort of scanning technology to take an LLM type approach to teaching robots how humans move? You know, if you were able to track within a room exactly how a human was to just go about tasks, just feed that into a big fuck-off model, and then use that to, I guess, you can't really work out sort of force application just by looking. That would be something you'd have to fit. Maybe you could put someone in a suit. I don't know. I'm wondering if we've seen so much progress using LLMs in the world of AI. Robotics seems to be something that's still kind of pretty janky. I'm wondering if there are any principles that can be taken from the world of LLM that can be applied to robotics.

**Dwarkesh Patel** (3:03)
I mean, that's a great question, and many companies are working on it. My understanding is that it's difficult for the fact that there's not as much data, just what you mentioned, that the kind of data you need of like, what did it feel like?

**Chris Williamson** (3:16)
There's no internet for human movement.

**Dwarkesh Patel** (3:18)
Exactly, right? And even video is limited. And even if you have the video, it's not... With language, you have this thing of you are exactly doing the thing which the online internet text is, right? You are predicting the next token in text. You can predict the next thing in a video frame. That's not the same thing as robotics. There's also additional challenges from what I understand around the fact that video is harder to process than text, it's just like a lot more data. There's latency overhead, so if it takes you a while to process language, that's fine, you can go a token at a time. The real world just moves very fast. You can try to solve these issues by going in simulation.
You can have a simulation where you're trying to move things around, and in that domain, you can train an AI to be good at robotics, but the real world is just very complicated. If I crumple this thing, why does it bend exactly the way it does? It's just very hard to get that in simulation. Yeah, I think robotics is tough.