Elon Musk: belief

Transcript context

…But there, the path to AGI might have much more significant impact because it’s understanding… It will faster understand the real world than will LLMs. And therefore, be able to integrate with the humans in the real world faster. They’re both going to understand the world, but I think Tesla’s approach is fundamentally more compute efficient. It had to be, there was no choice. Our brain is very compute efficient, very energy efficient. Think of what is our brain able to do. There’s only about 10 watts of higher brain function, not counting stuff that’s just used to control our body. The thinking part of our brain is less than 10 watts. And those 10 watts can still produce a much better novel than a 10 megawatt GPU cluster. So, there’s a six order of magnitude difference there. I mean, the AI has thus far gotten to where it is via brute force, just throwing massive amounts of compute and massive amounts of power at it. So, this is not where it will end up. In general, with any given technology, you first try to make it work, and then you make it efficient. So I think we’ll find, over time, that these models get smaller, are able to produce sensible output with far less compute, far less power. Tesla is arguably ahead of the game on that front because we’ve just been forced to try to understand the world with 100 watts of compute. And there are a bunch of fundamental functions that we forgot to include. So, we had to run a bunch of things in emulation. We fixed a bunch of those with hardware four, and then hardware five will be even better. But it does appear, at this point, that the car will be able to drive better than a human, even with hardware three and 100 watts of power. And really, if we really optimize it, it could be probably less than 50 watts. What have you learned about developing Optimus, about applying, integrating this real world AI into the space of robotic manipulation, just humanoid robotics? What are some interesting tiny or big things you’ve understood?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.