Liz Hoffman: You have described Prometheus, a new AI startup as a bet on physical AI for the industrial manufacturing.
Jeff Bezos: AI for atoms.
Liz Hoffman: AI for atoms. Can you give us some examples of what that actually is in practice?
Jeff Bezos: For the things we're used to, the LLMs and so on, they're very good symbol manipulators. But you really can't ask them to design an object for you in a detailed way. I'll give you a couple of examples, but the way we train them, these large, frontier models, is let them read the entire internet, so they have the corpus of human knowledge.
And they can solve problems that can be solved symbolically. So they're good at math, they're good at law, they're good at a lot of things. But there's a lot of things you can't learn by reading.
So I could read a thousand books on how to be a great gymnast, and I would still be a terrible gymnast. Or a thousand books on how to be a dancer, and I still couldn't dance. And so there are a lot of things like this, and engineering is one of them.
If I, give you a little thought experiment. Let you pick your six favorite physicists of all time, and pick Isaac Newton, and Einstein, and Maxwell. Take the six brilliant physicists, put them in a room together and say, "Build me a toaster oven."
They would have no idea how to do it. But understanding Maxwell's equations doesn't help you build a toaster oven at all. And this is true because knowledge operates at different levels of detail, so you have to train specialized models if you want to be able to do engineering things, and that's what we're doing at Prometheus.
and so it's very exciting. we have a brilliant team. We have two hundred and fifty people now.
Two hundred and fifty people. We raised something, just approximately fourteen and a half billion dollars. We've committed a lot of that to compute already.
It's very compute intensive. We build neural surrogates for numerical simulation. and we have one single multi-physics model.
And then we can do things, which we've already demonstrated. The team is just heads down, working hard. There's a new little miniature miracle every day, and it's incredibly fun to watch.
But, we're designing, analog chips, radios, and other things that would probably take a team of, I don't know, ten to twenty people six to nine months to do. and we're doing it in twelve hours. And similarly with valves.
One of our teammates there at Prometheus is from JPL and had done some work on the Mars helicopter. And so he asked, the Prometheus tools to redesign the Mars helicopter just with a prompt, and it was amazing how well it did and how far it went. And it started out with all these different designs.
Should it be a quadcopter? Should it be a counter-rotating helicopter? Should it be a fixed-wing aircraft?