Dr. Peter Lebedev

It could speak English at you, but it wasn't very good English.

And now we're at the point where it can fully code up full apps, full websites.

Earlier this year, in January, Opus 4.5, which is, again, the Claude model, was used to hack into the Mexican government.

almost fully autonomously.

Like 90% of the operation was done.

This was a single person using a commercially available product to hack into a government.

And like, this is where we're at right now.

So yeah, fairly concerning.

Yeah, Dr. Ty, that is a great question.

So on the first part that I want to say is that, yes, the current AI models are large language models where they're predicting the next word.

They're predicting, technically it's the next token, but you can think of it as predicting the next word.

But one of the things that happens is when you learn a lot about the world through language, is that you actually do end up developing some kind of world model.

As in, if you go, I'm holding out this pen, if I let it go, it will... The next sentence, the next word is drop.

So the AI system has at least a little bit of an understanding of gravity just from that.

So I really do think that it's a little bit controversial, but a lot of people do think that LLMs already have world models to at least some degree.

And the other thing right now is that we're also not just doing next token prediction with large language models.

We're doing reinforcement learning on top of this, where we basically reward the machine for getting the answer correct or achieving its goal, which is how all the AIs that like...

beat people at Go and beat people at Atari and a bunch of other video games were trained.

But yeah, this whole world model thing is, yeah, a really, really big part of AI research right now, so.

There are a bunch of people, including one of the pioneers, this wonderful, incredible woman called Fei-Fei Li, who is working on, yeah, basically that.