Cal Newport

It's better in specific tasks, or if I vibe code this, it looks better, I guess, and it seems more narrow.

And so, yeah, we're reaching an, there's a long answer to a short question, but we are reaching an asymptote on just pure fine-tuned LLMs as an engine for AI.

We're going to need more architectures.

It's going to take more time.

Yeah, but that won't happen because they tried and they don't know how to do that.

So it's not going to be just an LLM.

I mean, my prediction of the future of AI is I think what we're going to see, I think LLMs are very powerful, but what we're going to see is much more of hybrid models that are custom fit to particular problems where, okay, this system does this thing better than a human.

And in its guts, there's like an LLM in there, not a huge frontier model, but one that's like souped up and optimized for this particular type of thing.

But there's also like five or six other models.

There's an explicit world model.

There's a future predictor.

There's a policy network trained through reinforcement learning to try to evaluate situations to see what's good or bad.

There's a whole logic engine on top of this that hooks these together.

These are what I think the AI systems of the future are going to be like.

They're going to be bespoke, and there's going to be a ton of them.

So when we get to AGI, it's not going to be GPT-7 can do everything you ask it as well as a human.

It's going to be a world in which there's 10,000 different AI products, and you realize

Everything I can think of now, there's some product out there somewhere that can do this better than humans.

Just like there's AI that can play chess better than humans.

There's a different AI that can play Go better than humans.