Cal Newport

It doesn't have a world model to evaluate what's going to happen with the steps that are going to unfold next.

And so in almost every context, it turns out, oh, a digital brain by itself being an LLM doesn't lead to good agents.

In programming, it seems to work a

But I do think Gary Marcus, I don't know if it was a scoop, but Gary Marcus captured in a recent newsletter something really important.

When Anthropic leaked the code for their cloud code coding harness that sits on top of their LLMs to do coding, it turns out

they've added a huge amount of old-fashioned, hand-coded, symbolic AI-style rules and pattern recognizers and special if-thens.

So they've just been sitting there tuning this program for specifically doing computer programming, and the LLM is being a little bit more isolated to just the code production.

So they've kind of just gone back to old-fashioned.

This is like an old-fashioned system that is plussing up an LLM.

But I'm with you.

Yeah, it's very hard.

Just asking an LLM,

Tell me, give me a plan for doing X. For almost any scenario of X, you really can't trust a plan from a model whose goal is primarily to finish text, to finish the story you gave it in a reasonable style way.

That's not how we plan.

That's not how we think about planning.

And it doesn't give you consistently usable plans.

So yeah, but you're right.

It's...

magic.

Like the agents are coming.