Cal Newport

You can give it an input and it can give you a token.

You're talking about a program that is calling Mythos and then taking actions based on what it called.

And this is really what we're talking about with agents is the digital brains or LLMs.

And then you write a program

that will say to the LLM, give me a plan for doing X. And then the LLM spits out what seems like a reasonable text that seems like a reasonable plan.

And then you execute that plan.

The program executes that plan on behalf of the LLM.

And I wrote about this.

Yeah, and I wrote about this earlier this year.

LLMs are bad, you know, as a digital brain are bad planners.

It's not really, you're not going to get consistently usable plans because what an LLM is actually trying to do is finish the story you gave it.

So all it wants to do is produce a story

that sounds reasonable.

So it's giving you reasonable sounding plans.

Like, yeah, that's what a plan for doing this would more or less sound like.

But what it's not doing is actually doing step-by-step evaluations.

It doesn't have a clearly isolated goal that it's trying to measure how close you're getting to it.

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