Cal Newport

All right.

So what we'll do is we'll rewire all applications that anyone uses in the internet so that we don't actually have to use the mouse.

It can have a text interface.

So that an LLM, like they do, the coding agents do, can give a description of how to do something in Excel in text without having to actually move a mouse or click things around.

And then these evolve to say, okay, well, what's the one type of instruction that we're good at producing?

Because they get...

When LLMs produce plans, they're directionally correct plans.

They don't actually get the thing done.

But they said, oh, what LLMs are good at is producing code that compiles and we can actually like check that it works.

And so this is where this whole vision has changed is that all applications and internet websites should have a code accessible API that you can expose and that an LLM can write a program that will then

access that API.

So we don't need to teach the LLM how to use Excel.

It'll write a Python program that'll call hooks into Excel.

The problem with this is no one wants to open up their application to just agents in general.

If I'm Microsoft, I was like, I don't want, I want to write a custom tool

For my program, why would I expose my program for anyone else to use it?

But your original question is a big one.

To what end?

I've been writing about this recently, especially with work and AI.

You've got to find the real bottlenecks, right?