Cal Newport

And we've never encountered that before.

Like, I didn't realize that, you know, no, it's a business.

Yeah.

And the credulity with which we're taking these claims, like mythos is, I think the most important story there is, yeah, this is another example of what I wrote last summer about AI has hit a bit of a wall in the sense that all of the improvements that have come really since over the last two years have almost all been either on post-training or more importantly on the harnesses that you built.

So it comes in the software you're building.

What is a harness?

Think of it as like a computer program that can do stuff.

You can talk to it, it can do stuff, and it uses – it'll prompt or talk to an LLM as like its digital brain.

So the harness might actually be able to touch your file system, write the files, compile code, move things around.

But to figure out what actions to take, it will also then prompt an LLM and say, okay, what should I do next?

And you can put it on different – Is that just a wrapper?

Yeah, it's a wrapper.

All of the progress in coding agents since about a year has come, especially starting this fall, has come from better wrappers, better harnesses.

It's all in let's build better, just hand coding, no machine learning, no intelligence, no Skynet here, but just hand coding these programs that we'll call LLMs.

Let's just keep tuning and tweaking those to be better and better.

And of course, the programmers building those particular programs, they're building them to do their type of work.

So it's a field they understand really well.

So they can really just sit here and twist and tune.

And also like programmers are very adaptable.

They like tools and they'll adapt around the weaknesses or not.