What I'm watching next

A few ideas keep pulling me back in. Here’s what I’m reading about and trying in small ways.

Running AI on your own machine

Most people use AI through someone else’s service. You can also run smaller models on hardware you control. I’m interested in which jobs a smaller model can handle well, and running one yourself changes how you think about it. You see what it needs, where it slows down and where it struggles.

I’ve learned more from trying it than from reading about it, mostly by getting stuck and getting unstuck.

Small models learning from big ones

This one I keep coming back to: a large model can help train a smaller one, so the small one gets surprisingly good at a narrow job. A big, general teacher and a small, focused student.

That one feels familiar. I spend a lot of my week teaching.

Several agents working together

Instead of one AI doing everything, you can set up a few, each with one job. One plans, one does the work, one checks it. They pass the task along like a small team.

The question I’d ask is the same one I’d ask about any team: who owns what, and who has the final say? With people, the good answer is usually written down somewhere. I suspect the same will be true here.

Open questions

• Does any of this make a task easier once you count the setup and the checking?
• Which jobs are small enough for a small model?
• When a few agents pass work along, how do you tell which one got something wrong?

I’ll keep notes here as I try things.