Catch-all post
Here are some items that I haven't purged from my notes folder but also don't anticipate turning into full posts before the landscape changes again:
- If you're trying to learn math with AI, don't try to guess how it should teach you. Just tell it what you're trying to learn and let it decide where to start and what diagnostics to give you. I've gotten a lot of benefit from telling it to be less forgiving with me, but almost none from asking for syllabi or plans in advance.
- Here is Eric S. Raymond learning that he didn't miss what he thought he'd miss about the old way of programming. If you asked new Internet users in 2000 how they planned to use it, what they'd like about it, and what they'd miss about what it replaced, I think most of the answers would not be very accurate. Predictions about one's enjoyment of programming might not be much better.
- Today I made a little script, piggybacking on
pytest's collection, that shows me a random Python test in the codebase and opens Helix to it. This is a crude way of showing me what the AI has been up to, letting me improve those tests by hand, and generating material for improving the AI's instructions. Random or semi-random spot checks are a time-tested managerial technique, and there might be useful analogies to this in monitoring AI. - Here is Scott Hanselman on prompting and "semantic heavy lifting." Along the way, he gives an evocative definition: "Programming is the art of making the ambiguous incredibly specific through sculpting."
- There are lots of posts of the form: "I just spent X hours cleaning up after AI." I don't think these always prove what they think they prove. If you're replacing three weeks of work with AI, a full day of manual cleanup is still a very good progress-to-cleanup ratio, especially if you're still learning the technology. Now, I've gotten a lot of absolute junk from AI, and I'm sure lots of what seems like productive AI-assisted work turns out to be counterproductive. My point is simply: in many cases, if you take these AI-critical reports at face value, they seem to be evidence that AI worked quite well, not badly.
- Claude's
/fastmode is quite stunning. If you have free credits and haven't tried it yet, don't miss out. It's one thing to say that you know AI is going to get faster, and another actually to experience it. (Remember when the Internet was slow enough that deciding to load a Web page was making a significant time investment?) If nothing else, getting rewards faster encourages building the "you can just do things" muscle.