I have a writeup I want to publish about it.

- I felt like Tony Stark inverting Möbius strip to solve time-space navigation with AI.

- Any idea or gut feeling I had, I could verify very fast.

- I had to write a mini Genesis emulator to test the pipeline and it was very pleasant to direct the LLM to write it for me (with clean code and architecture).

Personally, I find LLMs really fun to work with, and the more capable they are the more fun it will get. It's one hell of a force multiplier.

It is fun in a totally different way than programming is. The "fun" or "joy" isn't coming from the sense of accomplishment of writing code or building software per se. Now it feels like it is just the joy of being able to prototype or debug or reverse engineer literally anything at a pace way beyond what you could've ever hoped to before. Have you ever dreamed of developing some software, but realized it was simply too big of a scope for one person? It might not be, anymore, even if you don't go full vibecoder.

The only real problem I ever had with this was the quality of the code. Sometimes it doesn't matter, like for pure prototypes or even reverse engineering, but sometimes it does. My heart sank a bit when I tried Fable to find that it was much better at long tasks and more ambitious in general, but I still didn't like the code much more than Opus 4.7 or 4.8. However, I'm pretty pleased with the code quality of GPT 5.6 Sol. This is a silly thing to care about but I love that by default it doesn't print code comments like a CVS receipt printer adding all kinds of pointless crap that is either apparent from context or irrelevant outside of the conversation you were having. Generally I've also found it to be super good at debugging, usually able to zero in on the actual problem with surprisingly little effort, even with not much context; I wish my debugging foo was that good. I gave Codex a spare machine to test something with real hardware and it has done a fantastic job making use of it.

Then of course there's the problem that I really don't like OpenAI or Anthropic very much. But it seems the news is good there too: even if we assume open weight Chinese models are at least less than 12 months behind on coding performance, then by next year we should have quite a lot of options for how to go about things. Evidence suggests 12 months is probably an over-estimation, and even now I use GLM 5.2 quite a lot at work since it is simply good enough. So... we're getting there too.

Obviously it's a little scary to think the world may need fewer programmers when you are one, and this is maybe not the way that I wanted programming to become accessible and democratized, but beggars can't be choosers. If the future is me being able to run an army of virtual programmers on a GPU cluster, I am interested in seeing what cool shit can be done with it. And when the hype and doom cycle finally ends, I hope everyone else will realize how fucking cool it is, too.