so on the limits of AI for autonomous progress for context: a few days ago, I wrote a prompt for a "NanoHVM" - a toy version of my runtime, with 16-bit terms - and asked the AI's to write an evaluator. after a few loops of human-AI collaboration, I got a surprisingly fast implementation sadly, the file was getting very big and ugly, so I wanted to make it smaller. I set up a codex/claude/gemini session, and asked them to "make it shorter". first prompt worked, so I kept repeating the same request, over and over, for several hours. eventually, the AI's hit a wall. no matter how much I kept asking, they couldn't shorten it past a certain threshold. I wanted to understand why, so I read the file. to my surprise, it was clearly not optimal - tons of redundancies and bad abstractions - yet, it barely changing anymore. commit after commit, it still included the same techniques, the same approach, and the same redundancies. there were only small tweaks, but no major rewrite was taking place, and that was needed for further gains. so, I asked the AI to do just that. "do a complete rewrite this time" "reason about it fundamentally" "you're free to change anything" "stop doing incremental changes" "replace X by a whole new approach" yet, this was futile. these generic requests had no effect at all. hours later, the file still implemented the same approach. the AI was stuck in a local minima, and couldn't get out of it. so, I gave up, and decided to join. I took some minutes to think about it, and wrote a short prompt - about 500 tokens long - with specific directions. "replace X by Y, do Z in this way". somehow these ~3 paragraphs made something click, they immediately got unstuck, worked for ~1h30m straight, and came back with a file that was nearly 2x smaller *and* 10% faster. and it was a really beautiful, high quality file so, I guess the lesson is: modern AI's can't make autonomous progress. if you leave them working unsupervised, they will get stuck. that's a fundamental limitation of how these things work, at least for now. yet, if, every few hours, you take 5-10 minutes to revisit them, understand where they are, and inject some human knowledge - then they will just keep going, doing massive amounts of work and getting incredible results, as you spend your weekend playing some games I wonder what's missing for them to be able to get unstuck without my intervention. it feels like my prompt had nothing special. all ideas on it are concepts these AI's know, and it feels like they absolutely could've come up with these ideas on their own. yet, they just don't... why? anyway, I plan to open source this cute thing later next week. it is not exactly useful for most of you, but if you, by chance, have been looking for the fastest 16-bit pattern-matching engine in the world, this file may be just what you've been looking for!

"caution, this is your last edit" no worries, I'm confident that was the last typo

oh, fuck