I Ran Claude in a Loop for 3 Months, and It Created a Genz Programming Language

One-shot a regexp, sure. One-shot a full compiler, not a chance.

Check out the repository - it has a full compiler that produces binaries, and a bunch of additional tooling. Doing that, even if he'd asked for a compiler for an existing language with no changes, is impressive.

It produced...something, alright. Something that can be replaced by a very small shell script.

3 versions... Yikes. But the output, while messy, is still impressive, and makes me wonder how much you could trim that down by giving more comprehensive instructions.

One of my fun experiments recently has been putting ChatGPT in conversation mode when I go for a walk. I recently had a 45 minute conversation where "we" fleshed out a multi-agent platform. I think a key is that you need to give each agent an "inner conversation" and criteria for when output from it gets copied to the other agents and the main chat, coupled with a process to regularly compact. I intend to set up a test system I want to run continuously, and given I enjoy working on compilers maybe I'll see how much cheaper you can do something like what OP did if you orchestrator a few agents with domain knowledge in specific areas.

I think I'd want to test a state of the art model, but it'd be fascinating to see how far you can get with Ollama as well - especially whether you can compensate for less smarts by just giving it far more runtime than I could afford with e.g. Claude.