Thank you for the kind words. Agreed that writing is process oriented and I'll have to work on that. This experiment has been on the top of my mind for a long time and it feels good to finally share it. It's been a long process.

Analysing the shape of the output is an excellent idea, it would be interesting to see which abstractions surfaced, and which did not. There are a lot of programming paradigms in high-level language that I'm curious to see if they are there or not.

And I do gloss over libmuffintop, and I think it might be worthy of its own post. It is the thinnest cross platform abstraction layer I could come up with. It's implemented in LLVM IR directly. The main idea is to remove as many abstractions as possible between the prompt and the machine. I wanted to see if the abstractions are necessary for agents, it seems that they are not.