[littlestymaar]

Yeah “Reinforced learning with compiler feedback” in addition to RLHF.

For that reason I suspect that Rust could do pretty well in terms of its writeablility by an LLM, but there's the problem of the relatively small training codebase in the wild which make LLMs perform quite poorly in Rust at the moment (Llama3 70B being far worse than even ChatGPT 3.5 on that, but even ChatGPT 4 can't write nontrivial Rust code right now). So it seems hard to even bootstrap the process.

[yosefk]

OTOH when it writes buggy code, you tend to get decent error messages when trying to use it. In this sense Rust + LLM is way better than C++ + LLM

[littlestymaar]

True, but having to fix all the stupid mistakes they make is kind of ruining the point of using an LLM in the first place.