RL is an "obvious" extension to code generating LLMs Try to compile, try to run, estimate reward, iterate. Possibly there is a cost issue, because sampling LL

RL is an "obvious" extension to code generating LLMs Try to compile, try to run, estimate reward, iterate. Possibly there is a cost issue, because sampling LLMs is more expensive than running starcraft?!