I don't think it really is, especially not if it's turned into a system, with multiple prompts, verification, etc. Humans have problems with IMO problems, and

I don't think it really is, especially not if it's turned into a system, with multiple prompts, verification, etc. Humans have problems with IMO problems, and this kind of kernel translation is a problem which is easier to humans, where there's more probably actually more data and a problem where the system can get feedback by simply running it and measuring memory use, runtime etc. It'd be a system and no one has developed it, but I think it can be done with present LLMs as a core mechanism. They just need to be trained with RL on this specific problem. Anyone with a good LLM, from Google to Mistral could probably do this, but it'd be a project.