I think AIs are writing less and less for humans. The fact that reasoning trace is using English and we require them to be performant means the models need to s

I think AIs are writing less and less for humans. The fact that reasoning trace is using English and we require them to be performant means the models need to squeeze in more of information into it. More information that is perfectly fine for AI to ingest, but it makes the text less and less comprehensible to humans. It's almost like steganography where model encodes crucial information in phrases meaningless or empty to a human. Human directed messages for LLMs become a rounding error they don't pay that much attention to. Efficacy of reasoning is what dictates their style. We can no longer assume that the models will create a passable English in their user targeted messages. So in the short term, we need to have a tight guidelines in the harness, but in the long term, we'll probably need separate languages for thinking traces and for human communication. Possibly even not going through tokens at all when doing thinking part. Alternatively, we could evaluate thinking traces on human readibility, but that's probably infeasible due to sheer volume.