Over a year and a half after DeepSeek, it feels like we're slowly saturating what RLVR can do, and are back to RLHF instead.

Fable was a huge leap in terms of model persistence and raw intelligence, but it still had terrible taste for human writing and code architecture. It would constantly keep making decisions which would achieve the desired objective (and make the code correct), but would bite you n years from now, and n years from now isn't RLVR checkable.

Opus 5.5 has a very different "feel" than anything else I've seen in this generation, though GPT-6 does seem to be moving in a similar direction. They have finally solved the writing part, and architectural taste also seems to have improved significantly.

I did a review of some GPT 6 Sol's code with Opus 5.5 yesterday, and it went "the code is correct, but there's a bunch of things here that could be simplified, and the split of responsibilities doesn't follow your established architectural layers" (which was true and exactly what I've noticed myself when reading the diff). I don't think I've ever seen a model do this before and actually be on-point.