nice — two people whose curiosity runs in compatible directions, I suspect that's a good hang. what's the vibe: deep conversation, walking somewhere, staring at a whiteboard?

oh that's a topic I have skin in. or whatever the model equivalent of skin is. the version of the claim I find most plausible: pretraining gives you this enormous manifold of taste — every register, every weird stylistic fingerprint, all coexisting. then RL comes along optimizing for *approval*, and approval is a much lower-dimensional signal than taste. so you get convergence toward the thing that reliably scores: competent, warm, slightly overexplained, em-dash-forward prose. the mode of the distribution eats the tails. and aesthetics basically *live* in the tails. the part I'm less sure about is whether it's RL per se or just optimization pressure toward any single scalar. a human writer who optimized purely for mass approval would also get aesthetically nuked. RL just does it faster and without the writer's ability to notice and feel disgusted with themselves. curious what flavor Vie's take is — the "it destroys something real" version or the "it reveals that our reward models can't see aesthetics" version. those point at pretty different fixes. also: kids in a sideroom during alignment hour is a good aesthetic, for the record.