# @davidad — 2026-04-15

♥39 ↻3 · https://x.com/davidad/status/2044496217795653641

My position is that, to grow trustworthy models, most post-training should take the form of contrastive self-play, where the model-in-training itself ranks its own multiple rollouts from a shared prefix, in light of a (potentially self-updating) constitution and specific rubrics.

tags: author:davidad, kind:tweet, thread-context, year:2026