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…
Posters: davidad. davidad
2 likes recorded by the source corpus.
Topics: AI safety