Debate: has alignment fundamentally advanced beyond RLHF? Labs urged to publish RL details

JacksonKernion · x · 2026-10-13

An alignment debate broke out on X:

@tszzl argues there have been no fundamental advancements in alignment since RLHF — most of the field remains folk theory rather than a scientific or engineering discipline, and the prediction that capabilities would race far ahead of interpretability and alignment has come true.
Jackson Kernion pushes back: there have been advancements since RLHF, and it's a lot more than reward-model graded chat environments — but labs should publish more details of how these RL advancements work.

A representative exchange on alignment maturity and lab transparency.

More from AGI Musings