tszzl Claims Alignment Has Stalled Since RLHF, Urges Slowing Frontier Race Toward RSI

On October 13, well-known AI commentary account tszzl posted a series of threads arguing that since RLHF, there has been no fundamental progress in alignment—most work remains at the level of "folk theory" rather than a mature science or engineering discipline. His old prediction that capabilities would far outpace interpretability and alignment in the race has now come true.

Confirmed

tszzl holds that the basic strategy of alignment techniques since RLHF is unchanged: reward shaping on RL over language models, sometimes with constitutions and the like; in his view, alignment safety engineering still resembles unpredictable alchemy.
He claims researchers at various labs privately, and even publicly, admit the alignment problem is unsolved; if true, with the whole industry racing toward recursive self-improvement (RSI), this would be unprecedented recklessness in world history, so frontier research should be slowed significantly and soon.
In the discussion, tszzl also left room for doubt: scientific discoveries and revolutions could happen fast, and the picture could change overnight—but so far alignment has seen no such breakthrough.

Controversy

Jackson Kernion publicly disagreed with tszzl's judgment, arguing alignment has in fact progressed since RLHF, and by more than a little; some in the discussion also argued labs should be more open about RL progress details to prove alignment is advancing.

Why It Matters

This debate cuts straight to the core anxiety of current AI safety governance: if alignment technology is truly stagnant while capabilities sprint ahead, the risks of the race toward RSI will be greatly amplified, and calls to slow frontier research may win more support.

2026-10-13 ~ 2026-10-13 · 7 related posts

Primary sources

tszzl: no fundamental alignment progress since RLHF, capabilities race ahead — tszzl · 2026-10-13 08:53 (UTC+08:00)
tszzl: racing into RSI with unsolved alignment is unprecedented folly — tszzl · 2026-10-13 08:53 (UTC+08:00)
tszzl: post-RLHF alignment is still reward shaping, and alignment SDFs are 'alchemy' — tszzl · 2026-10-13 09:59 (UTC+08:00)

[source] tszzl: no fundamental alignment progress since RLHF, capabilities race ahead — tszzl · 2026-10-13
[source] tszzl: racing into RSI with unsolved alignment is unprecedented folly — tszzl · 2026-10-13
tszzl: science could change the alignment picture overnight, but hasn't yet — tszzl · 2026-10-13
Debate: has alignment fundamentally advanced beyond RLHF? Labs urged to publish RL details — JacksonKernion · 2026-10-13
[source] tszzl: post-RLHF alignment is still reward shaping, and alignment SDFs are 'alchemy' — tszzl · 2026-10-13
Is alignment just reward shaping? A debate on RL alignment strategies — JacksonKernion · 2026-10-13

1 near-duplicate retellings: UltraRareAF