RLHF paper never described instruct model training; insiders reflect on invalidated takes
jd_pressman · x · 2026-09-15
A discussion of cases where facts learned after the time invalidated early analysis: the RLHF paper never actually described how its instruct models were made, and Altman admitted GPT-4o glazed too much and said he'd fix it, then left it that way for months. Blanche Minerva calls this insane and says a retrospective of post-hoc revelations that invalidate early analysis is needed.
More from AGI Musings
- UChicago AI panel: Doomsday Clock at 85 seconds, experts agree race-now-guardrails-later is risky — WillRinehart · 2026-09-15
- AI's $1.1T data-center bet: productivity must grow 2.7x by 2030 to break even — nordicinst · 2026-09-15
- Reddit thread challenges AI slowdown opponents: what exactly do you disagree with? — nemzylannister · 2026-09-15
- LeCun Endorses 'Realism Vaccine' Against the 'AI Doom Virus' — ylecun · 2026-09-15
- AI's Most Underrated Capability: Amplifying and Combining Human Intelligence — anonymousecateer · 2026-09-15
- What Would Adam Smith Make of AI? Azeem Azhar Builds an AI Persona to Find Out — Exponential View (Azeem Azhar) · 2026-09-15