Researcher slams OpenAI's redefinition of alignment as just being more useful
nabla_theta · x · 2026-09-24
nablatheta argues that redefining alignment to mean "more useful" has been a disaster for the alignment discourse, quoting OpenAI's 2026 claim that larger computational depth improves alignment because models generalize more flexibly—including from alignment data—and better-generalizing models are more aligned and hallucinate less. The post spotlights concerns that safety terminology is being diluted into product metrics.
More from AGI Musings
- Mistaking Language for Reality: Why 'Evolution's Goal' and 'AI Scheming' Are Labels, Not Truths — paraschopra · 2026-09-24
- Gary Marcus: companies can't handle current agents, let alone superintelligent AI — GaryMarcus · 2026-09-24
- 1800s: State Monopoly on Violence; 2030s: State Monopoly on ASI — MillionInt · 2026-09-24
- Yudkowsky's 'If Anyone Builds It, Everyone Dies' returns to NYT bestseller list — ESYudkowsky · 2026-09-24
- Why lab insiders are more AGI-pilled: they see models succeed on untrained tasks — kellerjordan0 · 2026-09-24
- Reddit debate: autonomous transport, not chatbots or robots, will make people richest — StrategicHarmony · 2026-09-24