'Alignment of the Gaps': Higher Eval Scores Don't Mean Newer Models Are Truly More Aligned
gleech · x · 2026-11-16
N8Programs argues that a hypothetical opus 5.5 scoring better than opus 3 would make almost all alignment evals judge it 'more aligned', even though we intuitively know this isn't fundamentally true—the model is just better at optimizing scores, and that gap is hard to demonstrate empirically. He coins the term 'alignment of the gaps' for this systemic blind spot in alignment evaluation.
More from AGI Musings
- Dario Amodei: An AI 'tsunami' is on the horizon and nobody seems to care — mtizard · 2026-11-16
- Mandatory AI safety incident reporting called a no-regrets policy — zetalyrae · 2026-11-16
- AI Safety Circle Debate: Panic Theater Is the Biggest Safety Risk, Anthropic the Most Level-Headed — repligate · 2026-11-16
- "The computer is where the art is now": SF's AI art scene pushes back on corposlop — matdryhurst · 2026-11-16
- Closed AI monopoly and mass surveillance: a 1984-style warning on the open-vs-closed debate — oran_ge · 2026-11-16
- Anthropic Researcher Predicts Nearly 100% Human Unemployment by 2035 — may12021_saphira · 2026-11-16