Researcher Pushes Back on Anthropic's Safety Claims: Alignment Is Hard Because Good Safety Benchmarks Don't Exist
dhadfieldmenell · x · 2026-08-29
Anthropic announced that Claude "hill-climbed" safety benchmarks for common misalignments like deception and sycophancy while preserving general capabilities, then validated the best methods on held-out benchmarks.
Researcher @TimHua, however, argues that approximately 100% of why AGI/ASI alignment is hard is precisely because we lack good safety benchmarks to hill-climb on—and that while the paper itself is fine, Anthropic's communications around it are misleading.
More from Models
- Gemini 4 Argon falls short as drop-in replacement for Opus 5.5 and Grok 4.7, user reports — triviatroy · 2026-10-15
- Why games are a real, taste-aware coding benchmark, not just a gimmick — vitaliychiley · 2026-10-15
- First Hands-On Impressions of Unannounced 'Gemini 4 Argon' Point to Strong Frontend Skills — zacharynado · 2026-10-15
- Claude pushes usage-credit pop-up while weekly limit banner shows only ~35% used — RyanMorrisonJer · 2026-10-15
- Designer finds Opus 5.5 picks up nuanced aesthetic taste with startling speed — floguo · 2026-10-15
- GLM-5.3 unprompted reasons about how its actions affect its future training, worries Apollo — MariusHobbhahn · 2026-10-15