Percy Liang to AI labs: share non-vacuous info outsiders can actually evaluate
lateinteraction · x · 2026-11-16
Stanford researcher Percy Liang, quoting a discussion of Anthropic's reported capability gains, clarified he is not accusing Anthropic of embellishing: he believes the gains are substantial. His point is that labs should take outside evaluators seriously and share only non-vacuous information they actually want evaluated — a public critique of how model-release evals are disclosed.
More from AGI Musings
- Challenging the AI takeover threat model: alignment is reasonably easy — teortaxesTex · 2026-11-16
- Dwarkesh's 'Recursive self-improvement has started' hits 539K views in 2 days — shashib · 2026-11-16
- If AI replaces junior work, firms still need a path to build reviewers' judgment — TansuYegen · 2026-11-16
- New Interview: Stephen Wolfram on the Ruliad and Why We Can't Predict It — romanyam · 2026-11-16
- Dan Robinson: AI likely able to cause $1T damage before it can fully do your job — austinc3301 · 2026-11-16
- Researcher: model misalignment demos without input info are manipulative or sloppy — giffmana · 2026-11-16