AI researcher: single-dimension "capabilities" evals have poor predictive power
joshua_saxe · x · 2026-11-10
AI researcher Joshua Saxe argues that discourse collapses AI into a single dimension—"capabilities", "progress", "intelligence"—obscuring the poor construct validity of offline evals, which are just averages over tasks.
His take: these averages have very poor predictive power over real-world outcomes, so the bad abstractions will evaporate. The future he sees is domain experts ramping up on AI and explaining back what AI trajectories mean in their own fields—how AI is changing a discipline, how to accentuate its affordances and mitigate risks, and what comes next. He calls this "AI social impacts as normal epistemology" and says it can't come fast enough.
More from AGI Musings
- Design engineer suspects AI agents could easily replace his job with good data — CreamPiBaroness · 2026-11-10
- Humanity's Greatest Achievement May Be an Intelligence That Surpasses Us — VraserX · 2026-11-10
- 430,000 AI Micro-Dramas in 8 Months: A Visitor's Take on China's AI Content Boom — JacquesThibs · 2026-11-10
- FOCS 2026 Grad Student Feedback: US Students More Optimistic, Many Feel No Agency — thegautamkamath · 2026-11-10
- Reddit essay: most of what people hate about AI comes from the humans steering it — teleport66 · 2026-11-10
- Underrated question: what do the diverse forms of recursive self-improvement look like? — repligate · 2026-11-10