Helen Toner: AI capabilities will stay extremely uneven, and that matters
hlntnr · x · 2026-09-09
Helen Toner, former OpenAI board member now at CSET, published the transcript of her talk "Taking Jaggedness Seriously" from The Curve conference in Berkeley.
- Her core claim: two things are true at once — models keep getting better, yet keep failing at confusingly simple tasks (e.g., GPQA scores climbing fast while basic failures persist).
- She argues this extreme unevenness is durable and should shape expectations: weak performance in one area doesn't imply slow overall progress.
- The post includes a 26-minute talk plus 25-minute Q&A video, and notes she has taken a new role at CSET, slowing her Substack cadence.
More from AGI Musings
- Kaj Sotala argues alignment should favor wise-advisor designs over untested CEV — xuenay · 2026-09-09
- OpenAI claims 10,000 agents solved Navier–Stokes Millennium Prize problem in 88 hours — scaling01 · 2026-09-09
- Evitable founder David Krueger on what a superintelligence takeover moment would look like — DavidSKrueger · 2026-09-09
- A text's AI share is either 0% or 100%, argues economist — paulnovosad · 2026-09-09
- Metaculus forecasters now expect 'weak AGI' later than pre-ChatGPT, Mollick pushes back — emollick · 2026-09-09
- Nikita Bier: AI-generated renovation plans are basically construction-ready now — aarthir · 2026-09-09