Dean Ball: Open-ended long-horizon domains hit a wall near human level — verifiability is the crux
deanwball · x · 2026-11-13
Dean Ball weighs in on the AI "walls" debate, arguing that any semantically open long-horizon domain plateaus near human performance: pretraining imitates humans, and RL only exceeds them when answers are verifiable.
Key points:
- Pretraining loss measures the map, not the territory. With fuzzy data, loss hits a floor; with systematically biased data, lower loss just reproduces the field's errors.
- Updating against the territory requires grading against the world, which is slow in most open, data-limited domains.
- Verifiability is the crux. Noisy data just needs calibration and should transfer; systematic bias is a different beast — correcting it requires knowing the direction and magnitude of skew, which is nearly impossible when data is sparse and costly to check.
More from AGI Musings
- Sequoia's Konstantine: consumer agents took off first, $1B+ annualized purchases — Konstantine · 2026-11-14
- Mensch Meets EU Council President: Europe Needs Faster Financing and Coordination to Lead AI — arthurmensch · 2026-11-14
- AI animator: without China's Seedance and open-source MiniMax H3, no decent generative model would survive Sora's exit — astralmatrix · 2026-11-13
- Burkov: Amazon Is Drowning in AI Slop and Readers Pay the Price — burkov · 2026-11-13
- Penrose's OR theory remains the only concrete scientific proposal for consciousness — JosephJacks_ · 2026-11-13
- Ghana's WhatsApp AI math tutor shows learning gains at ~$5 marginal cost per student — PeterDiamandis · 2026-11-13