Ben Todd: 3 factors explain AI hitting 10,000-hour math tasks without superexponential growth
ben_j_todd · x · 2026-11-15
Benjamin Todd explains why AI can now do math research that would take a human expert 10,000+ hours, despite METR's 80% time horizon being only 3h in April: (1) math was already ahead — 2h IMO problems vs 30min measured in July 2025; (2) math capability grows 1.5–2x faster than METR overall (40x/year vs 8x); (3) progress accelerated another 2x in 2026. Multi-agent systems added a further 10x boost. No superexponential needed — current trends are already extreme.
Related event: Ben Todd: AI Math Gains Explained, Superhuman AI Possible in 18 Months(2 posts)→
More from AGI Musings
- Gary Marcus: Doomers who can't tell an LLM from Claude Code are arguing in a muddle — GaryMarcus · 2026-11-15
- MotherDuck CEO: AI agents will collapse data teams into one Data Intelligence Engineer role — teodorio · 2026-11-15
- 2026 AI course guest lineup: Schulman on reward hacking, Pachocki on recursive self-improvement — himanshustwts · 2026-11-15
- Patrick Collison sparks debate: does post-scarcity AI break status games? — sebpaquet · 2026-11-15
- Runway CEO cites 1888 Kodak backlash as analogy for today's AI art haters — c_valenzuelab · 2026-11-15
- Ben Reinhardt: people fear sharing positive visions of the future — Ben_Reinhardt · 2026-11-15