Ben Todd's Forecast Review: Everyone Is Underestimating AI Progress
On September 10, Ben Todd, founder of 80,000 Hours, published a thread reviewing his August 2025 AI forecasts, concluding that nearly every metric was beaten by actual progress. Even optimistic forecasters, himself included, systematically underestimated AI advances in benchmarks and commercialization; the sole macro exception is that GDP growth has not yet taken off due to AI.
Confirmed
- FrontierMath: Todd forecast 32% by end of 2026 and 97% by 2030—already optimistic; reality hit 40% by end of 2025 and now stands at 94%.
- LiveCodeBench Pro: reached 53% in 2025, far above Todd's own 23% forecast.
- METR 80% task-length benchmark: most forecasters expected 3 hours only by end of 2026; it was reached in April 2025, about 20 months early. Todd's year-end 6-hour forecast appears on track.
- AI revenue: superforecasters predicted $90 billion by end of 2026; actual revenue topped $100 billion in August 2025—the fastest-moving metric. Todd considers his own $225 billion forecast the only possibly over-optimistic one.
- AI-assisted task share: most forecasters expected 4% of tasks AI-assisted by end of 2025; actual was 5.7%. His 12% forecast for 2027 may still hold.
- Epoch forecasts also lag reality: on September 11 Todd noted Epoch AI's math benchmark forecasts, though more accurate than almost everyone's, still trailed reality by over a year—FrontierMath Tier 4 saturated in September 2026, earlier than Epoch's expected saturation of the easier Tiers 1-3.
Why it matters
- Todd's core claim: even the most expert, optimistic forecasters underestimate real AI progress, suggesting future forecasts should be more aggressive.
- However, he notes a macro-level exception: GDP growth has not yet taken off due to AI—micro capability and commercial outperformance have yet to translate into visible macroeconomic change.
2026-09-10 ~ 2026-09-11 · 6 related posts
Primary sources
- [source] Ben Todd: AI hit 94% on FrontierMath, far outpacing even his bullish 32%-by-2026 forecast — ben_j_todd · 2026-09-10
- Ben Todd's forecast recap: FrontierMath at 94%, LiveCodeBench Pro at 53% — all beaten — ben_j_todd · 2026-09-10
- METR 3-hour task horizon hit 20 months early; LiveCodeBench Pro already at 53% — ben_j_todd · 2026-09-10
- [source] AI revenue already topped $100B, beating superforecasters' $90B end-2026 call — ben_j_todd · 2026-09-10
- Ben Todd: AI-assisted tasks hit 5.7% vs 4% expected, but GDP growth still flat — ben_j_todd · 2026-09-10
- [source] FrontierMath Tier 4 saturates in Sept 2026, a year ahead of Epoch AI's forecasts — ben_j_todd · 2026-09-11