Ben Todd scores his AGI-by-2030 forecasts: mostly right, but slower than expected
ben_j_todd · x · 2026-09-22
Ben Todd reviews his March 2025 essay "Will we have AGI by 2030?" and finds it held up well, though slightly slow.
What landed: all four drivers of progress (pretraining, RL, test-time compute, agent scaffolding) kept working. His end-2026 predictions vs reality:
- FrontierMath: "50% to saturated" → saturated
- METR 50% time horizon: 6h → 16h+
- Humanity's Last Exam: "40% to saturated" → 54%
- SWE-bench Verified & GPQA: saturated → saturated
- Frontier revenue: forecast 4x/year → actual 6x/year
He was proud of calling agents the next stage of scaling, identifying verifiability as the key axis, and not predicting mass unemployment. He argues that on maths, beyond-human-level problem solving may already be here.
What he'd change: he underweighted how compute shifting into RL would slow the pretraining trend and maybe overemphasized test-time compute; he now gives more weight to a gradual slowdown after 2028 putting AGI at 2030-2035 — though his probability of AGI before 2029 has gone up.
Related event: Ben Todd Reviews His AGI-by-2030 Forecast: Mostly on Track but Slower(2 posts)→
More from AGI Musings
- Podcast: Epoch AI researcher on RSI, robotics, China model gap, and open vs closed safety — natolambert · 2026-09-22
- Gary Marcus: near-term AI risk is deepfaked disinformation, not extinction — GaryMarcus · 2026-09-22
- Reality is moving at 70-90% the speed AI 2027 predicted, re-evaluation finds — ben_j_todd · 2026-09-22
- AI Risk Researcher: My Concern Stems From Ignorance, Not Prescience — danfaggella · 2026-09-22
- Google CISO Phil Venables: Defense Against AI Attacks Needs Autonomy, Not Just Speed — philvenables · 2026-09-22
- AI in GTM Will Decimate Teams That Were 'Tolerated, But Not Liked' — thedealdirector · 2026-09-22