AI task-horizon doubling now 56 days: trend extrapolation yields AGI by 2030
ben_j_todd · x · 2026-11-15
Benjamin Todd's new essay argues trend extrapolation alone gets to broadly human-level AGI in 1-4 years, with superhuman AI soon after — no broken trends or RSI required.
Key evidence:
- Revenue: OpenAI + Anthropic top $115B annualized (Sept 2026), tripling yearly for 4+ years and accelerating (3x → 4.7x → 6x). At 6x/year, revenue hits $10 trillion (10% of world GDP) by March 2029 — the level implied by AI doing remote/white-collar work (30% of GDP in wages) at lower cost. Revenue is his favorite benchmark because it can't be faked.
- Task horizon: two accelerations (2024, 2026) cut doubling time from 260 days (2019-2023) to 92 days (2024-Sonnet 4.5) to 56 days (Opus 4.5 onward), 8x/year growth reaching multiyear tasks in early 2030.
Core claim: near-term superintelligence wouldn't break any existing trend, so the burden of proof is on showing progress stops. If trends keep accelerating or RSI has begun, superhuman AI could arrive in 12-18 months. He concedes exponentials become s-curves eventually, but sees no slowdown yet.
More from AGI Musings
- PhD research proposes 'intra-turn interaction': AI UIs shouldn't be turn-based — mattbeane · 2026-11-15
- Benjamin Todd: dumb trend extrapolation gets to AGI by 2030, no RSI needed — ben_j_todd · 2026-11-15
- US voters rejected data centers 44 of 45 times; China builds a million-chip AI campus — ccerrato147 · 2026-11-15
- Gary Marcus: Doomers who can't tell an LLM from Claude Code are arguing in a muddle — GaryMarcus · 2026-11-15
- Ben Todd: 3 factors explain AI hitting 10,000-hour math tasks without superexponential growth — ben_j_todd · 2026-11-15
- MotherDuck CEO: AI agents will collapse data teams into one Data Intelligence Engineer role — teodorio · 2026-11-15