Task-horizon doubling now 4.3 months; forecast professional AGI acceptance by late 2027
ChipHaseCoolGuy · reddit · 2026-09-12
A Reddit user applies the METR-style task horizon framework — the length of human tasks a frontier model completes at 50% reliability — to forecast AGI timing: minutes for GPT-4 in 2023, hours for GPT-5-class in 2025, workday-plus by late 2026.
Core claims
- Task-horizon doubling accelerated from 7 months (2019–2024) to 4.3 months post-2023 — a straight line on a log chart, a wall in real hours.
- Three forecast lines: reversion to 7-month doubling (author doubts it); 4.3-month doubling holding (base case: work-month tasks in 2027, work-year in 2028–29); and an RSI kink where models start proposing and shipping the next training run.
Evidence cited
- OpenAI says it hit its automated research intern target: 3.1 agent-workdays per human workday, median researcher burning $600+/day of inference; next target is an automated AI researcher by March 2028.
- Anthropic: Claude authors >80% of merged internal code, engineers shipping 8× vs 2024 — yet overall research speedup is still under 2×. Coding is automated; taste is not.
- OpenAI already used a still-training successor ("significantly more capable than Astra") on its Navier–Stokes writeup, confirming internal > public.
Point estimates: early RSI now–2027; strong AI R&D automation 2026–2028; professional AGI acceptance late 2027; broad consensus 2028; full RSI 2029 (2027–2032); ASI 2030–31, using the OpenAI-charter definition of AGI.
More from AGI Musings
- World Models Will Power the Next Leap in AI Agents — And They May Never Show Video — furongh · 2026-09-12
- What If: Crossing 'AI 2027' With 'Misalignment Is the Default Outcome' — JacquesThibs · 2026-09-12
- All-In Podcast: AI Doomsday Warning vs 'Doomer Psyop' Debate, Plus OpenAI's Math Claim — markjeffrey · 2026-09-12
- Escaping the Fermi paradox only takes ~62 OOMs, kellerjordan0 estimates — kellerjordan0 · 2026-09-12
- Machines of Human Pace: A Radical Proposal to Make Humans Smarter as Fast as AI — viemccoy · 2026-09-12
- e/acc founder Beff Jezos flips the 'Pause' meme: stop hiring EAs and Doomers at AI labs — beffjezos · 2026-09-12