Safety researcher updates ASI timeline: ~25% odds by end of 2027

On November 10, AI safety researcher austinc3301 posted a systematic update to his ASI timeline thread, revising his judgment on the pace and risks of AI progress. His core conclusion: AI's acceleration of research is likely real, and the window for action is shorter than he previously thought. TL;DR: He admits he has been consistently optimistic but his timeline has moved up significantly. He estimates current AI-driven research acceleration at roughly 1.6–1.9x, or about 2.5x in "AI 2027" terms, and accordingly assigns about a 25% probability to ASI by end of 2027 and about 50% by end of 2028, urging governments to slow down AI before the window closes. Confirmed (poster's views) - He has changed his skeptical stance toward frontier labs' self-reported uplift data, now believing current acceleration is "very likely real and primarily AI-driven," on the order of 1.6–1.9x based on his own model. - He calibrates various estimates using "AI 2027" as a common unit (i.e., the factor by which AI makes algorithmic progress faster than the no-AI counterfactual): Ryan reads about 4x from METR data; his own estimate is about 2.5x with an 80% confidence interval of 1.3–4.6x. That interval broadly fits the AI 2027 timeline (2x by end of 2026, higher in 2027). - On ASI probabilities: assuming no slowdowns or policy interventions that change the trajectory, about 25% for ASI by end of 2027 and about 50% by end of 2028. He stresses that RSI (recursive self-improvement) is highly uncertain and timelines are extremely sensitive to takeoff model parameters. - He found the relevant Anthropic reports "pretty scary" and offered criticisms. Not yet confirmed - Sources disagree markedly on the acceleration multiplier (about 4x from the METR reading vs. the author's 2.5x); the exact multiplier and RSI trajectory remain unsettled, and the author's ASI probabilities are subjective personal estimates. Why it matters - A safety researcher who was previously skeptical of labs' self-reported data now concedes the acceleration is real, with probability estimates pinned to specific years—giving policy debates a testable reference point. He believes the next year or two are the critical window for action and calls on governments to slow down AI in time, directly connecting to the debate over the timing of AI regulation.

2026-11-10 ~ 2026-11-10 · 6 related posts

Full story(15 episodes)→

Primary sources