An AI researcher publishes detailed personal probability estimates for when AI will reach human-level and superhuman intelligence, grounding forecasts in observed progress patterns and identifying key uncertainties like recursive self-improvement and regulatory timelines.

What happened: The author defines AGI as AI capable of doing 90% of knowledge work jobs and estimates a 25% chance by 2027, 50% chance by 2034, and 75% chance by 2045. They argue AI is already smart enough in raw capability but limited by confusion, lack of agency, poor situational awareness, and hallucination—traits improving at rates the METR time horizon graph and related benchmarks suggest will cross human level around the early end of that schedule.

Why it matters: The author identifies the diffusion gap (time between AGI capability and actual deployment across jobs) as potentially longer than the capability gap itself, with a 50% chance it lasts less than 10 years. Regulation is cited as a major factor extending this timeline, though early-stage AI diffusion has already outpaced personal computers' early adoption in revenue growth—suggesting some precedent for faster rollout than skeptics expect.

What to watch: The author's median forecast has superhuman-level AI arriving (50% confidence) in less than 4 years after AGI—faster than the diffusion gap—meaning superintelligent systems could exist before human-range AI finishes deploying. Recursive self-improvement is flagged as the largest uncertainty in the model, with unknown effects on acceleration.