AI 2027 forecast faces fresh scrutiny

My take: AI 2027 is aggressive, but not reckless. 2027 was always the modal branch, and the newer benchmark data make the fast-progress case look stronger rather than dead.

• –METR's "50%/80% time horizons" are task-length thresholds, not plain accuracy scores; higher numbers mean models can handle longer tasks at that success rate.
• –TH1.1 and METR's later sensitivity note both show real movement, but also real noise: recent 50% horizons shifted, and plausible curve choices can move 50% and 80% estimates a lot.
• –The AI Futures Project's newer model pushed central coding-automation estimates later, to around November 2030 in the model median and roughly July 2032 all-things-considered, which is later than the original headline but still leaves 2027 in the distribution.
• –If the rumored Mythos model really is a jump in cyber and agentic capability, it would support AI 2027's thesis, but one release still wouldn't validate the whole AGI-to-ASI chain.