Quickly, There’s No Time

Open original

Zvi Mowshowitz rounds up how the METR report on Anthropic's AI progress pushed forecasters like the AI Futures Project and Paul Christiano toward much shorter timelines for superhuman AI researchers. He argues the updates are directionally right, that 2027 needs a real chance in anyone's plans, and that current safety plans aren't ready for it.

  • METR estimates Anthropic's AI progress is running about twice its historical rate, with perhaps 80% of that acceleration attributed to AI itself.
  • The AI Futures Project moved its median for a Superhuman AI Researcher from around 2030 to 2027, and Daniel Kokotajlo says the intelligence explosion now looks about five months ahead of the AI 2027 scenario.
  • Zvi puts the chance of a superhuman AI researcher in 2027 under 50% but not far under, and says anyone treating a 2027 superintelligence as possibly non-trivial but not baseline should plan for it while still having fallbacks if progress is slower.
  • Paul Christiano predicts possible AI takeover within 6 to 12 months and calls for a pause until alignment and security are much better.

The article is a reaction to the METR report, which estimated that Anthropic's internal AI development pace had roughly doubled, and to the timeline revisions that followed. The AI Futures Project, the team behind the AI 2027 scenario, moved its median estimate for a Superhuman AI Researcher (SAR), an AI that can take over AI research, from around 2030 to 2027, according to Eli Lifland. Daniel Kokotajlo had assigned roughly a 4% chance to the kind of benchmark acceleration that actually showed up. Paul Christiano, who gave three days of talks and a long Dwarkesh Patel interview, predicts rapid progress soon, with possible AI takeover within 6 to 12 months, and he calls for a pause until alignment and security are in much better shape. The author, Zvi Mowshowitz, thinks shorter timelines are directionally correct, but puts an SAR in 2027 under 50% without thinking it is far under 50%. He argues that anyone assigning zero or near-zero weight to superintelligence in 2027 is making a clear error, and he's irritated that many people are only now updating on evidence the labs themselves had been publishing for months. The piece spends real effort on what "timelines" even means, because people are using the word for different milestones. Zvi separates a Superhuman AI Researcher (SAR) from a Top-Expert-Dominating AI (TED-AI), which could automate all white-collar work, and from full superintelligence. He also distinguishes a Point of No Return, where human decisions can no longer change the outcome, from a visible AI takeover, and from the more extreme endpoint of human extinction. Recursive self-improvement, in his framing, has no single moment of onset; it is a gradual handover in which humans drift out of the loop and progress compounds. The practical upshot he draws is that a plan should not depend on the timeline being long, but also should not collapse if things take longer than 2027. Since small changes in the acceleration rate of an exponential produce large differences in when events occur, both overconfidence and complacency are risky. Zvi's main worry is that no one has a credible plan for making superintelligence go well if it arrives in 2027, and he says so plainly: "We need more time." He also flags that a coordinated slowdown, either among labs or through governments pacing or pausing frontier work, would push these timelines out. The article is long and the forecasting details are dense, so the summary version is the core: the trend lines moved sharply, the people closest to the data moved with them, and the gap between that and the state of safety planning is the thing he thinks deserves attention.

Ask about this article