After the METR report, forecasters shortened their AI timelines; Zvi wrote a commentary and cautioned against treating 2027 superintelligence as a baseline
Quickly, There’s No Time
Zvi wrote in LessWrong commenting on the updates in which many forecasters substantially shortened AI timelines after the release of the METR report, holding that the direction is right but the magnitude may be too large. The AI Futures Project predicts that superhuman AI researchers (SAR) will most likely emerge in 2027, with a 35-65% probability of TED-AI (capable of automating all white-collar work) emerging before 2028, and the largest source of evidence is METR's estimate that Anthropic's AI progress rate is about 2 times the historical rate (range of 1.3-3 times). OpenAI board member Paul Christiano said there is a substantial risk of AI takeover in the next 6-12 months, higher than his September estimate of only a 4% probability of loss of control within one year, and in a personal capacity called for a pause on development until safety and alignment improve substantially. Dwarkesh Patel updated his median expectation for human-level automated AI researchers to 2028.