AIFP Anticipates a SAR in 2027 and Calls for a Pause

A group of authors has revised the likely emergence of a "Superhuman AI Researcher" to 2027 and estimates a 35–65% probability of AI automating all office work before 2028. They base this shift in the timeline on a METR estimate of significant acceleration at Anthropic and advocate for an industry-wide pause followed by an international agreement.

Urgent Pause and Risks of Misaligned SAR

In the event of a SAR arriving in 2027, a pause or a significantly slowed pace of cutting-edge AI development would be necessary, preferably through a coordinated industry pause followed by an international agreement. They assert that companies are not in a position to properly align such a system. They also warn that using a misaligned SAR for self-recursive improvement could lead to an extremely unfavorable situation. Finally, they argue that if a SAR were trained soon, China could likely seize the training recipe and potentially the model weights.

Revised Timeline: SAR in 2027, 35–65% for TED-AI before 2028

The group now predicts that a Superhuman AI Researcher is likely to emerge in 2027, provided there is no substantial deliberate slowdown. They also assign a probability of 35 to 65% to the emergence, before 2028, of AIs capable of automating all office work. Previously, Daniel, described as the most optimistic of the authors, estimated a 50% probability of a SAR appearing before 2029. The SAR is defined as the point at which a company would choose fully automated AI researchers over maintaining the best human team without AI.

What Motivates These Revisions: METR's Estimate on Anthropic

The revisions are primarily based on a METR report. This report estimates that Anthropic is progressing about twice as fast as its historical rate, and notes that a factor between 1.3x and 3x would not be surprising. METR also claims that about 80% of the current acceleration can be attributed to AI. If these figures are accurate, the AI-driven acceleration would be ahead by 5 months of a scenario targeting 2027. They describe this level of AI-related acceleration as very unexpected and indicate that, among various factors considered, the METR report had the most influence on their reevaluation.