AI Futures Project and Paul Christiano Update Timelines to 2027-2028 Following METR Report

·2026.11.19 00:18

Key point

AI Futures Project forecasts a Superhuman AI Researcher (SAR) likely in 2027 and a 35-65% chance of Top-Expert-Dominating AI (TED-AI) before 2028, driven by METR data showing 2x faster progress at Anthropic.

1 / 12

Details

  • The AI Futures Project (AI 2027 Team) forecasts a Superhuman AI Researcher (SAR) likely in 2027 and a 35-65% chance of Top-Expert-Dominating AI (TED-AI) before 2028, based on METR estimates of 2x faster progress at Anthropic. The article author, Zvi Mowshowitz, contrasts this by stating he believes an SAR in 2027 is under 50% likely.
  • Paul Christiano, an OpenAI board member, predicts a 'material risk' of AI takeover within 6-12 months, updating his previous 4% one-year loss-of-control estimate after observing internal OpenAI pace and the METR report.
  • Zvi Mowshowitz argues that historical predictions of AI hitting a wall have failed 'approximately 7 of the last zero times', asserting that current barriers are not qualitatively different and that an industry-wide pause is urgently needed to manage the acceleration.
  • Dwarkesh Patel now expects AIs as good as the best humans at AI research in a couple of years (or months), citing the verifiability of AI research as a driver for rapid recursive self-improvement, while Dean Ball distinguishes between 'weak recursion' (historical norm) and 'strong recursion' (autonomous successor development).

Timeline Estimates and Metrics

  • METR attributes ~80% of current acceleration to AI, placing progress 5 months ahead of the original AI 2027 scenario.
  • Buck Shlegeris estimates a 1/3 chance of superintelligence within 1 year and a ~20% chance that misalignment issues prevent normal life (due to death, enslavement, or collateral damage) within that period.
  • Nate Soares (MIRI) maintains a >50% chance of reaching 2028, contrasting with the shorter timelines of Christiano and the AI 2027 team.

Recursive Self-Improvement (RSI) Definitions

  • Dean Ball defines 'strong recursion' as AI controlling successor development with little human oversight, which he argues is a bad idea and should be regulated.
  • 'Weak recursion' is described as technology being a key input for producing better versions of itself, a historical norm that Ball argues is not inherently dangerous.
  • Zvi Mowshowitz critiques Ball's distinction, arguing that AI doubling R&D speed is a precursor to strong recursion and that 'slow takeoff' is actually very fast.

Skepticism and Policy Responses

  • Zvi Mowshowitz identifies skepticism regarding AI hitting a wall, noting that such predictions have failed 'approximately 7 of the last zero times'.
  • The article calls for an industry-wide pause followed by international agreement, supported by figures including Nate Silver, Elizabeth Warren, and Ezra Klein.
  • Government intervention is viewed skeptically due to political incentives to avoid existential risk and focus on economic catch-up rather than safety.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.