AI 2027 Team Releases AI 2040 Future Scenario Projections

On July 10, researchers including Daniel Kokotajlo and Ryan Greenblatt, known for AI 2027, released a new report titled AI 2040. The report explores the long-term societal consequences of artificial intelligence and proposes a more positive alternative vision to the catastrophic scenarios of AI takeover or irreversible power concentration predicted in their previous work. Author Kokotajlo noted that this is merely their current best guess, hoping to inspire better plans before it is too late.

Core Scenarios and Tech Predictions

The report outlines multiple projection plans. "Plan A" envisions a smoothly progressing scenario where the US and China reach an agreement in 2029 to avoid an out-of-control race, leading to slower AI research, increased transparency, and a "citizen dividend" by 2033. Technologically, the report predicts the emergence of 100T-parameter models around 2028. However, the authors believe that a race towards Artificial Superintelligence (ASI), referred to as Plan D, remains the most likely actual outcome.

Governance Proposals and Controversies

Regarding governance, AI 2040 explicitly opposes the public release of frontier open-weight models, citing the risk of high-capability models being used to design catastrophic biological weapons. The report also highlights the massive impact of data centers on energy systems under high-AI-growth scenarios.

Community Reactions

The report sparked widespread discussion within AI safety and governance circles. Miles Brundage praised the value of its scenario projections but expressed a preference for a "CERN for AI" approach, suggesting the report's timeline for automated AI research might be conservative. Scott Alexander and others published beginner-friendly explanations to help disseminate the ideas. Meanwhile, critics like Sebastian Krier argued that the future depicted in the report exposes a severe "monoculture problem" within the AI safety community.

2026-07-10 ~ 2026-07-11 · 40 related posts

Full story(5 episodes)→

2 near-duplicate retellings: JeffLadish · scaling01