Executive summary: The author reviews their 2025 AI predictions, noting mixed calibration, and then offers quantified and qualitative forecasts for 2026 that predict substantial capability and revenue growth, modest increases in public salience, and a small but non-negligible chance of extreme outcomes like automated AI R&D or loss of control.
Key points:
The author reports being somewhat too bullish on benchmark progress and much too bearish on AI revenue in their 2025 predictions, notably underestimating annualized December revenue growth.
For 2026, they predict continued benchmark advances, including a longest-reported METR 50% time horizon of 24 hours and an Epoch ECI score of 169, implying high but slowing progress on established benchmarks.
They forecast combined annualized December revenue of $110 billion for OpenAI, Anthropic, and xAI, alongside modest increases in AI’s public salience and roughly stable net favorability.
The author lists several concrete tasks they expect AIs will likely be able to do by the end of 2026, including competent video game play, complex personal logistics, short-form video generation, visual novel creation, and solving the hardest problem in the 2026 IMO.
They also identify tasks they expect AIs will still not be able to do by the end of 2026, such as top-player-level gameplay in complex new games, organizing large weddings, producing festival-quality long films, generating long high-quality games, or publishing top-tier theoretical research from scratch.
Finally, the author assigns low but non-zero probabilities to extreme milestones by 2026, including near-full automation of AI R&D (10%), top-human-expert-dominating AI (5%), self-sufficient AI (2.5%), and unrecoverable loss of control (0.5%), while emphasizing that humanity remains unprepared even if these outcomes are unlikely.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.
Executive summary: The author reviews their 2025 AI predictions, noting mixed calibration, and then offers quantified and qualitative forecasts for 2026 that predict substantial capability and revenue growth, modest increases in public salience, and a small but non-negligible chance of extreme outcomes like automated AI R&D or loss of control.
Key points:
The author reports being somewhat too bullish on benchmark progress and much too bearish on AI revenue in their 2025 predictions, notably underestimating annualized December revenue growth.
For 2026, they predict continued benchmark advances, including a longest-reported METR 50% time horizon of 24 hours and an Epoch ECI score of 169, implying high but slowing progress on established benchmarks.
They forecast combined annualized December revenue of $110 billion for OpenAI, Anthropic, and xAI, alongside modest increases in AI’s public salience and roughly stable net favorability.
The author lists several concrete tasks they expect AIs will likely be able to do by the end of 2026, including competent video game play, complex personal logistics, short-form video generation, visual novel creation, and solving the hardest problem in the 2026 IMO.
They also identify tasks they expect AIs will still not be able to do by the end of 2026, such as top-player-level gameplay in complex new games, organizing large weddings, producing festival-quality long films, generating long high-quality games, or publishing top-tier theoretical research from scratch.
Finally, the author assigns low but non-zero probabilities to extreme milestones by 2026, including near-full automation of AI R&D (10%), top-human-expert-dominating AI (5%), self-sufficient AI (2.5%), and unrecoverable loss of control (0.5%), while emphasizing that humanity remains unprepared even if these outcomes are unlikely.
This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.