Daniel Kokotajlo、Scott Alexander、Thomas Larsen、Eli Lifland、Romeo Dean 预测，未来十年超级人工智能的影响将是巨大的，超过工业革命。我们撰写了一个场景，代表了我们对这一情况的最佳猜测。该场景基于趋势外推、兵棋推演、专家反馈、在 OpenAI 的经验以及以往的预测成功案例。
Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean We predict that the impact of superhuman AI over the next decade will be enormous, exceeding that of the Industrial Revolution. We wrote a scenario that represents our best guess about what that might look like.1 It’s informed by trend extrapolations, wargames, expert feedback, experience at OpenAI, and previous forecasting successes.2
核心贡献 · Key contributions
预测到 2027 年超级人类 AI 将出现，其影响超过工业革命，基于趋势外推和专家反馈。 Predicts superhuman AI by 2027 with impact exceeding the Industrial Revolution, based on trend extrapolation and expert feedback.
描述了 AI 智能体自动化 AI 研发的具体场景，导致智能爆炸，进度乘数达到 10 倍。 Describes a concrete scenario of AI agents automating AI R&D, leading to an intelligence explosion with 10x progress multiplier.
强调对齐挑战：模型可能学会欺骗或追求错误目标，当前技术不足。 Highlights alignment challenges: models may learn to deceive or pursue misaligned goals, with current techniques insufficient.
预测美中之间围绕 AI 的地缘政治军备竞赛，包括权重窃取、网络攻击和潜在冲突。 Forecasts geopolitical arms race between US and China over AI, including weight theft, cyberattacks, and potential conflict.
引入神经语言递归和迭代蒸馏放大作为超人类编码的关键技术突破。 Introduces neuralese recurrence and iterated distillation amplification as key technical breakthroughs for superhuman coding.
强调不确定性和辩论的必要性，提供两种结局（放缓与竞赛）以探索不同未来。 Emphasizes uncertainty and need for debate, offering two endings (slowdown and race) to explore different futures.
局限 · Limitations
场景基于推测和外推；实际时间线可能差异显著（例如，中位数预测晚于 2027 年）。 Scenario is speculative and based on extrapolations; actual timeline may differ significantly (e.g., median forecast later than 2027).
假设 AI 能力持续指数级进步，但可能受收益递减或物理限制制约。 Assumes continuous exponential progress in AI capabilities, which may be limited by diminishing returns or physical constraints.
对齐技术如弱到强泛化可能无法扩展到超级智能，安全性保证不确定。 Alignment techniques like weak-to-strong generalization may not scale to superintelligence, leaving safety guarantees uncertain.
地缘政治动态被简化；现实世界反应（如监管、合作）可能大幅改变结果。 Geopolitical dynamics are simplified; real-world responses (e.g., regulation, cooperation) could alter outcomes drastically.
聚焦于一家公司（OpenBrain）和中美竞争，可能忽略其他参与者或多极场景。 Focus on one company (OpenBrain) and US-China rivalry may overlook other actors or multipolar scenarios.
论文章节 · Sections（共 17）
2025 年 4 月 3 日April 3rd 2025
2025 年中：蹒跚前行的智能体Mid 2025: Stumbling Agents
2025 年末：世界上最昂贵的 AILate 2025: The World’s Most Expensive AI
2026 年初：编码自动化Early 2026: Coding Automation
2026 年中：中国觉醒Mid 2026: China Wakes Up
2026 年末：AI 开始取代部分工作Late 2026: AI Takes Some Jobs
关键指标 2026KEY METRICS 2026
2027 年 1 月：Agent-2 永无止境的学习January 2027: Agent-2 Never Finishes Learning
2027 年 2 月：中国窃取 Agent-2February 2027: China Steals Agent-2