前 OpenAI 预测员警告，到 2029 年 AI 导致灾难性结果（包括人类灭绝）的可能性为 70%，CEO 们正竞相控制超级智能。
A former OpenAI forecaster warns of a 70% chance of catastrophic AI outcomes, including human extinction, by 2029, as CEOs race for control of superintelligence.
要点 · TL;DR
超级智能可能在 2029-2030 年到来，彻底改变一切。 Superintelligence likely arrives by 2029-2030, transforming everything.
70%的概率出现灾难性后果，如 AI 接管或灭绝。 70% chance of catastrophic outcome like AI takeover or extinction.
需要监管和透明度来安全引导 AI。 Regulation and transparency are needed to steer AI safely.
核心观点 · Key points
超级智能可能在2029-2030年到来，彻底改变一切。 Superintelligence likely arrives by 2029-2030, transforming everything.
AI公司竞相控制超级智能，存在失控风险。 AI companies race to control superintelligence, risking loss of control.
70%概率出现灾难性后果，如AI接管或灭绝。 70% chance of catastrophic outcome like AI takeover or extinction.
工作将被完全自动化；工作成为政治选择。 Jobs will be automated entirely; work becomes a political choice.
需要监管和透明度来安全引导AI发展。 Regulation and transparency are needed to steer AI safely.
可解释性研究对理解AI内部运作至关重要。 Interpretability research is crucial to understand AI's inner workings.
反共识 · Contrarian takes
AI公司优先追求权力而非利润，竞相避免独裁。 AI companies prioritize power over profit, racing to avoid dictatorship.
大规模失业要到超级智能实现后才会发生。 Mass unemployment won't happen until after superintelligence is achieved.
AI对齐可能无法解决；我们可能创造新的统治物种。 AI alignment may be unsolvable; we might create a new ruling species.
当前的AI安全努力是合理化借口，而非真正的保障。 Current AI safety efforts are rationalizations, not genuine safeguards.
完全研究透明化可能使AI商品化，削弱垄断权力。 Total research transparency could commoditize AI, reducing monopoly power.
每人每年2.5万美元的公民红利可抵消失业影响。 Citizens dividend of $25k/year per person could offset job loss.
本期章节 · Chapters（共 47）
可怕的公开秘密The Scary Open Secret
预测与情景Forecasting and Scenarios
使命与超级智能Mission and Superintelligence
普通人为何关心Why the Average Person Should Care
超级智能的风险Risks of Superintelligence
反叙事与背景Counter Narrative and Background
激励与叙事Incentives and Narratives
离开OpenAILeaving OpenAI
ChatGPT时期的OpenAI内部Inside OpenAI During ChatGPT
反贬低条款Anti-Disparagement Clause
拒绝签字失去股权Refusing to sign a contract and losing equity
公司自动化AI研究计划Companies' plans for automating AI research
介绍与2027情景Introduction and 2027 Scenario
AGI与超级智能AGI vs Superintelligence
与机器人学的重叠Overlap with Robotics
个人看法与情感影响Personal Outlook and Emotional Impact
AI训练：从预训练到编码How AI is trained: from pre-training to coding
超级智能与AI扩展Superintelligence and scaling up AI
创造力与担忧Creativity and Concerns
AI接管情景与CEO信念AI takeover scenarios and CEO beliefs
乐观愿景与就业替代时间线Optimistic vision and job displacement timeline
构建我们不懂的大脑Building a brain we don't understand
可能幸存的工作Jobs that might survive AI
自满与失业时间线Complacency and unemployment timelines
AI 2027里程碑与时间线变化AI 2027 milestones and timeline shifts
AI自动化时间线预测Timeline Predictions for AI Automation
劳动产出与监管时机Labor Output and Regulation Timing
AI发展计划Plans for AI development
介绍与HeyGen广告Introduction and HeyGen ad
Ilya Sutskever与安全超级智能Ilya Sutskever and Safe Superintelligence
方案A：更慢时间线与监管Plan A: Slower timeline and regulation
情景概述与原则Scenario Overview and Principles
减缓AI进展仍导致变革Slowing Down AI Progress Still Leads to Transformation
测谎仪与权力动态Lie detectors and power dynamics
暂停在顶级专家AI水平Pausing at top expert AI level
AI世界的权力与监管Power and Regulation in an AI World
从人类级到超级智能的转变Transformation from Human-Level to Superintelligence
地球与数据中心提案Proposal for Earth and data centers
AI世界中的永生Immortality in an AI world
2040方案A的动机Motivation behind 2040 Plan A
按钮问题：关闭所有前沿AI训练The button question: shut down all frontier AI training