伊利亚·苏茨克沃参与创造了现代人工智能，从 AlexNet 到 GPT，但随着 AI 变得强大，他转而反对 OpenAI CEO 萨姆·奥尔特曼，担忧超级智能机器带来的生存威胁。
Ilya Sutskever helped create modern AI, from AlexNet to GPT, but as AI grew powerful, he turned against OpenAI's CEO Sam Altman, fearing the existential threat of superintelligent machines.
要点 · TL;DR
扩展计算和数据可预测地提升 AI 能力，但安全风险也随之增加。 Scaling compute and data predictably improves AI, but safety risks grow.
下一个词预测是通往通用智能的强大途径。 Next-token prediction is a powerful path to general intelligence.
伊利亚离开 OpenAI，专注于无商业压力的安全超级智能。 Ilya left OpenAI to focus on safe superintelligence without commercial pressure.
核心观点 · Key points
扩展算力和数据能带来AI能力的可预测提升。 Scaling compute and data yields predictable improvements in AI capabilities.
神经网络无需人类显式规则即可学习表征。 Neural networks can learn representations without explicit human rules.
下一个词预测是构建智能的强大方法。 Next-token prediction is a powerful approach to building intelligence.
随着模型能力增强且不透明，AI安全担忧加剧。 AI safety concerns grow as models become more capable and opaque.
超级智能若未与人类价值观对齐，将带来生存风险。 Superintelligence poses existential risks if not aligned with human values.
反共识 · Contrarian takes
AI研究人员并不完全理解大型神经网络的内部运作。 AI researchers do not fully understand how large neural networks work internally.
AGI的主要问题不是能力，而是权力集中。 The primary problem of AGI is not capability but power concentration.
构建安全的超级智能需要一家没有商业产品的公司。 Building safe superintelligence requires a company with no commercial products.
仅靠扩展规模，即使没有新想法，也能带来重大突破。 Scaling alone, without new ideas, can still lead to major breakthroughs.
真正的AI竞赛不在于用户或收入，而在于研究专注度。 The real AI race is not about users or revenue but about research focus.
本期章节 · Chapters（共 11）
引言与背景Introduction and Background
早年生活与教育Early Life and Education
遇见杰弗里·辛顿Meeting Geoffrey Hinton
神经网络早期研究Early Work on Neural Networks
扩展与ImageNet突破Scaling and the ImageNet Breakthrough
从视觉到语言：序列到序列学习From Vision to Language: Sequence-to-Sequence Learning