Ilya Sutskever 谈 AI 进展、对齐和时间线

Ilya Sutskever on AI Progress, Alignment, and Timelines

雅库布·帕霍茨基 Jakub Pachocki · Unsupervised Learning · 2026-04-09 · 约 59 分钟 · 原视频 ↗

本期速览 · Overview

OpenAI 首席科学家讨论模型进展、对齐挑战以及实现研究级 AI 能力的路径。

OpenAI's chief scientist discusses model progress, alignment challenges, and the path to research-level AI capabilities.

持续学习通过扩展预训练和强化学习实现，而非专门方法。
Continual learning is achieved by scaling pre-training and RL, not specialized methods.
思维链监控是关键对齐工具，必须在产品中隐藏以保持其可解释性价值。
Chain-of-thought monitoring is a key alignment tool that must be kept hidden in products.
AI 模型很快将能自主工作数天，需要新的监督范式。
AI models will soon work autonomously for days, requiring new supervision paradigms.

持续学习是核心目标，通过扩展预训练和强化学习来实现。
Continual learning is the core goal, achieved by scaling pre-training and RL.
思维链监控是关键的对齐工具，提供了可解释性。
Chain-of-thought monitoring is a key alignment tool, providing interpretability.
AI 在科学领域取得进展，模型在数学和物理中产生新想法。
AI for science is progressing, with models generating novel ideas in math and physics.
长期对齐挑战集中在训练分布之外的泛化。
Long-term alignment challenges center on generalization beyond training distribution.
模型很快将能自主工作数天，需要新的监督范式。
Models will soon work autonomously for days, requiring new supervision paradigms.

仅强化学习就足以实现持续学习；专门的持续学习实验室方向有误。
RL alone is sufficient for continual learning; specialized continual learning labs are misguided.
在产品中隐藏思维链对于保持其可解释性价值至关重要。
Hiding chain-of-thought in products is crucial to preserve its interpretability value.
当前模型已具有经济变革性，而不仅是研究新奇事物。
Current models are already economically transformative, not just research curiosities.
尽管强大 AI 的时间线缩短，但对对齐的乐观情绪反而增加。
Alignment optimism has increased despite shorter timelines to capable AI.
最大的社会风险是少数人运营的自动化组织导致的权力集中。
The biggest societal risk is concentration of power from automated organizations run by few.

引言与AI研究实习时间线 Introduction and Timelines for AI Research Intern
北极星转变：从数学基准到实际效用 Shift in North Stars: From Math Benchmarks to Real-World Utility
模型连接现实世界的复杂性 Complexity of connecting models to the real world
研究组织转型 Transition in Research Organization
算力分配策略 Compute Allocation Strategy
反思Anthropic在编程上的成功 Reflection on Anthropic's Success in Coding
编程代理的未来自主性 Future autonomy of coding agents
持续学习与RL扩展 Continual learning and RL scaling
长周期任务与模型改进 Long-horizon tasks and model improvement
AI科学：首个证明挑战 AI for Science and First Proof Challenge
AI科学：进展与挑战 AI for Science: Progress and Challenges
AI安全与思维链监控 AI Safety and Chain-of-Thought Monitoring
思维链与可解释性 Chain of Thought and Interpretability
长期对齐作为泛化 Long-term alignment as generalization
平衡研究与竞争 Balancing research and competition
流行度与长期研究的张力 Tension between popularity and long-term research
快问快答：过去一年改变的想法 Quick-fire questions: Changed mind in the last year
机器人技术时间线 Timelines for robotics
被忽视的社会影响 Under-thought societal impacts
用AI培养下一代 Raising the next generation with AI
结语与行动号召 Closing thoughts and call to action