里奇·萨顿：这个领域才是奇怪的，不是我

Rich Sutton: The Field Is Weird, Not Me

理查德·萨顿 Richard Sutton · Training Data · 2026-08-18 · 约 54 分钟 · 原视频 ↗

本期速览 · Overview

里奇·萨顿讨论了他关于持续学习的非传统观点，以及他创立 Oak Lab 的历程。

Rich Sutton discusses his unconventional views on continual learning and his journey founding Oak Lab.

世界无限复杂，智能体必须不断从经验中学习，而非仅依赖预先收集的数据。
The world is infinitely complex, so agents must continually learn from experience, not just pre-collected data.
合成数据受限于人类专业知识，无法替代真实世界的经验。
Synthetic data is bottlenecked by human expertise and cannot replace real-world experience.
真正的 AI 需要持续学习，而非像当前 LLM 那样训练后停止。
True AI requires continual learning, not just training then stopping, as current LLMs do.

世界是无限复杂的，因此智能体必须不断从经验中学习，而不是仅仅依赖预先收集的数据。
The world is infinitely complex, so agents must continually learn from experience rather than rely solely on pre-collected data.
合成数据受限于人类专业知识，无法替代从真实世界经验中学习。
Synthetic data is bottlenecked by human expertise and cannot replace learning from real-world experience.
持续学习对AI至关重要，需要能够在不发生灾难性遗忘的情况下更新权重的算法。
Continual learning is essential for AI, requiring algorithms that update weights without catastrophic forgetting.
苦涩的教训强调用学习方法扩展算力，而不是依赖人类知识。
The bitter lesson emphasizes scaling computation with learning methods, not relying on human knowledge.
当前的大语言模型受限于训练后停止学习；真正的智能需要持续学习。
Current LLMs are limited because they stop learning after training; true intelligence requires ongoing learning.

合成数据生成是一个大错误；它无法捕捉真实世界的无限复杂性。
Synthetic data generation is a big mistake; it cannot capture the infinite complexity of the real world.
监督学习不是人类或动物学习的主要方式；大多数学习来自经验。
Supervised learning is not the primary way humans or animals learn; most learning is from experience.
大语言模型只是智能的一个子集，而非全部；语言仅占智能的20-25%。
LLMs are only a subset of intelligence, not the whole; language is just 20-25% of intelligence.
这个领域才是奇怪的，不是我；持续学习就是学习，不是特例。
The field is weird, not me; continual learning is just learning, not a special case.
单一AI系统无法学习所有事物；需要具有不同经验的多个系统。
A single AI system cannot learn everything; multiple systems with different experiences are necessary.