图灵奖得主 Sutton：合成数据是大错特错

AI researcher Richard Sutton pushes back against a core strategy at the leading AI labs. Synthetic data won't solve the scaling problem for large language models, he says, and blames the sheer complexity of the world.

图灵奖得主 Richard Sutton 被视为强化学习领域的奠基人。他编写了该领域的标准教科书，指导过 David Silver 等研究者——后者后来参与了 AlphaGo 项目，并撰写了影响深远的 2019 年文章《苦涩的教训》（The Bitter Lesson）。在该文中，Sutton 认为，从长远来看，只有随算力扩展的 AI 方法才能胜出，比如搜索和学习，而不是依赖内置人类知识的方法。

In a recent conversation, he introduced his new company, Oak Lab, which he founded with his former student Khurram Javeed. He and his cofounder also talked about the limits of current training methods.

在最近的一次对话中，他介绍了自己的新公司 Oak Lab，这是他与昔日学生 Khurram Javeed 联合创立的公司。他和这位联合创始人还讨论了当前训练方法的局限性。

Why LLMs are only half a win

为什么大语言模型只赢了一半

For Sutton, large language models are both a good and a bad example of the Bitter Lesson. Good because they scaled enormously with compute and you could simply "drink in the internet." Bad because they hit a wall at exactly that point. The internet is finite, and the real world is "massively bigger than everything we stored on the internet." At that point, Sutton says, you lean too heavily on human knowledge, and that ultimately holds the systems back.

对 Sutton 而言，大语言模型既是苦涩教训的一个好例子，也是一个坏例子。好的一面是它们随算力大幅扩展，你可以简单地"喝下整个互联网"。坏的一面是它们恰好在那个节点撞上了墙。互联网是有限的，而真实世界"比我们存储在互联网上的所有东西都要大得多"。Sutton 指出，到了那个阶段，你过度依赖人类知识，而这最终会拖累系统。

Asked whether synthetic data could break through this bottleneck, Sutton doesn't mince words. "No, that's that's just a big mistake." The reasoning comes from the "Big World Hypothesis" that Javeed formulated and the group in Alberta has been working on for years.

当被问及合成数据能否突破这一瓶颈时，Sutton 直言不讳。"不，那是一个大错误。"这一推理来自 Javeed 提出的"大世界假说"（Big World Hypothesis），这也是阿尔伯塔团队多年来一直在研究的方向。

The world is too big for any simulation

世界太大了，任何模拟都装不下

The core assumption is that the world is infinitely complex and "massively more complex than your mind than any agents any agent." Any simulation of it is tiny, "microscopic." A small program can only ever produce a small world that doesn't match reality, with wrong friction values or an inaccurate model of a robot's motor behavior. Sutton also points out that the world contains many other agents whose inner workings simply can't be generated as synthetic data. "There's no way we can have synthetic data for other people's minds."

核心假设是，世界是无限复杂的，"比你的心智、比任何智能体都要复杂得多"。对世界的任何模拟都只是微小的、"显微级"的。一个小程序只能产生一个与现实不匹配的小世界，比如错误的摩擦系数或不准确的机器人电机行为模型。Sutton 还指出，世界上存在许多其他智能体，它们的内部运作根本无法作为合成数据生成。"我们不可能拥有他人心智的合成数据。"

A second objection is the human bottleneck. Who decides which synthetic data is good or bad? By Javeed's argument, that takes human experts. "You need human experts who know what's a good data set and what's a bad data set for that approach to scale. So it is bottlenecked by humans." Say you wanted to train a drone that moves like a bat using echolocation. You'd first have to hire domain experts. That makes the approach limited by human expertise, and it doesn't scale. Even with self-driving cars trained in simulation, humans end up fixing the gap between simulation and reality.

第二个异议是人力瓶颈。谁来决定哪些合成数据是好的或坏的？按照 Javeed 的论点，这需要人类专家。"你需要人类专家来判断什么是好的数据集、什么是坏的数据集，才能让这种方法扩展。所以它受制于人类。"假设你想训练一只像蝙蝠一样利用回声定位移动的无人机，你首先得雇佣领域专家。这使得这种方法受限于人类专业知识，而且无法扩展。即使是在模拟环境中训练的自驾车，最终也需要人类来弥补模拟与现实之间的差距。

Sutton's alternative is learning from your own experience

Sutton 的替代方案：从自己的经验中学习

Sutton's fix is to take humans out of the loop and let agents learn from their own experience. An agent should learn its own world model and keep correcting it, instead of relying on a frozen simulation model that humans built. "Simulators they make themselves."

Sutton 的解决方案是把人类从循环中移除，让智能体从自己的经验中学习。智能体应该学习自己的世界模型并不断修正它，而不是依赖人类构建的冻结模拟模型。"让智能体自己制作模拟器。"

Sutton also criticizes the fact that today's language models stop learning after training. "Their weights never change." What's needed instead is real continual learning, which in Sutton's view is just learning, since "all learning is continual," without wiping out old knowledge, the so-called catastrophic forgetting.

Sutton 还批评了当前语言模型在训练后停止学习这一事实。"它们的权重永远不会改变。"真正需要的是持续的真正学习，而在 Sutton 看来，这就是学习本身，因为"所有学习都是持续的"，不会抹去旧知识，即所谓的灾难性遗忘。

Sutton thinks this problem can be solved, in part with a method called "Continual Backprop" that his team published in Nature. He calls language models an "amazing scientific breakthrough," but only "like 20% or a quarter of intelligence."

Sutton 认为这个问题是可以解决的，部分可以通过他的团队在《自然》杂志上发表的一种叫做"持续反向传播"（Continual Backprop）的方法来实现。他称大语言模型是一项"惊人的科学突破"，但只是"大约 20% 或四分之一的智能"。

AI News Without the Hype – Curated by Humans

无炒作的 AI 新闻——人工精选

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

订阅 THE DECODER，享受无广告阅读、每周 AI 时事通讯、每年六期的独家"AI Radar"前沿报告、完整档案访问以及评论 section。

Read on for the full picture.Subscribe for free coverage.

继续阅读以获得完整图景。免费订阅无炒作报道。

Full access to every article on THE DECODER

完整访问 THE DECODER 的每一篇文章

Join the comments and community discussions

参与评论和社区讨论

A weekly AI news recap via mail

每周邮件 AI 新闻综述

6x/year: "AI Radar" — deep dives on the AI topics that matter most

每年六期："AI Radar"——深入探讨最重要的 AI 话题

Daily AI news, always up to date

每日 AI 新闻，始终最新

Our full ten-year archive

我们完整的十年档案

Covered by a team with 10+ years in AI

由拥有 10 年以上 AI 经验的团队报道