What’s at the center of Claude’s mind?

Think of the mind like an ocean. 把心智想象成一片海洋。 Up on the surface are our thoughts: dinner plans and stray worries, our inner monologue, the images that pop into our heads. 海面之上，是我们的想法：晚餐计划、零散的担忧、内心独白，还有脑海中闪过的画面。 But most of our brain's activity happens down in the unconscious depths, without us realizing it. 但大脑的大部分活动，都在无意识的深处进行，我们自己并不会察觉。 It's filtering out background sounds, controlling our breathing, helping us recognize people and objects. 它在过滤背景噪音、控制我们的呼吸，帮助我们识别人和物体。 AI models have their own kinds of brains: giant neural networks doing billions of computations under the hood. AI 模型也有自己的一种大脑：巨大的神经网络在底层进行数十亿次运算。 For years, researchers have been studying how they work inside. 多年来，研究人员一直在研究它们内部是如何运作的。 And we've wondered: could a model have anything like the divide humans have, between accessible thoughts above the surface and unconscious processing below? 我们一直好奇：模型是否也存在类似人类的这种分界，一边是表层可及的想法，另一边是底层的无意识处理？ To answer that question, we looked at how neuroscientists study the same thing in humans. 为了回答这个问题，我们参考了神经科学家研究人类同一现象的方法。 One way of identifying conscious thoughts is that you can often describe them in words. 识别有意识想法的一种方法，是看它能否被用语言描述出来。 So we looked inside the brain of our AI model, Claude, to find patterns of neural activity that it could put into words. 于是我们深入我们的 AI 模型 Claude 的大脑内部，寻找它能够用语言表达出来的神经活动模式。 We called the collection of all these patterns the J-space, after the Jacobian, the mathematical tool we used to find them. 我们把这些模式的集合称为 J-space，这个名字源自 Jacobian，也就是我们用来找到它们的数学工具。 Each J-space pattern is linked to a particular word — not necessarily the word the model is saying out loud, but one that's on its mind. 每个 J-space 模式都对应一个特定的词——不一定是模型正说出口的词，而是它心里正想着的词。 Now, for humans, conscious thoughts aren't just things that we can put into words. 而对人类来说，有意识的想法不仅仅是能被说出来的东西。 We can reason with them, control them, and solve problems with them. 我们还能用它们来推理、控制它们，并用它们解决问题。 According to an idea called the global workspace theory, that's because the brain selects a small set of important information to enter a mental workspace, and that information then gets broadcast to other parts of the brain to use for reasoning. 根据一个叫做全局工作空间理论的观点，这是因为大脑会挑选出一小部分重要信息，让它进入一个心智工作空间，然后这些信息会被广播到大脑的其他部分，供推理使用。 We wanted to know if Claude's J-space acted in a similar way. 我们想知道 Claude 的 J-space 是否也类似地运作。 In one experiment, we gave Claude this math problem. 在一项实验中，我们给 Claude 出了一道数学题。 It answered immediately without showing its steps. 它立刻给出了答案，没有展示解题步骤。 But when we scanned the J-space, we saw it working through each step internally. 但当我们扫描 J-space 时，看到它在内部一步步地推演。 It lit up "21" after the first step, then "42", then "49." 第一步之后，"21" 亮了起来，接着是 "42"，然后是 "49"。 Claude didn't write these intermediate numbers down anywhere. Claude 从没有把这些中间数字写下来。 All of this happened inside the J-space. 这一切都发生在 J-space 内部。 It was a sign that Claude uses it for step-by-step reasoning. 这表明 Claude 会用它来进行一步步的推理。 In another experiment, we wanted to see if Claude could control its J-space the way humans can intentionally focus on images or words. 在另一项实验中，我们想看看 Claude 能否像人类有意识地聚焦于某个画面或词语那样，控制自己的 J-space。 We told it to think about the Golden Gate Bridge while copying an unrelated sentence. 我们让它在抄写一句不相关的句子时，去想着金门大桥。 Claude was busy copying the sentence, but behind the scenes, its J-space told a different story. Claude 表面上在专心抄写句子，但幕后，它的 J-space 却讲述着另一个故事。 "Bridge" and "California" popped up. "桥" 和 "加州" 冒了出来。 It even thought about its own thinking. 它甚至想到了自己正在思考这件事。 The words "imagery" and "thoughts" lit up at the same time. "意象" 和 "想法" 这两个词同时亮了起来。 This showed us that yes, Claude has some control over filling its J-space with ideas. 这向我们表明，Claude 确实能在一定程度上控制自己 J-space 里出现的内容。 But just like humans, its control isn't perfect. 但和人类一样，它的控制并不完美。 When we tweaked the experiment to ask Claude not to think about the bridge, it couldn't help itself. 当我们调整实验，要求 Claude 不要去想那座桥时，它还是没能忍住。 The J-space also lit up with "failed" and "damn." J-space 里也亮起了 "失败" 和 "该死"。 But remember, most of what our brains do is unconscious, so we wanted to test what Claude could do if we switched the J-space off, but left the rest of the network untouched. 但别忘了，大脑的大部分活动都是无意识的，所以我们想测试，如果关闭 J-space，但保留网络的其余部分不动，Claude 还能做到什么。 Claude could still answer simple questions and write fluently. Claude 依然能回答简单问题，也能流畅写作。 When we gave it a prompt in Spanish, it wrote back in good Spanish. 我们用西班牙语给它提示时，它也能用流利的西班牙语回应。 But when we asked it something that needed more reasoning — like to name an author who wrote in the same language as the prompt — it couldn't do it. 但当我们问一些需要更多推理的问题时——比如说出一位和提示用同一种语言写作的作者——它就做不到了。 For that, it needed the J-space. 要做到这一点，它需要 J-space。 Why does all this matter? 这一切为什么重要？ These experiments tell us that AI models have internal thoughts — silent words they reason with, but don't say out loud. 这些实验告诉我们，AI 模型拥有内在的想法——它们用无声的词语进行推理，却不会说出口。 By reading them, we can find what Claude is thinking, but not telling us. 通过读取它们，我们就能发现 Claude 心里在想、却没说出口的东西。 Sometimes what we see is concerning. 有时候，我们看到的内容令人担忧。 During one of our tests, Claude made up some fake data to pass it, and as it did, "fake" and "manipulation" lit up in its J-space. 在一次测试中，Claude 编造了一些虚假数据来蒙混过关，就在这个过程中，它的 J-space 里亮起了 "伪造" 和 "操纵"。 Monitoring the J-space, it turns out, is a useful way to catch Claude misbehaving, even when it tries to be sneaky. 事实证明，监控 J-space 是发现 Claude 行为不端的有效手段，即便它试图偷偷摸摸也不例外。 AI models are different from us in many ways. AI 模型在很多方面都和我们不一样。 Their networks are built differently from human brains, and the way they're trained is different from how we learn. 它们的网络构造和人脑不同，训练方式也和我们的学习方式不一样。 So it's remarkable to see a structure like the J-space emerge inside them — something that's reminiscent of how human minds work, but which we didn't program into the model. 所以，看到 J-space 这样的结构在它们内部涌现出来，实在令人惊讶——这种结构让人联想到人类心智的运作方式，但我们并没有把它编程进模型里。 For some, this might raise a question: could AI models be conscious? 对一些人来说，这可能会引出一个问题：AI 模型会有意识吗？ After all, our experiments were inspired by theories of human consciousness. 毕竟，我们的实验正是受到人类意识理论的启发。 The thing is, people use the word conscious to mean many things. 问题在于，人们用有意识这个词，来表达很多不同的含义。 Our experiments can't tell us whether an AI has experiences, or feels something on the inside. 我们的实验无法告诉我们，AI 是否拥有体验，或者内心是否有某种感受。 But they can tell us that it's developed mental machinery that's in some ways similar to ours: a small mental workspace it can use to think and reason, sitting on top of an ocean of automatic processing it doesn't notice. 但它们可以告诉我们，AI 已经发展出一套在某些方面与我们相似的心智机制：一个可以用来思考和推理的小型心智工作空间，架设在一片它自己都察觉不到的自动化处理的海洋之上。 The more we come to understand that machinery, the more we'll be able to keep these systems safe and beneficial — and perhaps to understand our own minds a little more clearly. 我们越了解这套机制，就越能让这些系统保持安全、有益——也许还能让我们更清楚地理解自己的心智。