# Meet Yann LeCun's Lab and the AI World of 2030
- Transcript ID: `1307e6710a4ed84d`
- Source tier: `official_authored_text`
- Original language: `en_or_source_language_unlabeled`
- Translation status: `agent_translated_reviewed` (141/141 chunks)
- Original source: https://nebius.science/stories/meet-yann-lecuns-lab-and-the-ai-world-of-2030
- Raw evidence: `data/transcripts/web/yann-lecun-nebius-science-2026.json`
> 中文是派生翻译，英文是留存的源文本。中文未复核时，以英文和原始音视频为准。
## paragraph 1
**中文**
他正重注押在预测式架构上。于是，我们请他也来扮演一次“预测架构师”。
**English / Original**
He’s betting big on predictive architectures. We asked him to play the predictive architect.
## paragraph 2
**中文**
“这里没有秘密。”Yann LeCun 在一块写满潦草字迹的白板旁摆好姿势拍照时说道。这间洒满阳光、以橡木色调为主的厨房，同时也是他新创办的 AI 企业的中枢。即使真有秘密，他补充说，外人反正也看不懂。
**English / Original**
No secrets here, Yann LeCun says as he poses for a photograph next to a scribblings-laden whiteboard in a sunlit, oak-toned kitchen that doubles as his new AI venture’s nerve center. Even if there were secrets, he adds, no outsider could read them anyway.
## paragraph 3
**中文**
去年离开 Meta 后，这位图灵奖得主、纽约大学教授把注意力转向巴黎，在时尚的 Sentier 街区一个蓬勃发展的初创企业集群中创办了 AMI Labs。AMI 是 Advanced Machine Intelligence（先进机器智能）的缩写，藏身于一家香水店和一家成衣精品店之间，正在打造这位法裔美国科学家所称的“面向现实世界的 AI”。
**English / Original**
Since leaving Meta last year, the Turing Award winner and New York University professor has turned his attention to Paris, where he founded AMI Labs among a burgeoning cluster of startups in the chic Sentier district. Here, tucked away between a perfume shop and a prêt-à-porter boutique, AMI, short for Advanced Machine Intelligence, is building what the French-American scientist calls “AI for the real world.”
## paragraph 4
**中文**
白板上的标记暗示了这句话的含义：在一幅看起来像玫瑰的涂鸦上方，是一张预测式 AI 方案的双编码器草图。旁边写着对比损失函数的公式，用来引导系统的猜测趋向正确答案。附近还有一张标注着“瓶子，打开”和“瓶子，关闭”的向量图——这是他常用的例子，用来说明世界模型必须理解什么，才能预测各种动作的结果。
**English / Original**
The markings on the whiteboard hint at what that phrase means: Just above a doodle of what looks like a rose, there’s a double-encoder sketch of a predictive AI scheme. Next to it is a formula for a contrastive loss function that guides a system’s guesses toward correct answers. Nearby is a vector diagram labeled “bottle, open” and “bottle, closed” — his go-to example of what a world model must grasp to predict the outcomes of various actions.
## paragraph 5
**中文**
如果机器能够预见物理事件的后果——例如瓶子即将倾倒并洒出液体、机器即将发生故障，或人体正变得容易患病——就可能催生应用似乎无穷无尽的工具。LeCun 说，他的系统最终将使机器人无需数百万次训练演示，就能第一次尝试便应对全新而复杂的任务。目前，给 AMI 厨房里的酒柜添酒的仍然是人类。
**English / Original**
Being able to foresee the consequences of a physical event, like when a bottle is about to tip over and spill, a machine is about to malfunction, or a body is becoming susceptible to disease, could lead to tools with seemingly limitless applications. LeCun says his system will eventually allow robots to handle new and complex tasks on their first try, without millions of training demonstrations. For now, it’s humans who stock the wine fridge in AMI’s kitchen.
## paragraph 6
**中文**
LeCun 笑着说，世界模型就像早期的深度学习，是本季的“青少年性话题”：“人人都声称自己在做，但没人知道它究竟是什么。”他很乐意解释世界模型——这种实验性 AI 系统与如今主导行业的大语言模型有根本不同。这不只因为 LeCun 善于传播科学，也因为世界模型是他最新、或许也是最大的一次押注。
**English / Original**
Like deep learning in its early days, world models are this season’s teenage sex, LeCun says with a grin: “Everybody claims they’re doing it, but nobody knows what it is.” He’s happy to explain world models, experimental AI systems that are fundamentally different from the LLMs now dominating the industry. And not just because LeCun is a gifted communicator of science, but also because world models are his latest — and perhaps biggest — bet.
## paragraph 7
**中文**
正是这种信念促使他在担任 Meta 首席 AI 科学家十余年后离职。在他看来，这家社交媒体巨头过度投入大语言模型，而他认为单靠大语言模型无法通向人类水平的智能。如今，LeCun 的愿景几乎与整个领域相左：业界仍在凭借他所怀疑的那种架构不断跨越里程碑，而他自己的架构仍留在实验室里。如此投入还伴随着个人代价。他工作越多，留给那些数十年来滋养其创造性另一面的爱好——巴洛克音乐和电子音乐——的时间就越少。
**English / Original**
That faith prompted him to leave Meta after more than a decade as chief AI scientist. To his mind, the social media giant was just too committed to large language models, something he believes will not, on its own, lead to human-level intelligence. Today, LeCun’s vision sets him against nearly the entire field, which continues to hit milestone after milestone using the very architecture he doubts, while his own remains in the lab. And then there’s the personal toll such commitment entails. The more he works, the less time he has for the pursuits that have fed his creative alter ego for decades, namely baroque and electronic music.
## paragraph 8
**中文**
LeCun 所看到的未来，值得承担这种风险和取舍：一台能够在物理世界中规划、推理和行动的机器；它学习的是行为本身的结构，而非行为外观的表层形式。在 10 亿美元投资和海量算力的支持下——Nebius 是关键供应商之一——LeCun 基于联合嵌入预测架构（Joint Embedding Predictive Architecture，JEPA）的模型正开始取得可喜成果。他说，再过几个月，AMI 就会开始在实际基准上测试这些模型。
**English / Original**
What LeCun sees coming is worth both the risk and the trade: a machine that can plan, reason, and act in the physical world, learning the structure of its behaviour, rather than the surface of how it looks. Backed by $1 billion in investment and an ocean of compute, with Nebius being a key provider, LeCun’s models based on the Joint Embedding Predictive Architecture (JEPA), are starting to produce promising results. A few more months, he says, and AMI will start testing them on practical benchmarks.
## paragraph 9
**中文**
他的模型会预测，他本人也会。LeCun 的世界如今围绕预测和里程碑运转。就连 AMI 的白板上也自豪地写着一个里程碑：“2027 年 1 月”，并被圈了起来。因此，当他坐下来接受这次采访时，对话自然而然地以未来时展开。室外，今夏第一波热浪正笼罩法国；AI 世界内部的新闻也同样火热。
**English / Original**
His models predict. So does he. LeCun’s world now revolves around forecasts and milestones. Even AMI’s whiteboard boasts one: “Jan. 2027,” circled. So when he sat down for this interview, the conversation naturally unfolded in the future tense. Outside, the first heat wave of the summer was settling over France. Inside the AI world, the news was just as hot.
## paragraph 10
**中文**
人们经常问你 AI 的未来，大多数人想知道十年后会是什么样。这很自然：一个把大部分时间用于开发联合嵌入预测架构的人，必然会被视为“预测架构师”。但我想，即便对你来说，预测十年也有些要求过高。
**English / Original**
People often ask you about the future of AI, and most want to know what it will look like in ten years. That’s only natural: someone who spends most of his time developing a Joint Embedding Predictive Architecture is bound to be seen as a predictive architect. But even for you, I imagine ten-year forecasts are asking a bit too much.
## paragraph 11
**中文**
是的。
**English / Original**
Yes.
## paragraph 12
**中文**
不过三年就正合适。
**English / Original**
But three is just fine.
## paragraph 13
**中文**
三到五年，是的。
**English / Original**
Three to five, yes.
## paragraph 14
**中文**
我想请你描绘一下并不遥远的未来——2030 年。如果说 2022 年标志着现代生成式 AI 的爆发，而今天我们正经历大规模 AI 训练时代，你会怎样称呼 2030 年？
**English / Original**
I’d like to ask you to paint a picture of the not-so-distant future — 2030. If 2022 marked the breakout of modern genAI, and today we’re living through the era of massive AI training, what would you call 2030?
## paragraph 15
**中文**
面向现实世界的 AI。
**English / Original**
AI for the real world.
## paragraph 16
**中文**
而且，2022 年的那场革命其实更早就发生了，至少对研究界的人而言如此。直到 2022 年末，它才成为面向公众的现象。公众对大语言模型如此热情，令所有人都感到意外，也让许多公司意识到这里存在市场。
**English / Original**
And the 2022 revolution really happened before, at least for people in the research community. It became a public-facing phenomenon in late 2022. That was a surprise to everyone that the public was so enthusiastic about LLMs. And it made a lot of companies realize that there was a market.
## paragraph 17
**中文**
但理解和生成语言是一回事，理解并控制现实世界则复杂得多。AI 领域几十年来一直知道这一点，它被称为莫拉维克悖论。领域内有些人被大语言模型的进展迷住了，忘记莫拉维克悖论依然存在。
**English / Original**
But understanding and producing language is one thing, while understanding and controlling the real world is much more complicated. And we’ve known this in AI for many decades. It’s called the Moravec paradox. Some people in the field are hypnotized by the progress of LLMs and forget about the Moravec paradox, which is still with us.
## paragraph 18
**中文**
这些会做数学、会写代码的系统非常有用，但这不是通往人类智能的道路，就像 20 世纪 90 年代初的神经网络也不是一样。我认为，未来四年左右，我们会看到另一场革命，它将为现实世界的事物打开大门，而不再只限于语言。这难得多，而这正是我想看到的革命。我相当确信它会发生。
**English / Original**
These systems that do math and write code are super useful. But this is not a path to human intelligence. Not more than neural nets were in the early nineties. I think, in the next four years or so, we’re going to see another revolution that will open the doors to real-world things, not just language. That’s much more challenging. And that’s a revolution I want to see. I’m pretty sure it’s going to happen.
## paragraph 19
**中文**
你曾说，拥有猫或老鼠那样聪明程度的 AI 系统就会是一项真正的成就。你对 2030 年的估计是什么？猫、老鼠，还是昆虫？
**English / Original**
You’ve said that AI systems as smart as a cat or rat would be a real achievement. What is your 2030 estimate? Cat, rat, insect?
## paragraph 20
**中文**
大概就是那个水平，就对物理世界的理解而言，约为猫或老鼠的量级。我们低估了动物有多聪明，因为我们把自己对智能的印象与操纵语言的能力联系起来，而这完全是错的。
**English / Original**
Something like that, on the order of a cat or a rat in terms of their understanding of the physical world. We discount how smart animals are because we correlate our impression of intelligence with language manipulation. And that’s just wrong.
## paragraph 21
**中文**
智能包括理解世界中发生了什么、世界将如何演变、你的行动可能产生什么后果，从而让你能够规划一系列行动来达成特定目标。而这是大语言模型完全做不到的。
**English / Original**
Being intelligent includes understanding what happens in the world, how the world is going to evolve, what the consequences of your actions might be, which would allow you to plan a sequence of actions to arrive at a particular goal. And this is something that LLMs can’t do at all.
## paragraph 22
**中文**
它们本身没有什么不好，也很有用，只是缺失了若干关键部分。我们缺少的就是世界模型，即能够预见世界中将发生什么以及某个行动会产生何种后果的模型。
**English / Original**
There’s nothing bad about them. They are useful. It’s just they’re missing major pieces. And that thing we’re missing is world models. Things that can anticipate what will happen in the world and the consequences of an action.
## paragraph 23
**中文**
2030 年达到猫的水平，离人类水平仍然有些远，不是吗？
**English / Original**
Cat level in 2030, that’s still a bit far from the human level, no?
## paragraph 24
**中文**
并不远，其实非常接近。
**English / Original**
It’s not. It’s really close.
## paragraph 25
**中文**
问题是：什么底层技术能让系统运行这些世界模型，并用它们进行规划？我认为未来几年内，我们会掌握一种很好的技术。
**English / Original**
The question is: what is the underlying technique that will allow a system to run those world models and use them for planning? I think we’ll have a good technique for that within the next few years.
## paragraph 26
**中文**
我们已经取得了一些很有希望的初步成果。大问题是，一旦有了这种技术，怎样让它在机器人、控制工业流程的系统，或通过观察学习复杂现象模型的系统中实际运行？这一步大概会推进得相当快。
**English / Original**
We already have preliminary results that are promising. The big question is, once you have that, how do you make it operational in a robot or a system that controls an industrial process, or a system that learns a model of a complex phenomenon by observation? That’s probably going to be pretty fast.
## paragraph 27
**中文**
但接下来还有一个问题：要打造能以人类同等水平学习和产生抽象概念的系统，需要哪些要素？我们不知道这要花多长时间。也许一旦掌握训练世界模型的基本技术，就很容易扩大规模，让它们处理更复杂的概念。
**English / Original**
But then there’s the question of what ingredients you need to build a system that can learn and produce abstractions at the same level as a human. And we don’t know how long that’s going to take. It could be that once we get the basic technique for training world models, it’ll be easy to scale them up and get them to work with more complex concepts.
## paragraph 28
**中文**
也可能比我们预想的困难得多，需要十年甚至更久。不过，这一概念会在两到三年内产生实用应用。
**English / Original**
Or it could be a lot more difficult than we thought. And it may take a decade or more. But there are going to be useful applications of this concept within two to three years.
## paragraph 29
**中文**
稍后我们会再谈世界模型。我想问问最近政府对一些 AI 模型施加的限制。你预见到这件事了吗？你认为它可能标志着未来几年 AI 行业更广泛危机的开始吗？
**English / Original**
We’ll come back to world models in a moment. I’d like to ask about the recent government-imposed restrictions on some AI models. Did you see this coming? Do you think it could mark the beginning of a broader crisis for the AI industry over the next few years?
## paragraph 30
**中文**
嗯，危机其实早已存在。有一种“安全主义”的论述，大意是 AI 很危险，正因为危险，所以不能交到所有人手中。这类论述在其他领域存在已久。
**English / Original**
Well, the crisis was already with us. There is a discourse, a safetyist discourse. Of the type, AI is dangerous, and because it’s dangerous, it can’t be put in the hands of everyone. This discourse has been around for a long time in other domains.
## paragraph 31
**中文**
印刷机发明时也出现过类似论述。它让人们能够获取知识，而这被视为危险。15 世纪起的两百年间，奥斯曼帝国禁止使用印刷机印刷阿拉伯文。据一些历史学家称，这实际上是奥斯曼帝国相对西方衰落的一个重要因素。
**English / Original**
There was a similar discourse when the printing press was invented. It gave people access to knowledge, and that was considered dangerous. In the 15th century, for 200 years, the Ottoman Empire banned the use of the printing press for Arabic. And according to some historians, that was actually a major factor in the Ottoman Empire’s decline relative to the West.
## paragraph 32
**中文**
今天，AI 是传播知识的一种方式。认为应该限制人们获取知识、智能或教育，是蒙昧主义。我在哲学立场上反对它，并认为它极其危险。
**English / Original**
Today, AI is a way to disseminate knowledge. The idea that you should limit access to knowledge, intelligence, or education is obscurantism. I’m philosophically opposed to it. I think it’s super dangerous.
## paragraph 33
**中文**
如果 AI 被美国西海岸或中国的少数几家私营公司控制，或像我们刚刚看到的那样被少数几个政府控制，那将非常可怕。这意味着民主的终结、社会自由主义的终结。你的思想以及你获得的全部信息，都将受少数人控制。
**English / Original**
If AI is controlled by either a few private companies on the West Coast of the US or China, or by a few governments, like we just saw, it’s horrible. It’s the end of democracy, the end of social liberalism. Your thoughts and all the information you get would be controlled by a few people.
## paragraph 34
**中文**
让所有人都能使用 AI，或许会伴随一些风险。但与权力集中和信息受控这一大得多的风险相比，那些风险小得多、也轻微得多。
**English / Original**
There might be some risks attached to giving access to AI to everyone. But those risks are much, much smaller than the far greater risk of concentrating power and controlling information.
## paragraph 35
**中文**
你认为到 2030 年，事态会演变成你刚才描述的景象吗？
**English / Original**
Do you see it evolving into the picture you just described by 2030?
## paragraph 36
**中文**
不会。我认为对此存在抵抗。大多数国家可能会联合起来，打造开源、开放访问、开放权重的开放 AI 平台。事实上，市场也在推动这种趋势，因为 AI 正在成为基础设施。
**English / Original**
No. I think there is resistance to this. Most countries are probably going to band together to produce open source, open access, open weight, open AI platforms. And in fact, there is also a market push for that because AI is becoming an infrastructure.
## paragraph 37
**中文**
互联网就是这样发展的，对吧？当某种事物成为基础设施，市场就会推动它变得开放和免费，因为这样可移植性更强、更安全，也更容易部署。所以我认为 AI 也会发生同样的事情；这只是时间问题，也在一定程度上取决于地缘政治竞争。
**English / Original**
That’s what happened to the internet, right? When something is an infrastructure, there’s pressure from the market for it to become open and free because it’s more portable, it’s more secure, and it’s just easier to deploy. So I think the same thing is going to happen for AI; it’s a question of time, and it’s a bit of a question of geopolitical rivalries.
## paragraph 38
**中文**
在我看来，除美国和中国之外的各国政府都应该全面拥抱开源。那基本上是它们取得主权的入场券。奇怪的是，美国政府——即便是现任政府——相当支持开源，反倒是行业并不支持。
**English / Original**
In my opinion, governments aside from the US and China should completely embrace open source. That’s basically their ticket to sovereignty. And it’s weird because the American government, even the current one, is fairly pro-open source; it’s just that the industry is not.
## paragraph 39
**中文**
而中国目前实际上产出了最好的开源模型，这正在造成一种非常奇特的效应：人们因为中国模型是开放的而转向使用它们。因此，这种压力确实存在。
**English / Original**
And China actually produces the best open source models right now, which is causing a very strange effect. People are switching to Chinese models because they’re open. So that pressure exists.
## paragraph 40
**中文**
我们转到 JEPA。它与当前模型有何不同？
**English / Original**
Let’s switch to JEPA. How is it different from current models?
## paragraph 41
**中文**
JEPA 的基本思想是：我们观察的是一种复杂现象，不是离散 token 序列，不是语言，而是来自工业系统的传感器数据、视频，或连续、高维且含噪的信号。你不能用生成式 AI 去理解它。原因在于，其中大部分数据不可预测：既有噪声，也有高度复杂的依赖关系。如果你尝试训练一个系统预测视频接下来会发生什么，它不可能预测每一个细节。对于一段给定的视频，合理的后续有无限多种。
**English / Original**
The basic idea of JEPA is that we have a complex phenomenon that we observe, not a sequence of discrete tokens, not language, but sensor data from industrial systems, videos, or continuous high-dimensional noisy signals. You cannot use generative AI to understand it. And the reason is that most of that data is unpredictable. There’s noise, and there are highly complex dependencies in it. And if you try to train a system to predict what’s going to happen in a video, there’s no way it can predict every detail. There’s an infinity of plausible continuations for a given video segment.
## paragraph 42
**中文**
当你告诉人们，处理现实世界数据的方法不是使用生成式 AI 时，他们很难相信。但这意味着什么？它意味着你不去预测信号的每一个细节，而是找到一种能让你作出预测的输入抽象表征。如果我拿起一个瓶子并把它碰倒，水就会洒出来。你无法逐个细节预测水会流到哪里，但也没有必要；你只需知道水会洒出来。
**English / Original**
When you tell people that the way to handle real-world data is not to use generative AI, they have a hard time believing you. But what does that mean? It means that instead of predicting every detail of the signal, you find an abstract representation of the input that allows you to make predictions. If I take a bottle and knock it over, it spills the water. There’s no way you can predict in every detail where the water is going to flow, but you don’t need to. You only need to know it’s going to spill.
## paragraph 43
**中文**
你可能想粗略预测它会朝哪个方向倒、什么东西会被弄湿，等等，但你不可能预测每个细节。即便模拟复杂系统也是如此。模拟器按理说要捕捉所有细节，但事实上并不能。
**English / Original**
You might want to predict roughly in which direction it’s going to fall, what’s going to get wet, and so on. But there’s no way you can predict every detail. The same is true even when you simulate a complex system. Simulators are supposed to capture all the details, but in fact, they don’t.
## paragraph 44
**中文**
我们从不会在那种细节层面模拟。我们会抽象掉那些细节，寻找一种让我们能够预测的表征。因此，要作出预测就必须找到抽象，这一思想至关重要。而生成模型做不到这一点，必须使用 JEPA。
**English / Original**
We never simulate at that level. We abstract away those details and find a representation that allows us to make predictions. So the idea that you need to find abstractions to make predictions is absolutely crucial. And you cannot do this with a generative model. You have to use a JEPA.
## paragraph 45
**中文**
这里存在一个技术难题。训练生成模型时，你确切知道要预测什么，因为你能观察到视频或信号的后续。但使用 JEPA 时，你要求它预测由系统自身产生的表征。问题在于系统可能发生坍塌：它可以选择忽略输入，并始终输出一个恒定表征。因此，训练 JEPA 的全部复杂性都在于防止这种坍塌。
**English / Original**
There’s a technical difficulty with this. When you train a generative model, you know exactly what to predict because you can observe the continuation of your video or your signal. But with a JEPA, you ask it to predict a representation that the system itself produces. And the problem is that the system can collapse. It can choose to ignore the input and produce a constant representation. So all the complexity of training a JEPA is in the prevention of that collapse.
## paragraph 46
**中文**
你希望 JEPA 到 2030 年实现什么？希望它具备哪些能力？
**English / Original**
What do you want JEPA to achieve by 2030? What do you hope it will be capable of?
## paragraph 47
**中文**
有些应用属于相对短期。我们正在与多家潜在产业合作伙伴洽谈，涉及提前预测故障或异常之类的事情。
**English / Original**
Some applications are relatively short-term. We’re talking to a bunch of potential industry partners. And they have to do with things like predicting failures or faults in advance.
## paragraph 48
**中文**
如果你有一个能很好描述复杂系统如何演变的模型，也许就能预测系统何时会失效，这称为预测性维护。你还可以检测异常动态：如果某处损坏，你会看到系统的演变方式不同寻常，这称为新颖性检测。JEPA 架构可以做到这些。
**English / Original**
If you have a good model of how a complex system evolves, maybe you can predict when the system will fail. That’s called predictive maintenance. You can also detect unusual dynamics. If something is broken, you’ll see that the way the system evolves is unusual. That’s called novelty detection. JEPA architectures can do this.
## paragraph 49
**中文**
接下来是控制问题。JEPA 可以转化为世界模型。世界模型所做的是：给定 t 时刻的世界状态，再给定你可能采取的行动或干预，它能够预测系统在 t+1 时刻、或一毫秒后、十毫秒后、一小时后的状态。换句话说，它能预测你所采取行动的后果。如果拥有这样的世界模型，你就能规划一系列行动来达到特定状态。这就是经典最优控制。
**English / Original**
Then there is the issue of control. A JEPA can be turned into a world model. What a world model does is this: given the state of the world at time t, and given an action or intervention you might take, it can predict the state of the system at time t + 1, or a millisecond later, or 10 milliseconds later, or an hour later. In other words, it can predict the consequences of the action you took. If you have such a world model, you can plan a sequence of actions to arrive at a particular state. That’s classical optimal control.
## paragraph 50
**中文**
假设你有一台喷气发动机、一座化工厂或一条制造生产线。你通过传感器、视频和其他测量手段观察系统的完整状态，然后对系统进行干预并观察结果。利用这些数据，你训练出整个系统的整体性现象模型。一旦拥有这个模型，就可以用它驱动系统，使各项指标得到优化。
**English / Original**
Imagine you have a jet engine, a chemical plant, or a manufacturing line. You observe the entire state of the system through sensors, video, and other measurements. Then you intervene in the system and observe the result. From that data, you train a holistic phenomenological model of the entire system. And once you have it, you can use it to drive the system toward optimized metrics.
## paragraph 51
**中文**
对喷气发动机而言，这可能意味着最大限度延长寿命或提高可靠性、尽量减少二氧化碳排放，或最大限度提升效率。制造生产线的目标有所不同，但原理相同。这是最优控制新时代的开端。
**English / Original**
For a jet engine, that might mean maximizing longevity or reliability, minimizing CO₂ emissions, or maximizing efficiency. For a manufacturing line, the objectives are different, but the principle is the same. This is the beginning of a new era in optimal control.
## paragraph 52
**中文**
这些就是你设想三年内落地的 JEPA 实际应用吗？
**English / Original**
These are the practical implementations of JEPA you envision in three years?
## paragraph 53
**中文**
可能比那更早。大约一年内我们就会有原型。我们会与少数合作伙伴一起，看看它可以应用于哪些工业领域。有些领域容易部署，有些会更困难。我们希望适用领域的数量会相当多。
**English / Original**
Probably before that. We’ll have prototypes within a year or so. We’ll be working with a few partners to see which industrial areas this can be applied to. Some areas will be easy to deploy in, others will be more difficult. We hope there will be a fairly large number of them.
## paragraph 54
**中文**
再往后，这大概会催生更通用的智能系统。将来，当你需要一个系统理解并控制某种特定现象时，只需对它稍加预训练、接入系统，它很快就会弄清怎样以最优方式控制该系统。最终，这会带来家用机器人、自动驾驶汽车等许多应用中的智能系统。
**English / Original**
Further down the line, this will probably lead to more general intelligent systems. When you’ll need a system that understands a particular phenomenon and can control it, you’ll simply pre-train it a little, plug it in, and within a short time, it will figure out how to control that system optimally. Eventually, that will lead to intelligent systems for things like domestic robots, self-driving cars, and many other applications.
## paragraph 55
**中文**
机器人行业眼下有一个大秘密。很多公司正在制造机器人，包括人形机器人，但没有一家知道怎样让它们足够聪明，从而具备普遍实用性。你可以收集足够的数据，让机器人胜任某一项特定任务；但归根结底，你希望机器人第一次听到要求时就能理解一项新任务——这是人类和许多动物完全能够做到的。
**English / Original**
There’s a big secret in the robotics industry right now. Many companies are building robots, including humanoid robots. None of them knows how to make them smart enough to be generally useful. You can collect enough data to make a robot useful for a particular task. But ultimately, you want a robot to apprehend a new task the first time you ask it—something humans, and many animals, are perfectly capable of doing.
## paragraph 56
**中文**
仅靠模仿训练的系统做不到这一点。每当你希望它执行新任务，都必须从人类那里收集海量演示，再训练系统进行模仿。但如果系统无法预见自身行动的后果，其能力就会受到根本限制。
**English / Original**
A system trained purely by imitation won’t be able to do that. Every time you want it to perform a new task, you’ll have to collect huge numbers of demonstrations from humans and train the system to imitate them. But if the system can’t anticipate the consequences of its own actions, its capabilities will remain fundamentally limited.
## paragraph 57
**中文**
你在 2026 年发表的论文《LeWorldModel》表明，这种架构方法可以在实践中奏效。下一个重大挑战是什么？
**English / Original**
Your 2026 paper, LeWorldModel, shows that this architectural approach can work in practice. What’s the next big challenge?
## paragraph 58
**中文**
有好几个挑战。其中之一是找到防止系统坍塌的良好配方，就像我前面解释的那样。我们已经有一种效果很好的技术，并在超大规模视频上进行了测试，但它也有一些缺点。它基于蒸馏：把完整视频交给一个编码器，把该视频的受损版本交给另一个架构相同的编码器。损坏方式可以是局部遮蔽，例如遮住视频的后半段或其中很大一块。
**English / Original**
There are several. One is coming up with a good recipe to prevent the system from collapsing, as I explained earlier. We already have a technique that works very well. We’ve tested it on video at a very large scale. But it has some downsides. It’s based on distillation. You show the full video to one encoder, and a corrupted version of that video to another encoder with the same architecture. The corruption can be partial masking; you can mask the second half of the video or a large chunk of it.
## paragraph 59
**中文**
然后让两个版本分别通过两个编码器，并训练预测器和编码器，根据部分受损的视频预测完整视频的表征。为防止坍塌，看到完整视频的编码器不会通过梯度下降来更新权重，而只是从另一个编码器那里取得权重。
**English / Original**
You then run both versions through the two encoders and train the predictor and the encoder to predict the representation of the full video from the partially corrupted one. To prevent collapse, the encoder that sees the full video doesn’t edit its weights by gradient descent. It actually just gets its weights from the other guy.
## paragraph 60
**中文**
这就是它被称为蒸馏的原因。第一个编码器的输出充当目标，用来训练预测器和另一个编码器。不知为何，它并不会坍塌，而我们对此还没有完全理解。因此理论仍有些不稳固，但方法确实有效。至少，我们已经有一个能够学习优质特征的可用系统实例。它能生成非常好的视频表征；如果把这些表征作为动作识别等下游任务的输入，就能取得当前最先进的性能。
**English / Original**
That’s why it’s called distillation. The output of the first encoder serves as the target for training the predictor and the other encoder. And it doesn’t collapse for some reason that we don’t completely understand. So the theory is still a little flaky, but it works. At least we have one working example of a system that can learn good features. You get very good video representations from that, and you get state-of-the-art performance if you use those representations as input to downstream tasks for action recognition.
## paragraph 61
**中文**
这很好，但我们希望有一种不使用这种奇怪蒸馏方法的技术，转而采用我们所谓的正则化器，使表征中的信息量最大化。我们有几个候选方案。《LeWorldModel》使用了一种名为 SIGReg 的技术，非常有希望。
**English / Original**
That’s great, but we’d like to have a technique that doesn’t use this weird distillation, that uses what we call regularizers to maximize the information content in the representations. We have a few candidates. LeWorldModel uses a technique called SIGReg, which is a very promising one.
## paragraph 62
**中文**
它仍有一些我们希望解除的约束。因此，我们目前投入大量时间寻找正确配方，并通过它在下游任务上的表现来衡量效果。这还需要几个月的反复试验。这正是 AMI 目前最主要的工作。
**English / Original**
It still has a few constraints we’d like to lift. So right now we’re spending a lot of time trying to find the right recipe and measuring it by its performance on downstream tasks. That will require several more months of playing around. That’s what we’re mostly doing at AMI these days.
## paragraph 63
**中文**
你们怎样组织 AMI 的研究工作？
**English / Original**
How do you organize research in AMI?
## paragraph 64
**中文**
AMI 目前约有 50 人，主要是工程师和科学家。
**English / Original**
AMI currently has about fifty people, mostly engineers and scientists.
## paragraph 65
**中文**
目前很大一部分重点放在工程上：软件、存储、数据加载器、数据清洗和 GPU 基础设施。当然，我们把 Nebius 作为 GPU 算力的主要供应商之一。
**English / Original**
Right now, there’s a lot of focus on engineering: software, storage, data loaders, data cleaning, and GPU infrastructure. And of course, we’re using Nebius as a major supplier of GPU compute.
## paragraph 66
**中文**
另一项工作则聚焦于完善训练分层 JEPA、并将其用作世界模型的方法论。层级是一个极其根本的概念；事实上，它正是深度学习的根源，也是深度学习之所以“深”的原因。我们只能通过层级来理解世界。
**English / Original**
And then there is another effort, which is focused on solidifying the methodology to train hierarchical JEPAs and to use them as a world model. Hierarchy is a very fundamental idea. In fact, it’s at the root of deep learning. The reason deep learning is deep. We can only understand the world through hierarchy.
## paragraph 67
**中文**
训练第一层时，系统从传感器取得信号，然后生成一种抽象程度不高的表征，用于作出非常短期的预测。比如以每秒 10 或 15 帧的视频预测接下来的几帧，这并不是很复杂的任务。但这种预测式训练得到的视频表征，会保留下一些帧中可预测的内容，同时消除所有只是噪声的细节。随后，在下一层，你把第一层的输出拿来，训练它作出更长期的预测。
**English / Original**
You train the first layer, the system gets the signal from the sensors, and then produces a not very abstract representation that makes very short-term predictions. You predict the next few frames in a video at 10 frames or 15 frames per second. Not a very complicated task. But what you get from this predictive training is a representation of the video that contains what can be predicted in the next frames while eliminating all the details that are just noise. Then, at the next layer, you take the output of the first one and train it to make predictions over a longer term.
## paragraph 68
**中文**
训练系统作长期预测时，它必须消除大量细节，因为大多数细节都无法预测。你迫使系统预测的时间跨度越长，得到的表征就越抽象。这非常重要，一方面因为人类通过层级理解世界，机器也应如此；另一方面，如果能把这些分层 JEPA 训练成世界模型，就能实现分层规划。
**English / Original**
And when you train a system to make long-term predictions, it has to eliminate a lot of details, because most details you cannot predict. And the longer term you force the system to make predictions, the more abstract representations you get. This is very important because we understand the world through hierarchy, and machines should do the same. But also because if we can train those hierarchical JEPAs to be world models, it would enable hierarchical planning.
## paragraph 69
**中文**
所以这属于更多研究工作。其中很多成果会发表，大量研究代码会开源。基础技术会公开，因为我们希望人们帮助我们，也希望他们在这个领域提出好想法。
**English / Original**
So that’s more research. A lot of it will be published. Much of the research code will be open source. The basic techniques will be out there because we want people to help us. We want them to come up with good ideas around this domain.
## paragraph 70
**中文**
我认为，假定我们能独占好想法会极其自负。许多公司认为自己垄断了好点子，所以转向保密。我们没有那么傲慢。我们认为，建造智能机器大概是这个时代最重大的科学与技术挑战之一。即便是我们这样资金充足的初创企业，也不垄断好想法。
**English / Original**
I think that would be incredibly presumptuous. A lot of companies believe that they have a monopoly on good ideas, which is why they go secret. We’re not that arrogant. We think that building an intelligent machine is probably one of the biggest scientific and technological challenges of our times. And even within a well-funded startup like ours, we don’t have a monopoly on good ideas.
## paragraph 71
**中文**
我们工作中的专有部分，在于怎样把这些技术应用于具体工业场景。那里需要一些不会发表的东西：可扩展基础设施、领域专长，以及如何让这些模型在特定条件下充分发挥性能。如果你愿意把它称为护城河或秘方，它就在基础研究的下游。基础研究是开放的。
**English / Original**
The proprietary part of what we’re doing is how we apply this to specific industrial applications. That’s where you need stuff you don’t publish: scalable infrastructure, domain expertise, how you make these models sing in particular conditions. That’s where a bit of the moat, if you want, the secret sauce, is downstream from the basic research. The basic research is open.
## paragraph 72
**中文**
你预计未来三年智能体生态系统会发生重大变化吗？还是说，它现在或多或少已经呈现出最终形态？
**English / Original**
Do you envision any major changes to the agentic ecosystem over the next three years? Or is it already more or less taking the shape it’s going to have?
## paragraph 73
**中文**
不，我认为几年内它会变得非常不同，主要是因为当前智能体系统缺乏世界模型。构建一个无法预测自身行动后果的智能体系统，是非常糟糕的主意。
**English / Original**
No, I think it’s going to look very different within a few years. Mostly because of the lack of world models in current agentic systems. It’s a very bad idea to build an agentic system without it having the ability to predict the consequences of its own actions.
## paragraph 74
**中文**
当前智能体系统采用的替代办法是照着配方执行。它们无法预测自己行动的后果，而这是一种糟糕的智能体构建方式。它在一定程度上有效，却无法泛化；系统只是盲目套用配方。而且这种方式的生产成本也很高，因为你需要收集大量数据。因此，未来几年智能体系统将发生一场重大革命。
**English / Original**
The alternative to that is what current agentic systems are doing, which is applying a recipe. They have no way of predicting the consequences of their actions. And that’s a terrible way to build an agentic system. It works to some extent, but it’s not generalizable. They just apply the recipe blindly. It’s also very expensive to produce because you need to collect a lot of data. So there’s going to be a big revolution in the agentic system over the next few years.
## paragraph 75
**中文**
能够在所有条件下运行的完全自主 L5 级自动驾驶汽车，到 2030 年有可能实现吗？
**English / Original**
Fully autonomous, level-5 self-driving cars that can operate under all conditions. Possible by 2030?
## paragraph 76
**中文**
是的，完全有可能。
**English / Original**
Yes, entirely possible.
## paragraph 77
**中文**
那能把碗碟放进洗碗机、又不会摔坏它们的家用机器人呢？
**English / Original**
And a home robot that loads my dishwasher without breaking the dishes?
## paragraph 78
**中文**
那还要一段时间。至于能像水管工一样工作、修理你家里各种问题的机器人，则要久得多。那会是最后被自动化的工作，因为它要求对底层现象和操作有非常充分的理解。距离实现还有很多年，甚至几十年。
**English / Original**
That’ll take a while. And the robot that will do the same job as a plumber and fix the problems in your house, that’s much longer. That’s the last job to be automated. Because that requires a very good understanding of the underlying phenomenon and manipulation. That’s years away, if not decades.
## paragraph 79
**中文**
你认为我们很快会看到 AI 从人类无法联系起来的模式中提出全新科学假设吗？
**English / Original**
Do you think we’ll soon see AI proposing novel scientific hypotheses from patterns no human would connect?
## paragraph 80
**中文**
显然，把 AI 应用于科学和技术进步拥有巨大潜力。人们寄予很大希望，但对 AI 在材料科学、生物医学等领域应用的投资可能还不够。
**English / Original**
Clearly, the application of AI to science and technological progress has enormous potential. And there’s a lot of hope and probably not enough investment in the use of AI for things like material science and biomedicine.
## paragraph 81
**中文**
JEPA 之类的技术到一定阶段会相当有用。我们如今看到，在科学上取得成功的所有系统，比如 AlphaFold、ESM——ESM 团队过去在 Meta，其中两人是我的学生——都是专用模型。它们不是大语言模型。它们使用 Transformer，也使用自监督学习，但不是大语言模型，也不试图以自回归方式预测下一个 token。
**English / Original**
Techniques like JEPA would actually be quite useful at some point. What we’re seeing now is that all of the systems that have been successful in science, like AlphaFold, like ESM, which is a group that used to be at Meta, two of them are students of mine — those are all specialized models. They’re not LLMs. They use transformers, they use self-supervised learning. But they’re not LLMs. They’re not trying to predict the next token autoregressively.
## paragraph 82
**中文**
未来的演变是，不再需要为每个领域进行大量模型专门化，模型本身将能够适应领域。只要你有足够数据，即便完全没有标注也可以。未来几年内这可能会发生，也确实是许多相关工作背后的希望。
**English / Original**
The evolution will be that you won’t need extensive model specialization for each domain. Instead, the model will be able to adapt to the domain. If you have enough data, even if it’s completely unlabeled. That could happen within the next few years. That’s certainly the hope behind a lot of these efforts.
## paragraph 83
**中文**
科学论文中的“AI 垃圾”又如何？虚构的参考文献、由 AI 写成并被大型会议或严肃期刊接收、随后被撤稿或没有撤稿的论文——同行评审体系有崩溃的风险吗？
**English / Original**
What about the AI slope in scientific papers? Hallucinated references, papers being written by AI and accepted to big conferences and serious journals, and then retracted, or not. Is there a risk of the peer-review system collapsing?
## paragraph 84
**中文**
既是，也不是。同行评审体系极不完善。我写过关于它以及如何修复它的檄文。大约 15 年前，NIPS 会议讨论怎样改革这一体系时，我写过一篇，提出了我认为同行评审应该如何运作的方案；它不同于今天的做法。
**English / Original**
Yes and no. The peer-review system is extremely imperfect. I’ve written pamphlets about it and about how to fix it. I wrote one about 15 years ago, when the NIPS conference was discussing how the system could be changed. I made a proposal for how I think peer review should work, which is different from how it works today.
## paragraph 85
**中文**
我们确实需要同行评审，因为需要一种方式让研究人员评论同行的论文，有些匿名，有些公开。我的提案本质上讨论的就是怎样组织这件事。
**English / Original**
We do need peer review, because we need a way for researchers to comment on papers by their peers, some anonymously, some openly. My proposal was essentially about how to organize that.
## paragraph 86
**中文**
论文应该是自成一体的实体。现在，你为某个特定出版场所、期刊或会议写论文，双方存在排他关系。如果论文被拒，你就改投别处，评审过程又从零开始，这很荒谬。
**English / Original**
Papers should be self-contained entities. Right now, you write a paper for a particular venue, a journal, or a conference, and there’s an exclusive relationship between the two. If the paper is rejected, you submit it somewhere else, and the review process starts from zero, which is ridiculous.
## paragraph 87
**中文**
我认为应该有一个开放市场，论文直接公开放在那里。随后，人或组织可以自行决定是否评审、是否认可它们。我不认为应当存在排他关系，所有事情都应公开进行。
**English / Original**
I think there should be an open marketplace, where papers are simply out there. Then people or organizations can choose whether to review them and whether to approve them. I don’t think there should be exclusive relationships. It should all happen in the open.
## paragraph 88
**中文**
再说 AI 垃圾。是的，外面确实有糟糕的论文，但人类一直也在写很差的论文。在某种程度上，我认为 AI 正在提高论文质量。把 AI 当作写论文的工具，我看不出有什么问题。但如果你使用 AI 却不核实所写内容，就把论文提交出去，之后论文因你没有检查而被发现存在重大缺陷，那么你的声誉就应该一落千丈——至少暂时如此。
**English / Original**
Now AI slop. Yes, there are bad papers out there. But there have always been terrible papers written by people. To some extent, I think AI is improving the quality of papers. I see nothing wrong with using AI as a tool to write papers. Now, if you use AI and you don’t verify what is written, and you submit the paper, and then the paper is discovered to have a major flaw because you didn’t check, your reputation should go down the tube, at least temporarily.
## paragraph 89
**中文**
今天已经存在类似现象。大型实验室的负责人会署名实验室产出的许多论文，但未必读过其中每一篇。这样可以吗？两者并没有太大差别。
**English / Original**
There’s already a similar phenomenon today. Directors of large labs have their names on many papers coming out of their labs, but they haven’t necessarily read every one of them. Is that okay? It’s not very different.
## paragraph 90
**中文**
归根结底，生产端需要做好质量控制，评审端则需要良好筛选。大多数论文基本无人理会，而这本来就是应有的结果。没有理由阻止人们访问某篇论文，而传统同行评审实际上正是这么做的。每篇论文都应该可以获取。
**English / Original**
Ultimately, it’s a question of quality control on the production side. Then we need good filtering on the review side. Most papers are largely ignored, which is what should happen. There’s no reason to prevent access to a paper, which is effectively what traditional peer review does. Every paper should be available.
## paragraph 91
**中文**
arXiv 已经在这样做。在此基础上，还应由社区借助 AI 工具对论文进行评价。随后，人们可以利用这些信息形成自己对论文的看法。
**English / Original**
That’s what arXiv already does. On top of that, there should be community-based evaluation of papers, assisted by AI tools. Then people can use that information to form their own opinion about a paper.
## paragraph 92
**中文**
你认为未来几年 AI 将怎样改变教育？儿童依赖 AI 学习是否存在风险？它会不会干扰儿童自然学习和发展的方式？
**English / Original**
How do you think AI will transform education over the next few years? Do you see any risks in children relying on AI for learning? Could it interfere with the way children naturally learn and develop?
## paragraph 93
**中文**
我看到的是机会。纵观历史，每当一种新的通信技术出现，人们都会说：“哦，现在……”以 20 世纪 70 年代计算器的出现为例。老师们说：“孩子们以后不再做心算了。他们不需要，因为会有计算器。”
**English / Original**
I see an opportunity. I mean, throughout history, when a new communication technology comes in, people say, “Oh, now…” Take the emergence of calculators in the 1970s. Teachers said, “Children aren’t going to do mental arithmetic anymore. They won’t need to because they’ll have calculators.”
## paragraph 94
**中文**
确实如此。但通常在孩子学会心算之前，计算器不会进入课堂；只有学会以后才允许使用。如今，每个孩子都能接触到本质上是一台超级计算器的 AI。这难道不会改变整个局面吗？
**English / Original**
That’s true. But calculators are generally kept out of the classroom until children have learned to do arithmetic in their heads. Only then are they allowed to use calculators. Now every child has access to what is essentially a super-calculator—AI. Doesn’t that change the equation?
## paragraph 95
**中文**
会，但它也会改变你在教育孩子时把时间花在哪里。
**English / Original**
Yes, but it also shifts what you spend time on when teaching the children.
## paragraph 96
**中文**
我在法国读八年级时，有一次全国考试。数学科目中，我们必须学习向量、点积等内容，也要学会使用计算尺，以及根据印刷表格对对数和三角函数进行插值。现在你不再需要做这些，它们已从课程中消失。但在当时，这些是考查孩子的重要内容。AI 会产生同样的作用：它会改变中小学生以及后来大学生需要学习的内容。
**English / Original**
When I was in the eighth grade in France, there was a national exam. In mathematics, we had to learn vectors, dot products, and stuff like that. But also how to use a slide rule and interpolate logarithms and trigonometric functions from printed tables. You don’t need to do this anymore. It disappeared from the programs. But at the time, these were major things we were testing kids on. AI will have the same effect. It will change what schoolchildren and university students later need to learn.
## paragraph 97
**中文**
包括硅谷一些人在内，有人传播一种观点：以后不再需要上大学，因为 AI 会替你完成一切。这是我听过最愚蠢的话。事实恰恰相反，AI 正在增加对更高学历的需求。如今，AI 越来越能处理具体实现和解决较低层次的问题，但你仍然需要能够提出正确问题、设计正确系统，并向 AI 系统分配任务的人。
**English / Original**
There’s this idea, disseminated by some people, including in Silicon Valley, that you won’t need to go to college anymore because AI will do everything for you. That’s the most stupid thing I’ve ever heard. It’s exactly the opposite. AI is increasing the demand for more advanced degrees. What happens now is that AI can increasingly take care of implementing things and solving lower-level problems. But you still need people who can ask the right questions, design the right systems, and assign tasks to AI systems.
## paragraph 98
**中文**
AI 不会自己发明一项任务或一个新产品；这仍掌握在人类手中。每个人都在成为一支 AI 智能体团队的管理者。这需要更多教育，而不是更少；需要有远见的人，需要理解某个领域将走向何方、市场是什么样子的人。我们看到，对博士人才的需求正在增加。如今最酷、而且广受认可的工作，是成为 AI 研究员；这大多需要博士学位。
**English / Original**
They are not going to invent a task or a new product; that’s in the hands of humans. Every human is becoming the manager of a team of AI agents. That requires more education, not less. It requires people with vision, people who understand where a field is going, what a market looks like. What we see is an increasing demand for people with PhDs. The coolest job you can have today, and it’s very widely shared, is a researcher in AI. And mostly that requires a PhD.
## paragraph 99
**中文**
AMI Labs 最早的一批合作伙伴之一是医疗初创公司 Nabla。你认为未来几年哪些临床领域将主要受到 AI 改造？
**English / Original**
One of the first partners of AMI Labs is a medical startup, Nabla. What clinical domains do you see mostly transformed by AI in the coming years?
## paragraph 100
**中文**
很多领域都会，但会非常困难，因为在医疗保健中部署任何东西，都需要临床研究和政府机构批准。医疗保健领域已经有很多 AI 应用，主要并非大语言模型，其中大多数用于图像理解。Nabla 更多处理医生本该完成的行政任务，例如填写表格和对诊断进行分类，就像医生的副驾驶。
**English / Original**
A lot, but it’s going to be very difficult because deploying something in healthcare requires clinical studies and clearance from government agencies. There are already many AI applications in healthcare, mainly not LLMs. Most of them do image understanding. Nabla is more about the administrative tasks that doctors are supposed to do, filling forms, and classifying diagnoses. It’s like a copilot for doctors.
## paragraph 101
**中文**
但最终，你会看到 AI 越来越多地应用于诊断、数据分析和理解。它在医学影像中的影响已经很大，未来还会变得极其巨大。现在，当你接受 MRI、CT 扫描或 X 光检查时，系统需要生成可在平面屏幕上显示的图像，供人类解读。
**English / Original**
But eventually you’ll see AI being applied more and more to diagnostics, data analysis, and understanding. In imaging, the impact is already big, but it’s going to become enormous. Right now, when you go through an MRI, CT scan, or X-ray, the system needs to generate an image that can be displayed on a flat screen. So that a human could interpret it.
## paragraph 102
**中文**
但原始数据并不是一组图像，而是一个体数据，我们的眼睛其实并不适合解读它。AI 系统无需观看图像，可以直接处理原始数据。例如，MRI 的初始数据并非图像，而是某种频谱域数据。你可以不经过成像过程，直接把它输入 AI 系统；成像过程实际上还会降低数据质量。
**English / Original**
But the original data isn’t a bunch of images. It’s a volume, and our eyes are not really designed to interpret that. An AI system doesn’t need to look at an image. It can work directly with the original data. For an MRI, for example, it’s not an image at first. It’s some spectral domain. You can feed that directly into an AI system without going through the imaging process, which actually degrades the quality of the data.
## paragraph 103
**中文**
这一领域会取得大量进展。AMI Labs 更专注于世界模型这一侧。假设你有一名患糖尿病等慢性病的患者，应该采取怎样的治疗方案，才能使其血糖保持在受控范围？答案想必因患者而异。人们的生理状况、肠道微生物组各不相同，对食物的反应也不同。
**English / Original**
There’s going to be a lot of progress in this area. At AMI Labs, we’re more focused on the world-model side. Let’s say you have a patient with a chronic disease like diabetes. What course of treatment should you follow to keep that patient’s blood sugar under control? Presumably, the answer is specific for every patient. People have different physiologies, different gut microbiomes, and respond differently to food.
## paragraph 104
**中文**
如果你拥有患者生理的个性化动力学模型，也就是该患者的世界模型，就可以用它来规划治疗方案。类似地，假设你有某位患者的一枚干细胞；患者患有 1 型糖尿病，而你想把该干细胞转化为能够分泌胰岛素的胰腺 β 细胞。你应该以不同蛋白质的形式向细胞发送哪些信号，使它自行转化为 β 细胞？这就是一个世界模型问题。
**English / Original**
If you had a personalized dynamical model of a patient’s physiology, a world model of that patient, you could use it to plan a course of treatment. Similarly, imagine you have a stem cell from a particular patient. Let’s say the patient has type 1 diabetes, and you want to turn that stem cell into a pancreatic beta cell that produces insulin. What messages should you send to the cell, in the form of different proteins, so that it turns itself into a beta cell? It’s a world-model problem.
## paragraph 105
**中文**
那么 2030 年的你自己呢？你预计自己的研究重点或时间安排方式会有所不同吗？
**English / Original**
What about yourself in 2030? Do you expect your research focus, or the way you spend your time, to be different?
## paragraph 106
**中文**
好吧，我这个月满 66 岁。到 2030 年，我就 70 岁了。我只希望到那时，我的大脑还没有变成白酱。我年纪太大了，不能再把时间花在自己并不擅长的事情上。这就是我没有担任 AMI Labs 首席执行官的原因。
**English / Original**
Okay, so I’m turning 66 this month. By 2030, I’ll be 70. I just hope my brain hasn’t turned into béchamel sauce by then. I’m too old to spend time doing things I’m not the best at. That’s why I’m not the CEO of AMI Labs.
## paragraph 107
**中文**
我是执行董事长，但我的主要工作是提供科学方向，激励人们研究我认为最有前景的问题，然后亲自参与其中一些项目。这也是我在 Meta 最后几年里所做的事。我不做管理，也不太擅长管理。我感到时间正在流逝，不想把时间花在那上面。首席执行官 Alex LeBrun 比我擅长得多。
**English / Original**
I’m the Executive Chairman, but most of what I do is provide scientific direction, inspire people to work on the problems I think are the most promising, and then get my hands dirty on some of the projects. That’s what I was doing during my last few years at Meta. I don’t do management. I’m not very good at it. I see time running out, and I don’t want to do that. Alex LeBrun, who is the CEO, is much better at it.
## paragraph 108
**中文**
我人生和职业生涯的目标，一直是揭开智能之谜。作为工程师，做到这一点的最佳方式就是建造智能机器。最终目标是增强人类智能，把这些益处带给全人类；我认为这本质上是一件好事。
**English / Original**
My goal in life and in my career has been to uncover the mysteries of intelligence. As an engineer, the best way to do that is to build intelligent machines. Ultimately, the goal is to amplify human intelligence and bring those benefits to humanity, which I think is intrinsically good.
## paragraph 109
**中文**
所以我希望自己能在这个方向上取得尽可能多的进展，或帮助其他人取得尽可能多的进展。直到最近，做这件事最好的环境还是 Meta 的 FAIR。现在则是 AMI Labs，因为我们需要加速进展、扩大规模并加以应用；在初创企业中做这些，比在一家只把相关应用视为边缘兴趣的大公司里更好。
**English / Original**
So I’d like to make as much progress, or help others make as much progress as possible in that direction. Until recently, the best environment to do this was at FAIR at Meta. Now it’s AMI Labs, because we need to accelerate progress, scale it up, apply it, and it’s better to do this in a startup than in a large company whose interest in those applications is only peripheral.
## paragraph 110
**中文**
你从哪些领域获取 AI 研究所需的思想输入？
**English / Original**
What are the places you get intellectual input from for AI research?
## paragraph 111
**中文**
神经科学、认知科学和物理学。神经科学的原因显而易见。哺乳动物、鸟类和灵长类动物的心智是如何组织的？是什么让它们学习得如此之快？就像我前面说的，青少年为何能在 15 小时内学会开车？大概是因为世界模型。我们脑中有世界模型吗？有，毫无疑问；它就是前额叶皮层。
**English / Original**
Neuroscience, cognitive science, and physics. Neuroscience, for obvious reasons. There are questions about the organization of the mind in mammals, birds, and primates. What enables them to learn so fast? As I was saying before, how is it that adolescents can learn to drive in 15 hours? Probably because of a world model. Do we have a world model in our heads? Yes. There’s no question we do, it’s the prefrontal cortex.
## paragraph 112
**中文**
还有许多我们不理解的事情。我们拥有定义自己所设任务的目标函数，其中一些由进化硬编码。饥饿就是一个例子：它告诉你饿了，应该采取行动，却不告诉你该怎么做。进化规定的是目标函数，而不是实现这个目标的行为。
**English / Original**
There are a lot of other things we don’t understand. We have objective functions that define the tasks we set for ourselves. Some of them are hardwired by evolution. Hunger is an example. It tells you you’re hungry and that you should do something about it, but it doesn’t tell you how. Evolution specifies the objective function, not the behavior to fulfill this objective.
## paragraph 113
**中文**
你必须通过学习掌握行为。大多数动物靠自己学习，人类则会得到父母及其他同类的帮助。
**English / Original**
You have to learn that. Most animals learn it by themselves. Humans get help from parents and other members of the species.
## paragraph 114
**中文**
因此，神经科学一直是我的巨大灵感来源。例如，卷积网络受到视觉皮层架构的启发。它们也有数学依据，但灵感来自神经科学。同样，鸟类启发了飞机。细节不同，但飞行的底层原理相似。
**English / Original**
So neuroscience has always been a huge source of inspiration for me. Convolutional networks, for example, were inspired by the architecture of the visual cortex. There’s mathematical justification for them as well, but the inspiration came from neuroscience. In the same way, birds inspired airplanes. The details are different, but the underlying principles of flight are similar.
## paragraph 115
**中文**
认知科学是另一个灵感来源。儿童怎样学习自己的世界模型？这里存在争论。很多知识仅靠观察就能学到。我在纽约大学有位同事 Karen Adolph，研究婴儿的运动学习。她告诉我，我完全错了——婴儿主要通过互动学习。我不同意。我认为我们能通过观察学到很多，而我们的世界模型也印证了这一点。
**English / Original**
Cognitive science is another source of inspiration. How do children learn their world model? There’s a debate there. A lot is learned simply by observation. I have a colleague at NYU, Karen Adolph, who studies motor learning in infants. She tells me I’m completely wrong—that babies learn mostly by interaction. I disagree. I think we can learn a lot by observation, and we see that in our world models.
## paragraph 116
**中文**
不仅 AI 受到认知科学启发，AI 也会反过来影响认知科学。深度学习已经如此。如今，神经科学家用于理解视觉皮层怎样处理图像的一些最佳模型，就是深度神经网络，尤其是 Transformer。这是深度学习对神经科学的一项重大贡献。
**English / Original**
Not only is AI inspired by cognitive science, but it also has an effect on cognitive science. That has been the case with deep learning. Now, some of the best models neuroscientists have for understanding how the visual cortex processes images are deep neural networks, particularly transformers. That has been a big contribution of deep learning to neuroscience.
## paragraph 117
**中文**
认知科学中还有快速学习的概念。儿童学习语言所需的数据远少于大语言模型。我经常举例说，大语言模型通常用 10¹⁴ 字节进行预训练，约合 30 万亿个 token，或大约 20 万亿个单词。这相当于 40 万年的阅读材料。
**English / Original**
Cognitive science also has the idea of fast learning. Children learn language with way less data than LLMs. I often cite that LLMs are typically pretrained with 10¹⁴ bytes, about 30 trillion tokens, or around 20 trillion words. That’s about 400,000 years’ worth of reading material.
## paragraph 118
**中文**
儿童到四岁就能学会语言。四年中清醒时间总计约 1.6 万小时。人类确实在海量数据上接受训练，但大部分数据来自视觉、触觉和听觉。语言出现得较晚，而且只占数据中极小一部分。然而，仅凭相对少量的语言输入，我们就能以惊人的速度学会语言。这是怎样发生的？
**English / Original**
Children learn language by the age of four. The total amount of wake hours in four years is roughly 16,000 hours. Humans are trained on enormous amounts of data, but most of it comes through vision, touch, and audition. Language comes late, and it’s only a tiny fraction of data. Yet we learn incredibly quickly from comparatively little linguistic input. How does that happen?
## paragraph 119
**中文**
我与认知科学领域的同事广泛讨论过这个问题，包括 Meta 的 Emmanuel Dupoux，以及 Alison Gopnik 等人。
**English / Original**
That’s something I’ve discussed extensively with colleagues in cognitive science, including Emmanuel Dupoux at Meta and people like Alison Gopnik.
## paragraph 120
**中文**
然后是物理学。为什么是物理学？因为机器学习中推断所依据的数学，与统计物理完全相同；两者使用的是同一套数学。概念也一样：能量、概率分布、吉布斯分布。
**English / Original**
Then there’s physics. Why physics? Because the mathematics underlying inference in machine learning is identical to statistical physics, it’s the same mathematics. The concepts are the same: energy, probability distributions, Gibbs distributions.
## paragraph 121
**中文**
但联系不止于此。你可以把神经网络视为一个物理系统。要理解它，就必须理解它的动力学，而我们在这方面拥有的最佳工具来自统计物理。
**English / Original**
But it goes beyond that. You can think of a neural network as a physical system. To understand it, you need to understand its dynamics, and the best tools we have for that come from statistical physics.
## paragraph 122
**中文**
这也与 JEPA 有关。我一直主张，理解任何事物的唯一方式，是抽象掉不必要的细节，从而能够作出预测。这是所有科学、尤其是物理学的根本原则。我可以用量子场论描述此刻我们之间发生的一切，但那完全不实用，因为它会预测无数我们根本不关心的细节。描述此刻我们之间所发生之事的正确层次，是心理学。
**English / Original**
There’s also the connection with JEPA. I’ve argued that the only way to understand anything is to abstract away unnecessary details so that you can make predictions. That’s fundamental to all of science, especially physics. I could describe everything happening between us right now in terms of quantum field theory. It would be completely impractical because it would predict countless details we simply don’t care about. The right level of description for what goes on between us right now is psychology.
## paragraph 123
**中文**
想想抽象层级。在物理学中，有量子场、粒子、原子和分子；在生物学中，有蛋白质、细胞器、细胞、生物体、社会和生态系统。每一层都通过忽略其下层的细节来描述现实。这种分层表征的概念至关重要。我认为几年后它会变得如此显而易见，以至于人们会奇怪我们过去为什么对此如此视而不见。
**English / Original**
Think about the hierarchy of abstractions. In physics, you have quantum fields, particles, atoms, and molecules. In biology, you have proteins, organelles, cells, organisms, societies, and ecosystems. Every level describes reality by ignoring the details of the levels below. This notion of hierarchical representations is absolutely fundamental. I think it will become so obvious in a few years that people will wonder why we were so blind to it.
## paragraph 124
**中文**
说到未来，哪些科幻作家对你的思考影响最大？
**English / Original**
Speaking of the future, which science fiction writers have influenced your thinking the most?
## paragraph 125
**中文**
Arthur C. Clarke 是其中之一。其他作家的作品我觉得很有娱乐性，但未必描绘了可信的未来，更像是在提出有趣的问题，其中很多问题的政治性强于科学性。
**English / Original**
Arthur C. Clarke. He’s one of them. The others I find entertaining, but not necessarily describing a believable future, more like asking interesting questions, many of which are more political than scientific.
## paragraph 126
**中文**
有一位法国作家写过一篇短篇小说。他描绘了一个社会：人们家中的一切都是灰暗、乏味的，没有任何个性。人人都戴着增强现实眼镜，让世界看起来很美好，但事实上那不是真实世界。
**English / Original**
There is a short story by a French author. He describes a society where everything in people’s houses is gray and bland. Nothing has any kind of character. Everybody wears augmented reality glasses that make the world look great, but in fact, it’s not real.
## paragraph 127
**中文**
那是一篇法国短篇小说。这个装置叫 Oreyeux——在法语中意为“耳朵—眼睛”。那些人显然完全受政府控制，因为政府可以创造任何想让他们看见的现实，并把它美化，尽管那并不真实。所以，它算是《黑客帝国》的前身，只是没那么极端。
**English / Original**
It’s a French short story. This gadget is called the Oreyeux — ear-eyes in French. Those people are obviously being completely controlled by the government, because they can create whatever reality those people want and make it look nice, even though it’s not. So it’s a predecessor to The Matrix, but less extreme.
## paragraph 128
**中文**
这很有趣。我无法想象人类会真的转向那种模式，因为它违背人性。但如果你的 AI 助手通过智能眼镜——就像我现在戴的这种——为你服务，而它来自美国西海岸或中国的少数几家公司，那确实存在危险。民主会陷入严重困境，文化会陷入严重困境，语言多样性也会陷入严重困境。世界上大多数政府都不会接受这种情况。这基本上等于把你的所有电视频道和报纸都交给一家外国实体。
**English / Original**
It’s interesting. I can’t imagine humanity will ever move to that mode. That’s contrary to human nature. But there is a danger if your AI assistant, using your smart glasses, like the ones I’m wearing right now, comes from a handful of companies on the West Coast of the US or China. Democracy is in big trouble. Culture is in big trouble. Linguistic diversity is in big trouble. Most governments in the world would not accept it. It’s basically like giving away all your TV channels and newspapers to a single foreign entity.
## paragraph 129
**中文**
Arthur C. Clarke 的作品之所以给我启发，是因为他大量探讨生命的意义：智能是什么？我们为什么在这里？人类的未来是什么？人类的过去是什么？智能是怎样出现的？这些主题都可以在《2001：太空漫游》中看到，那部作品有点像他钟爱主题的浓缩。我觉得这些很有启发。
**English / Original**
What I find inspiring about Arthur C. Clarke is that a lot of his writing is about what life is about. What is intelligence? Why are we here? What’s the future of humanity? What’s the past of humanity? How did intelligence pop up? You see this in 2001: A Space Odyssey, which is a bit of a concentrate of all the themes that he likes. I find that inspiring.
## paragraph 130
**中文**
你一直公开发表逆主流的立场。这在职业和个人层面给你带来了什么代价？随着 AI 被赋予如此重的政治和经济含义，做一名异议者是变得更容易，还是更困难？
**English / Original**
You consistently publish a contrarian position. What does that cost you, professionally and personally? And is it getting easier or harder to be a dissenter as AI becomes so politically and economically loaded?
## paragraph 131
**中文**
嗯，我并不是为了唱反调而唱反调。我对这个领域会怎样发展有一些想法。如果我相信这些想法，就应该谈论它们。至于这些想法与行业主流相悖，我为什么要在意？
**English / Original**
Well, I’m not a contrarian just to be a contrarian. I have some ideas about where I think the field will evolve. And if I believe in those ideas, I should talk about them. Now, if those ideas go against the mainstream of the industry, why should I care?
## paragraph 132
**中文**
也许人们不会听我的。有些人在我把东西发到社交媒体上时，会觉得我是个白痴，或认为我不了解 AI 的进展，这颇为讽刺。但过去 AI 经历过一轮又一轮浪潮，我见过人们的兴奋，也见过兴奋如何消退。
**English / Original**
Maybe people won’t listen to me. Some people, when I put something on social media, think I’m an idiot or that I’m ignorant of the progress of AI, which is kind of ironic. But I’ve seen the excitement and the deflation of that excitement for various waves of AI in the past.
## paragraph 133
**中文**
人们总是低估，以足够可靠、能够产生经济价值的方式部署某种东西有多难。专家系统就是如此。当然，有些应用确实产生了经济价值，但事实证明，把人类专家的知识转化为规则和事实来构建应用，要复杂、昂贵得多。因此，这只对少数应用可行。
**English / Original**
People have always underestimated how difficult it is to deploy something in a way that’s reliable enough to be economically useful. That was the case for expert systems. Some applications, of course, had economic value, but it turned out to be much more complex and expensive to build applications by transferring the knowledge of a human expert into rules and facts. So it was only practical for a small number of applications.
## paragraph 134
**中文**
随后是 20 世纪 80 年代末至 90 年代初神经网络的第二波浪潮。当时的困难是算力和数据获取。只有少数应用拥有足够数据来取得进展，基本上包括文档理解、光学字符识别（OCR），以及一定程度上的语音识别——尽管语音识别直到最近才具备商业可行性。此外，还有一些应用可以使用很小的神经网络和相对少量的数据，例如欺诈检测、信用卡使用，以及某些控制问题。
**English / Original**
Then there was the second wave of neural nets in the late '80s and early '90s. What turned out to be difficult was compute power and access to data. There was only a small number of applications for which enough data was available to make progress. That was basically document understanding, OCR, and, to some extent, speech recognition—although that didn’t become commercially practical until recently. Then there were applications where you could get away with very small neural nets and relatively small amounts of data, like fraud detection, credit card use, and some control problems.
## paragraph 135
**中文**
20 世纪 60 年代初，人们研究感知机和 ADALINE 时，也出现过同样的情况。那些早期模型让人以为，他们既然已经弄清了怎样训练单层系统，最终就能建造智能机器。但到 60 年代末，人们已经清楚，这些技术无法扩展到多层系统。
**English / Original**
You saw the same thing in the early '60s, when people were working on the Perceptron and ADALINE. Those early models led people to think they were eventually going to build intelligent machines because they had figured out how to train a single-layer system. But by the late '60s it became clear that these techniques couldn’t be extended to multilayer systems.
## paragraph 136
**中文**
人们仍继续研究那些思路，只是改变了工作名称。其中一个分支走上工程路线，成为自适应滤波器，对技术产生了巨大影响。你的手机就使用了许许多多自适应滤波器，它们是 20 世纪 60 年代发明技术的直接后代。另一个分支演变为当时所谓的统计模式识别，最终成为今天神经网络和机器学习的基础。
**English / Original**
People kept working on those ideas—they just changed the name of what they were doing. One branch went down the engineering route and became adaptive filters. That had a huge impact on technology. Your cell phone uses adaptive filters, tons of them. They’re the direct descendants of techniques invented in the ‘60s. Another branch became what was then called statistical pattern recognition, which eventually became the foundation of neural nets and machine learning as we know them today.
## paragraph 137
**中文**
这种兴奋周期总会出现：“我们有了这项新技术。十年后，地球上最聪明的实体将是一台机器。”70 年前这是错的，今天依然是错的。
**English / Original**
There’s always this cycle of excitement. “We have this new technique. Ten years from now, the most intelligent entity on the planet will be a machine.” It was false 70 years ago. It’s still false.
## paragraph 138
**中文**
随后，预期会被重新校准，那些技术则被纳入工程工具箱。抛开炒作，这正是大语言模型今天正在经历的事情。之后，我们还会看到另一波浪潮。
**English / Original**
Then expectations get renormalized, and those techniques are brought into the engineering toolbox. That’s what’s happening with LLMs today, beyond the hype. We’ll see another wave.
## paragraph 139
**中文**
你预计这些预期何时会被重新校准？
**English / Original**
When do you expect these expectations to renormalize?
## paragraph 140
**中文**
对大语言模型而言，这会在两年内发生。但之后还会出现另一波浪潮，或许基于世界模型，也可能基于 JEPA；它有望扩展 AI 技术的适用范围。随后，我们会在此基础上开发出许多东西，又碰到其他限制，再等待另一个概念性想法带领我们迈出下一步。
**English / Original**
For LLMs, it’s going to happen within two years. But then there’ll be another wave, perhaps based on world models and maybe JEPA, which will hopefully extend the applicability of AI techniques. Then we’ll develop a lot of things on top of that, and hit other limitations. And we’ll have to wait for another conceptual idea to take us to the next step.
## paragraph 141
**中文**
那么，要达到人类水平的智能，我们究竟需要多少次这样的概念突破？这正是我们不知道的。它就像爬山：我们能看到眼前这座山，却不知道它后面还有多少座山。只有登上山顶时才会知道，而后面的山很可能更高。
**English / Original**
Now, how many such conceptual breakthroughs do we need to reach human-level intelligence? That’s what we don’t know. It’s like climbing a mountain. We can see the mountain in front of us, but we don’t know how many mountains are behind it. We won’t know until we get to the top. And the following ones are probably taller.