记忆与持续学习：AI 的下一个前沿

欢迎来到 Training Data。今天我们非常荣幸邀请到 Engram 的联合创始人 Dan Biderman 和 Jesse Lin。Engram 是一家专注于记忆与持续学习的新实验室，这两个是当今 AI 研究中最热门的话题。我和 Shawn 很高兴今天能与你们深入探讨这个话题。

Welcome to Training Data. We are delighted to have Dan Biderman and Jesse Lin, co-founders of Engram today. Engram is a new lab focused on memory and continual learning, two of the hottest topics in all of AI research today. Shawn and I are delighted to dig in on this topic with you today.

太棒了，很高兴来到这里。

Awesome. Happy to be here.

好的，那我就开始吧。Engram 的网站上写着：“我们不以预训练或后训练的视角看待世界。我们的模型始终在训练。” 这是什么意思？

Great. So I'm going to kick off. The Engram website says, "We don't see the world through the lens of pre-training or post-training. Our models are always training." What does that mean?

我认为如今的模型显然知道很多事情，它们非常聪明。但我们觉得，当前让这些模型更有用的瓶颈并非原始智能，而是理解新的、不断变化的上下文。无论是你正在做的新任务，还是某个工作场景下的特定上下文。如何像预训练和后训练那样，将这些信息深度嵌入模型权重？这正是我们致力于解决记忆与持续学习这些根本问题的原因，它们其实是同一枚硬币的两面。如何让模型学习新事物，并将其深度融入模型权重？

So I think models today obviously know a lot of things. They're incredibly smart. But we kind of think the bottleneck for making these models more useful these days is not really raw intelligence but understanding new and evolving context. Whether it's a new task that you're doing or a particular context for a job or something like this. How do you bake that into the model weights the same way that pre-training and post-training makes that into the model weights very deeply? And this is kind of why we think of ourselves as working on these fundamental problems of memory and continual learning, which are really two sides of the same coin. How do you make the models learn new things and bake them deeply into the weights of the model?

那么你们的前提是，将记忆作为单独的数据库或塞进上下文窗口的东西，并非真正的记忆，也不是真正的持续学习？

And is your premise then that memory as a separate database or separate thing that you've shoved into the context window is not true memory and is not true continual learning?

我认为所有这些工具都会融合在一起。如今，人们解决这些问题的方式是上下文工程。你使用一个巨大的提示词，也许与模型进行多轮、数小时的对话，并重新组织上下文以更好地理解你的意图。我们认为工具使用、上下文工程这类方法会发挥作用，但目前一个未被充分利用的工具是采用前沿实验室用于让模型擅长前沿数学或代码的同类训练流程、框架或工作流，并将其应用于每个领域、每个上下文，比如公司内部。

I think all of these tools will kind of come together. So these days, the way that people are solving these problems is with context engineering. So you take a huge prompt, maybe you keep talking to the model over many turns and hours, and reorganize the context to better understand what you're trying to do. And we think these kinds of things like tool use, context engineering will play a part, but I think an under-leveraged tool these days is using the same kind of training pipeline or framework or workflow that the frontier labs are using to make these models really good at frontier math or code, but applying that to every kind of domain, every kind of context that you have, like in a company.

是的。对我来说，就像个人一样，记笔记和贴便利贴是非常有价值的事情，我们永远不应该抛弃它。但每当第二天回到工作中时，我们的大脑中总会留下某种记忆痕迹，一些关于事情应该如何进行以及应该关注哪里的新信息。所以这两者应该结合起来。而当前的解决方案更多是外部化记忆。这有两个问题：一是我们所有人、每个人每天生成的 token 量很快将达到数千万。因此，仅仅保存、搜索和重新读取这些 token 将非常昂贵，而且对模型来说也会非常困难、非常混乱，除非我们在方法上取得重大突破。

Yeah. And to me, it's like as an individual, taking notes and having sticky notes is a very valuable thing. We should never discard this, but whenever we get back to business the next day, we always have some sort of trace of memory in our brain, some new information about how things should be and where we should look. So these two things should come together. And current solutions are more kind of externalized memory. And this has two issues. One is that the amount of tokens we will all collectively, individually generate is going to be in the tens of millions of tokens per day soon. So just keeping it and searching through it and rereading it is going to be pretty expensive, but it's also going to be pretty hard and pretty confusing for the models unless we have major breakthroughs in how we do it.

对 Sean 来说的 token 量？

Of tokens for Sean?

不错。

That's good.

看情况。

Depends on the day.

能否请你介绍一下 Engram 的架构、产品以及它是如何工作的？

Could you maybe tell us a little bit about the Engram architecture, the Engram product and how it works?

是的，从高层来看，我们试图处理任何上下文。假设存在各种不同的工作空间。我们正在与 Notion、Microsoft 和 Harvey 等合作伙伴合作，这些平台上有人们长期进行大量工作的场所。那里有丰富的上下文，既包括团队已编写的文档，也包括人们在这些产品中与智能体日益频繁的互动。我们进行对话、提供反馈，并致力于让模型深度理解这些上下文。不仅仅是测试时读取文件，而是像在公司工作多年的员工那样真正理解。你从高层了解公司的各项举措、做事方式，掌握了如何运行招聘流程或如何在公司内完成某项任务，并能像公司其他任何人一样高效运作。因此，我们正在这些工作空间内训练每个团队的模型，这些模型深度理解那些上下文，并能随着时间推移在人们关心的事情上不断改进。从技术层面来说，我们通过将知识训练到权重中来实现这一点。我们大量使用适配器微调。适配器有多种类型，我认为人们已经研究了几十年，无论是 LoRA、前缀微调还是稀疏架构。所有这些工具都可供我们使用。然后要确定正确的数据是什么。如何将任何原始文档或交互转化为对模型有用的训练信号。同样，我们现在有各种工具，如监督微调、强化学习、在线策略蒸馏等，这些都是该领域已经发展出来的。我们试图将这些组件整合成一个模型，使其能够持续学习人们关心的事物。

Yeah, at a high level, I think what we're trying to do is take any context. There are all these different workspaces, let's say. So we're working with partners like Notion, Microsoft, and Harvey that have these places where people are doing a lot of work over a long period of time. There's all of context, both in terms of documents that you've already written as a team, as well as now people are interacting with these agents more and more in these products. We're having conversations, giving them feedback, and figuring out how to have a model that deeply understands that context. So not just reading the files at test time, but really understanding it the way that an employee that's worked at your company for years has. So you kind of understand at a high level, oh, these are the initiatives across the company. This is the way that we do things. You've studied how to run the hiring pipeline or how to do this kind of thing within the company, and can operate just as well as anybody else can in the company. And so what we're doing is training per team models within these workspaces that deeply understand those contexts and can improve with time on the things that people care about. So the way that we do this at a technical level maybe is training these into weights. So we do a lot of adapter fine-tuning. Adapters are many types. I think people have looked into this for decades at this point, whether it's LoRAs or prefixes or sparse architectures. I think all of these tools are at our disposal. And then figuring out what the right data is. So how do you turn any kind of raw document or interaction into useful training signal for the model. So again, we have a variety of tools now like supervised fine-tuning, RL, on-policy distillation, all of these things that the field has developed, and trying to fit these pieces together into a model that learns continuously on the things that people care about.

是的。这并非赌工具不存在。我们的模型始终假设某些知识是外部化的，某些工具始终存在。但你需要做的是——这也是困难之处——判断哪些需要内化，哪些可以外化。即使对于外部化的内容，许多个人和公司也有自己定制的工具和做事方式。并非每个人都拥有前沿模型训练时所用的相同 bash CLI 工具。如何让所有这些工具更好地理解你的定制设置，我认为这本身就是一件有趣的事情。

Yeah. And it's not a bet that tools are not there. Our models always work under the assumption that some knowledge is externalized. Some tools are always there. But what you need to do is you need to figure out, and that's the hard task, is what needs to be internalized and what can be externalized. And for even stuff that's externalized, many individuals and companies have their own bespoke tools and ways of doing things. Not everyone has the same bash CLI tools that the frontier models are training on, and how to get them all to better understand your bespoke setup, I think, is its own interesting thing.

那么，前提是不是我的 Notion 智能体会是一个定制智能体，经过 Laura 微调，或者通过某种适配器调优，从而能持续学习我 Notion 工作区中添加的新内容？这是前提吗？

And so, is the premise then that my Notion agent will be a custom agent that is Laura fine-tuned or, you know, some way with an adapter tuned so that it's constantly learning on new content that's added into my Notion workspace? Is that the premise?

是的，他们正在与许多模型合作，是所有前沿模型的早期用户，而且可能会继续这样做。

Yeah, and they're working with many models and they're the early users of all the frontier models and they're probably going to keep doing that.

这种方法对前沿模型有效吗，还是说它们是封闭的？

Does this approach work on the frontier models or are they closed?

所以，我们需要白盒访问权重，对吧？我们可以与拥有闭源权重的公司合作，与他们一起做这件事。但对我们来说，用开源模型做最容易。不过，只要是 Transformer 模型，我们都能对它进行操作。

So, we need white box access to the weights, right? So, we can partner with companies that have closed-source weights and do this with them. But it's easiest for us to do it with open-source models. But any model that's a transformer model, we can do our thing to it.

那么，当人们比较使用你前后的情况时，权衡是什么？是不是他们不再发送那么多上下文了？所以，权衡就像你前期消耗更多算力，将公司的做事方式学习到权重中，然后在每次推理时向模型发送更少的上下文。这是大致的权衡吗？

And what's the trade-off then when people are comparing the before and after using you? Is it that they're no longer sending so much context? And so, the trade-off is like you burn more compute up front to learn your company's way of doing things into the weights and then you're sending less context to the model on every inference pass. Is that the rough trade-off?

这是一方面。你不需要研究、重读，也不需要写庞大的系统提示，这些本身就能将 token 推理消耗降低两个数量级。不是 50%，而是可能减少 100 倍的 token，因为很多事情，尤其是与人、团队、组织和优先级相关的事情，你无法在一份文档中找到，除非你真的把它管理得井井有条并记录一切。这类事情模型可以通过在部分数据上训练来隐式学习，并在 100 个 token 内回答，而最好的前沿模型需要消耗 10 万个 token。所以，这些例子很有趣。还有质量，有些任务对当前一代模型来说并非超自然，我们觉得会持续存在 3 到 6 个月的差距，某些定制化的事情人们正在探索，模型对它们还不够好。模型最终会变得很好，但如果你能以非常轻量的方式自主学习，那段时间内它就能在能力上带来价值。

That's one thing. The fact that you don't have to research things and re-read things and the fact that you don't have to write monstrous system prompts, they're already that can give you two orders of magnitude reduction in token inference consumption. It's not like 50% or it can be 100x fewer tokens because many things, especially things that relate to people and teams and organization and priorities, these are things that you can't really find in one document unless you really have it regimented and document everything. These kinds of things the model can kind of implicitly learn by training on some of the data and answer within 100 tokens what the best frontier models would consume 100,000 tokens doing. So, these kinds of examples are interesting. And also the quality, there are tasks that are not supernatural for the current generation of the models and we kind of think there's going to be consistently this gap of like 3 to 6 months ahead where certain things are bespoke that people are just exploring, the models are not fully great for them. The models will at some point be great for them, but if you can autonomously learn in a very lightweight way, it will give value in that time in terms of capabilities.

为什么在工作区层面训练，而不是在个人层面，比如？

Why train on the workspace level versus the individual level, for example?

对我们来说都可以。只是从团队开始更容易，他们在如何收集上下文以及多年积累的上下文数量上更有纪律性，我们容易从那里开始。但每个人的电脑和手机有一天都会成为我们技术的有用目标，我认为去那里会非常有趣。我们只是觉得现在大量的信息沉积在协作知识工作的团队中。

Either is fine for us. It's just easier to start with teams of people have more discipline in how they collect context and in the amount of context they have over years and it's easy for us to start there, but every person's computer and every person's phone one day is a useful target for our technologies and I think it will be very interesting to go there. We just think the big deposits of information are now in teams of people collaborating in knowledge work.

大型语言模型中内置了这么多事实记忆，这是特性还是缺陷？有一种观点认为，模型只是记住“法国首都是巴黎”这个事实实际上是一件坏事。我们更希望模型做的是抽象地学习国家和首都城市的概念，而不是在权重中记住所有这些事实。所以，我很好奇你如何看待将记忆与学习分离，当前模型是如何做的，以及你打算如何处理。

Is it a feature or a bug that there is so much fact memorization basically built into large language models? And there's a school of thought that the models just memorizing the fact that the capital of France is Paris is actually a bad thing. And what we would prefer for the models to do is abstractly learn the concepts of countries and capital cities. But not to memorize all these facts in the weights. And so, I'm curious what you think about disentangling memorization versus learning, how it's done in the models today, and then how you're thinking of approaching it.

是的。我认为这是一个非常有趣的问题。在某种程度上，你需要记住一些东西，才能将它们组合成更复杂的概念。我觉得缺失的部分是弄清楚什么值得记住。即使现在，当你考虑学习新知识时，看看很多学术基准，它们关注的是如何学习非常具体的事实，比如某个非洲国家一座桥的长度。这其实不是你想让模型投入能力的地方，也不是我们投入能力的地方。所以，如果你看人类记忆，它是有损的，因为智能的未来一部分在于压缩重要的东西，并将其与不重要的东西分开。所以，我认为你无法像有些人想的那样，将事实学习与非事实学习或技能学习分开。如果你拿一个模型，剥离所有事实，只留下纯粹的核心，那模型会非常不自然。它不知道基本的东西。而你需要那些基本知识。但为什么需要呢？为什么不能查事实，然后直接……如果你看模型如何思考，如果需要回忆基本事实才能进行下一步思考，那你就走不远。也许这是一种高层直觉，但这也是我们认为训练非常重要的原因之一。为了对事物进行越来越复杂和深入的思考，你需要内化一些东西，以便将它们组合成更抽象的概念。

Yeah. I think it's a really interesting question. To some extent you kind of need to remember stuff in order to compose them into more complex concepts. I think the thing that's kind of missing is figuring out what's important to remember. And I think even now when you think about learning new knowledge, if you look at a lot of these academic benchmarks, it's like how can we learn very specific facts like the length of a bridge in this African country. And that's not something that you really want the models to devote capacity for, and it's not something that we devote capacity to. So I think if you look at human memory, it's lossy because part of the future of intelligence is compressing what's important and separating that from what's not important. And so I think you can't really separate fact learning from non-fact learning or skill learning as some people would like to think. If you take a model and strip out all the facts and just have it like the pure core, it's very unnatural as a model. It doesn't know basic things. And you kind of need that. But I think why do you need that? Like why can't you look up facts and then just have... I think if you look at how the models think, if you need to recall basic facts in order to take the next step in your thinking, you can't get very far. Maybe that's like a high-level intuition, but it's part of the reason why we think training is really important. In order to think more and more complex and deep thoughts about things, you kind of need to internalize something so that you can compose them into more abstract concepts.

是的，以前有过一些努力，但很难规模化，试图将两者分离，并以允许模型检索和搜索事物而不内化它们的方式预训练模型。只是我们现在集体爬山所知道的配方就是这种事实预训练。我认为这种方法的神奇或神秘之处在于，传统上在计算机科学中，我们有数据库作为自己的课程，有算法，数据库就像关于世界的事实和首都之类的东西，存储它们，查询它们。还有算法，如何高效地操作信息并以样本高效的方式得到答案。我认为深度学习的神奇之处在于，这两者现在混在一起了，我们需要所有这些聪明的可解释性专家来尝试将它们分开。

Yeah, and there have been efforts before that were hard to scale to try and disentangle the two and pre-train the models in a way that allows it to retrieve and search for things and not internalize them. It's just the recipe we know to hill climb on collectively right now is this fact pre-training stuff. And I think the magic or the mystery of this approach is that traditionally in CS we would have databases as its own curriculum and we would have algorithms and the databases is like facts about the world and capitals of whatever store them query them. There's also algorithms of how do you efficiently manipulate information and get some answers in a sample efficient way and I think the magic of deep learning is that these two things are now mushed together and we need all these smart people on topic interpretability to try and break them apart.

我认为，在 AI 融入经济的过程中，我们看到这些事物正在逐渐分离：公司拥有自己的上下文，它们精心处理并工程化这些上下文，而通用模型对这些上下文完全陌生，模型只是在上面操作。但对我们来说，显然需要某种程度的融合——至少以一定的节奏，让事实、故事和细节融入模型。这也有缺点，因为像国家首都这样的东西，虽然变化不频繁，但还有很多其他事实在不断变化。把它们直接印刻到权重里是很有挑战性的。

And I think a lot of what we're seeing now in the adoption of AI into the economy is that these things are gradually separating again where companies have their own context and they really handle them with care and engineer them with care and there's a generic model that's completely a stranger to these contexts and that the model is operating on them but it for us it's clear that there needs to be a certain convergence at least with some cadence where the facts and the stories and the details are getting mixed into the model. It has disadvantages as well because if you have to, you know, capitals of countries, they can change but it's not very frequent but there's many other facts that are changing all the time. Just imprinting them into weights is a challenging thing to do.

我明白了。所以你是说，试图在这里把算法和数据库分开是一种错误的二分法。真正重要的是如何区分哪些值得记住，

I see. So you're saying it's a false dichotomy to try to separate algorithms from databases here. What really matters is like how to distinguish what's important to remember

对。

Yeah.

哪些不重要。

versus what's not important.

正是。这是一个开放性问题。

Exactly. And it's an open question.

你们在排序方面有从中获得什么启发吗？

And are you guys taking any inspiration from that in terms of ranking?

我觉得非常松散。只是说，这可能是一个缺失的阶段：你拿到一个上下文，然后深度内化它。现在的情况是，一切都在测试时发生。你看着用户给你的上下文，然后即兴思考。但同样，你走不了太远，或者也许能走一段，但过程中会犯错。如何把这些消化回模型，以便下次做的时候能正确执行并取得更大进步？

Very very loosely I think. Just the idea that that's kind of a phase that's missing maybe where you take a context and you deeply internalize it. Right now it's like everything happens at test time. You look at the context that the user gives you and you do some thinking on the fly. But again like you can't get very far or you can get so far maybe and like you make mistakes along the way. Like how do you digest that back into the model so that next time you do it you do it the right way and make even more progress.

对。那梦是什么呢？说我们要构建一个像梦一样的 AI，听起来有点疯狂。梦没有太多连贯性。但有趣的是，在梦中，我们看到东西，自言自语，实验我们在世界和社交情境中能做什么、不能做什么，而且它严重偏向社交内容，对吧？所以，对我们正在构建的东西来说，我们给模型时间，让它从实际交互中退回来，实验它的能力。它能在环境中做什么？它知道什么？它能多快处理那些我们在夜晚梦到的极端情况？

Yeah. And then what are dreams? Dreams are pretty crazy things to say we want to build an AI that's like our dreams sounds a little bit like a nut thing to do. There's not a lot of coherence there. But what's interesting there is like what happens in our dreams, we see things, we talk to ourselves, and we experiment with the affordances of what we can do and can't do in the world and social situations and then, you know, it's heavily biased towards social stuff, right? So, for us too with things we're building is, you know, we give the models the time to then go back, retreat from the actual interaction, and experiment with its affordances. What can it do in an environment? What does it know? How fast can it handle these kind of tail extreme things that same ones that we dream about at night.

你们有学术背景吗？比如，有什么典型例子激发了这个问题，或者目前取得了什么成功？

Do you guys come from academic backgrounds? Like, what's a canonical example that motivates this problem or, you know, like or that's a win so far?

对。我有一个例子，也许 Jesse 可以再给一个。假设一个 AI 实验室，比如 OpenAI，必须在一周内赢得某个数学奥林匹克竞赛。他们会构建所有数学教科书的目录，让人标注哪些章节要看、哪些图表要看，还是会实际收集这些资料，合成一些训练数据，启动一个训练任务，在 5-6 天后看看结果，开始评估等等？对任何训练过模型的人来说，很明显有一种更优越的方式来整合想法和能力，这涉及训练的神奇之处。我们清楚，这在数学、编程、网络安全等高价值领域必须发生。我们只是认为，这种神奇之处可以以有趣的方式落到更多人手中。

Yeah. I have one example, maybe Jesse can give another one. A hypothetical one, for example, imagine one of the AI labs, say, OpenAI has to win some math Olympiad in a week time from now. Would they construct a catalog of all the math textbooks and really have people annotate which chapters to get and which graphs to see or will they actually collect this, synthesize some training data, launch a training job, see where it lands in 5-6 days, start evaluating it, and stuff like that? So, it's obvious for anyone who's trained model that there's superior way to integrate across the ideas and capabilities, and it involves this kind of magic of training. And we are clear that this has to happen in those high-stake domains of math and coding and cyber and stuff. We just think much of this magic can actually end up in the hands of many more people in interesting ways.

比如，为什么最终产品不是由基础模型实验室独占？你们如何在巨头之间周旋？

Like, why isn't it just the foundation model labs that own the end product here? You know, like, how do you go between giants?

对。我认为我们的世界观与前沿实验室的世界观有些不同。他们想要一个越来越大、越来越智能、跨领域的模型。而我们设想的世界是每个人都拥有自己的模型。人们想学的很多东西要么是私有的——永远不会出现在后训练数据集中，要么甚至是冲突的——比如我做任务的方式与另一家公司或个人不同。我认为，很多这些东西很难用我们几十年来在机器学习中使用的工具训练进模型，那些工具需要非常干净的监督、真实的奖励信号、良好的环境，然后训练模型使用工具更好地完成编程任务。而现实世界中发生的很多事情非常模糊，很难说清什么才是好的。所以，我认为这些东西非常个性化，与前沿实验室对整个训练流程以及长期存在的模型类型的看法不太一致。

Yeah. So, I think like the world view that we have is a bit different from the Frontier Lab worldview, where it's like we want one model that's bigger and bigger, that's more and more intelligent across a variety of domains. Instead, how we see it, like we kind of imagine this world where everybody has their own model. A lot of the things that people want to learn are either private, like things that will never see the light of day in a post-training data set, or even conflicting, like, oh, the way that I want to do the task is different from how another company or another individual wants to. And I think a lot of these things we're already seeing are hard to train into the models with the same tools that we have used for like decades in machine learning, which is like you have really clean supervision, you have like ground truth reward signals, and you like create a nice environment, and you like train the model to like use the tools to better accomplish this like coding task. And instead, a lot of the things that actually happen out in the world are very ambiguous, or like it's hard to say like what makes something good. And so, I think a lot of these things are very specific to individuals, and I think very kind of misaligned or not very aligned with how the Frontier Labs think about the whole training pipeline and what kind of models will exist in the longer term.

对，补充一点。前沿实验室的 P0 是什么？在座有些人跟他们很熟。是达到 AGI，得到一个在编程和数学上极其强大的通用模型，然后用它来自动化经济，或解决密码学、国防等领域的长期难题。要推动这一点，需要更多预训练、更大的模型、更多数据、更多强化学习、更多推理时算力，这很清楚。这是 P0，大部分资金和人才都投在这里。当然，他们都在考虑记忆和持续学习，但目前更多是产品层面的努力。我们认为这值得单独关注，需要突破。大约一个月前，Demis 在一个活动上明确表示，我们需要在这些话题上取得新突破，显然他们也在思考。我们只是专注于这一点。我们认为，关于数据在哪里、谁拥有模型的激励机制非常有趣。如果你能从许多人类或组织那里大规模学习，而不必派人肩并肩地与他们合作，那将是一个巨大的解锁。

Yeah, and to add to it, I think, you know, what is the P0 for the Frontier Labs? And some of you here are pretty close with them. It's getting to AGI, getting this one generic model that's extremely capable in coding and math, and then using it to automate the economy or to solve really hard, long-term problems in cryptography and defense or whatever. And it's pretty clear what needs to happen to push this, you know, some more pre-training, bigger models, more data, more RL, more inference time compute, that kind of stuff. That's P0, that's where the majority of expenditure and talent goes. And definitely all of them are thinking about memory, and all of them are thinking about continual learning. It's just more of a product kind of effort right now. We think it deserves its own attention. And we think breakthroughs need to happen there and Demis and this event about a month ago said pretty clearly that we need new breakthroughs around these topics and obviously they're thinking about them. We're just focusing exclusively on this. And we think certain things around incentives of where the data is and who owns the model are pretty interesting. So, if you could learn from many humans or organizations at scale without necessarily sending someone to work with them shoulder to shoulder, that would be a pretty big unlock.

也许还有一点：我认为世界上很多事情需要改变。一是需要新的研究突破，二是需要新的训练基础设施——比如为每个人提供小模型，而不是一个大模型、一次大运行。

And maybe another point on that is like I think a lot of things need to look different in the world. So one is there needs to be new research breakthroughs, two is new infrastructure for training like, you know, small models for everybody rather than like one big model, one big run.

第三点，我认为是一种将研究与产品结合的不同方式。目前，前沿实验室的研究人员训练模型，然后将其抛给产品团队，由他们通过提示或工程在核心模型之上构建新的产品界面。但在模型持续训练的世界里，用户提供的输入与模型的学习来源——即训练信号——紧密相连。因此，研究与产品之间需要形成一个更加紧密的集成循环。虽然我们专注于攻克核心研究挑战（这也是我们的背景），但我们也非常注重尽快部署，以便从真实世界的反馈中学习。

And then the third I think is a different way of combining research and product. Right now, researchers in frontier labs train the model, then throw it over the fence to the product team, who prompt or engineer new product surfaces on top of the core models. But in a world where models are always training, the inputs users provide are intricately tied to what the model has learned from—the training signal. So there needs to be a much more integrated loop between research and product. While we're focused on tackling core research challenges, which is our background, we're also very focused on deploying this as quickly as possible to learn from real-world feedback.

是什么促使你研究这个问题？

What motivated you to work on this problem?

我认为这显然是 AI 领域的重大挑战之一。如今每个人都在谈论它，因为模型太聪明了。还有什么没解决的呢？关键在于边缘学习——学习那些让模型变得有用的剩余部分。这不再仅仅是原始智能的问题，而是学习新事物的能力。而且这个问题非常基础，因为它回归到了真正理解模型为何如此优秀。目前，模型通过预训练偶然学到了很多东西，我们并不真正理解原因。互联网就像一份礼物，提供了包含编码、写作等各种例子的多样化数据——事情就这样发生了。现在，要攻克持续学习，我们需要弄清楚预训练甚至后训练中的哪些因素使模型能够以这些神奇的涌现方式泛化，并控制这一过程，以便拥有私有数据的公司能让模型像了解法国首都或编写 Python 一样学习那些数据。所以，这是一个非常有趣的问题。

I think it's obviously one of the grand challenges in AI. Everybody is talking about it these days because the models are so smart. What else is left? It's about learning at the edges—learning the remainders of what makes these models useful. It's not just about raw intelligence anymore; it's about learning new things. And it feels very fundamental because it goes back to really understanding what makes the model so good. Right now, models incidentally know a lot from pre-training, and we don't really understand why. The internet was a gift with diverse data containing examples of coding, writing, and all these other things—it just happened that way. Now, to crack continual learning, we need to figure out what about pre-training or even post-training enables models to generalize in these magical emergent ways, and control that process so that a company with private data can make the model learn that data as well as it knows the capital of France or how to write Python. So it's a really fun problem to think about.

Dan，你来自神经科学领域，对吗？

And Dan, you came from the neuroscience world, is that right?

是的。我最初对意识、人类状况等问题感兴趣。模型有意识吗？我没有比你能读到的更深入的想法。我认为没有，但聪明的人思考这个问题很重要。我对人类如何思考和感知感兴趣。正如以色列心理学家阿莫斯·特沃斯基所说，他对人工智能不感兴趣，他对自然愚蠢感兴趣。所以我也是从类似的角度出发，试图了解人和动物如何体验世界。渐渐地，我的兴趣转向了统计学和 AI，我发现记忆和持续学习中的许多相同问题非常紧迫。当前系统中的解决方案与生物学中的相去甚远。我不是那种认为机器应该像动物或人类大脑的人——我不这么认为。计算机可以做很多比我们更好的事情。但人类记忆有着非常不同的特性。如果你想存储整个代码库，你会用电脑——甚至不需要 AI 就能无损存储。但人类大脑是在信息容量的约束下进化而来的，拥有模糊的表征，这些表征可以被抽象、形成连接，并为第二天提供信息。当前系统除了通用的预训练步骤外，没有这些。我对如何构建这种能力并从中学习非常感兴趣。

Yes. I was initially interested in questions around consciousness, the human condition, and things like that. Are the models conscious? I don't have any advanced thoughts beyond what you'd read. I don't think so, but it's important that smart people are thinking about it. I was interested in how humans think and perceive. As the Israeli psychologist Amos Tversky used to say, he's not interested in artificial intelligence; he's interested in natural stupidity. So I started similarly, trying to see how people and animals experience the world. Gradually, my inclinations took me to statistics and AI, where I found that many of the same problems of memory and continual learning are really urgent. The solutions in current systems are far from what we have in biology. I'm not one of those who say machines should be like animal or human brains—I don't think so. Computers can do many things better than us. But human memory has very different properties. If you want to store a whole codebase, you use a computer—you don't even need AI to store everything losslessly. But the human brain evolved to work under constraints of information capacity, with fuzzy representations that can be abstracted, form connections, and inform the next day. Current systems don't have that beyond the generic pre-training step. I was really interested in ways to build that in and learn from it.

这更像一个哲学问题。你提到大脑有不同的区域和协同处理单元。现代计算机架构有 CPU、GPU、内存、不同的协处理器。根据苦涩的教训，你认为 LLM 会收敛到一个主导的协处理器——所有计算都发生在语言模型的 GPU 等价物中？还是这些模型在内部涌现性地构建多个协处理器？对于记忆，你认为模型会自行构建擅长记忆的大脑等价物，还是需要独立的架构？

This is more of a philosophical question. You mentioned the brain has different real estate, co-processing units. Modern computer architecture has CPUs, GPUs, memory, different co-processors. With the bitter lesson, do you think LLMs converge to one dominant co-processor—like all compute happens in the GPU equivalent of a language model? Or do these models build multiple co-processors emergently inside? For memory, do you think models will build whatever brain equivalent is good at memory, or is there a need for a separate architecture?

正是。记忆是涌现属性吗？

Exactly. Is memory an emergent property?

正是。还是说……

Exactly. Versus like...

是的，几乎一切。我们所需的智能是否都会随着更好的训练数据和更多的算力而涌现？

Yeah, and almost everything. Is everything we need in intelligence just emergent with better training data and more compute?

从当前 AI 部署的表面来看，远不止 GPU。我们看到沙盒激增，模型在其他计算机上运行并尝试各种事情。

I would say, on a superficial perspective of current AI deployment, it's way more than just GPUs. We're seeing sandboxes exploding, models operating on other computers, trying things.

在模型架构层面，而不是……

On the model architecture level rather than on the...

所以其他实验——之前有很多不同架构的实验，比如状态空间系列，以更高效地处理极长上下文。所有这些方法最终都是一种权衡，通常是在记忆和准确性之间——这里的记忆是计算机意义上的，而非行为认知意义上的。与 Transformer 注意力机制在序列长度上的二次方内存占用不同，这些模型有……

So other experiments—there have been many previous experiments with different architectures, like the state space family, to handle very long context more efficiently. The thing with all these methods is that it ends up being a trade-off, usually between memory and accuracy—memory in the computer sense, not behavioral cognitive sense. Instead of the quadratic memory footprint of transformer attention in sequence length, these models have...

有些人声称它们具有次二次方复杂度。

Some are claiming they have sub-quadratic.

是的，有些人声称，有些确实做到了。一些最好的中国模型有受那些状态空间架构启发的层，并且成本不是二次方的。问题是，在我们的实践中，我们发现它们总是为了记忆而牺牲准确性。天下没有免费的午餐。

Yeah, some are claiming, some do have it. Some of the best Chinese models have layers inspired by those state space architectures and are not quadratic in cost. The thing is, in our hands, we find that they always compromise accuracy for this memory. There's no free lunch.

我们要说的是，如果你真正贯彻了苦涩的教训，你要做的就是思考：我怎样才能消耗更多算力？我怎样才能把它消耗在从未见过的新语境上？所以我们和其他人一样彻底贯彻了苦涩的教训，我们并不认为 AGI 的整体方向会很快终结。我们只是认为还有更多算力可以扩展。如果我真的想了解 Shawn 以及 Shawn 的工作和语境，仅仅重读文件是不够的，尤其是对于像你这样特别的人。我们得为这家伙训练一个 100 万亿参数的模型。

And what we're saying is, look, if you're really bitter lesson peeled, what you want to do is think: how can I burn more compute? And how can I burn it on new context that I have not seen before? So we're as bitter lesson peeled as anyone else, and we are not betting that the overall direction of AGI is going to end anytime soon. We just think there's more compute to scale. And if I truly want to understand Shawn and Shawn's work and Shawn's context, just re-reading files is not going to make it, especially for a special person like you. We got to train a 100 trillion parameter model for this guy.

“特别”是贬义的。

Special is derogatory.

我们得为这家伙训练一个 100 万亿参数的模型。

We got to train a 100 trillion parameter for this guy.

是的。Cosine Cosine。嗯，你发现人们最关心他们的模型学习什么？是记住关于组织的事实吗？是记住比如“啊，不，我们是这样做的 CI”吗？还是说人们实际上希望什么？然后这可能影响到你如何做记忆槽的排名等等。

Yep. Cosine cosine. Um, what are you finding that people care most about their models learning? Like is it memorizing facts about the organization? Is it remembering like, ah, no, we do CI this way? Is it like what are people actually hoping that? And then maybe this feeds into how you do the ranking of memory slots and all that.

是的，我认为如果你看看人们现在在应用层花时间做什么，很多都是试图让模型为你的用例工作得好。比如，哦，我希望模型用我的品牌风格设计我的网站。这是现在很常见的例子，但人们用智能体做很多不同类型的任务，比如学习如何运行工作流，或者比如你特定的写作方式。所以有很多种事情。老实说，我认为当我们思考这些方法时，回到事实和技能之间的区别，其实根本没有区别。我认为这些方法对此是无关的。

Yeah, well, I think if you look at what people are spending their time in the app layer doing these days, it's a lot of just trying to make the model work well for your use case. Like, oh, I want the model to like, let's say, design my website with my brand style. That's a very common example these days, but there are many different kinds of tasks that people do with agents, like learning how to run a workflow, or your particular way of writing, let's say. So there are many kinds of things. And honestly, I think when we think about these methods, going back to this distinction between facts and skills, there really is none. I think the methods are kind of agnostic to that.

是的，对我来说很自然：几乎所有的应用层基本上都是一个前沿模型，包裹在带有搜索工具等的循环中。他们对我们感兴趣的是找到一种方式，以更快、更高效且更具上下文的方式与他们的数据交互。所以几乎所有人都是这样：我们希望将公司知识编码到更高效的东西中，这样我就不用研究了。我们希望模型能有针对性地知道应该把某件事分派给谁。我们只是向他们展示，通过相当轻量级的训练，这些东西可以成为模型的本能。他们不需要那些非常复杂的长期仓库循环来解决这些问题。所以从某种意义上说，这有点像 RAG 杀手。再说一次，我们总是可以做 RAG，总是可以检索，但人们感兴趣的是：以这种方式与非常大的数据平面交互并自动化非常重复的事情。

Yeah, to me it's like the natural thing: almost all the app layers are basically a frontier model wrapped in a loop with search tools and stuff. And what they're all interested in doing with us is finding ways to interface with their data in a way that's faster, more efficient, and also more contextual. So almost all of them, it's like we want to have our firm knowledge encoded in something that's more efficient, so I don't have to research. We want to have the model know in a targeted way who's the person I should triage a thing to. And we're just showing them that with pretty lightweight training, these things can become instinctual to the models. They don't have to have these very involved long repo loops to solve them. So in a sense, it's like a RAG killer kind of thing. Again, we can always do RAG and we can always retrieve, but that's the thing that people are interested in: interfacing with very large data planes and automating very repetitive things this way.

是的。我想深入探讨一下这个 RAG 杀手的事情。抱歉老调重弹，但我还没有完全理解它。

Yeah. And I want to double click on this RAG killer thing. And I'm sorry to beat the dead horse, I just don't fully grok it yet.

嗯。

Yeah.

前提是使用 RAG 和更新模型权重之间存在某种权衡吗？是说你应该两者都做吗？比如哪些类型的事情应该放在权重中，哪些应该外部化到 RAG？

Is the premise that there's some trade-off between doing RAG versus updating your model weights? Is the idea that you should be doing both? Like what types of things should be done in the weights versus what types of things should be externalized to RAG?

我认为这是一个未解决的问题。我不认为有人有答案。我们都在研究它。这也是生物记忆的根本问题：什么应该内化，什么不应该。我确实认为，比如，你需要内化一年前住过的酒店的房间号吗？可能不需要，不在你的神经组织中。可能写下来就好。但你现在需要内化家里的密码吗？可能未来几年内把它印在某个地方是有用的。所以，是的，这如何转化为知识工作和产品？这仍然是我们正在摸索的事情，我们尝试采用尽可能少用启发式的方法。在数据上运行过滤器并说“我要保留这个，丢弃那个，训练这个，训练那个”更容易。但作为人类，我们看 TikTok，接触很多垃圾信息，大脑仍然能够学习，不会完全失控。我们认为模型也应该如此。

I think it's an unsolved problem. I don't think anyone has an answer to it. We're all working on it. It's also the fundamental question of biological memory: what should be internalized versus what not. I do think that things like, do you need to internalize the room number in a hotel that you were in a year ago? Probably not, not in your neural tissue. Probably that's good to write down. But do you need to internalize the password to your home right now? Probably it's useful for the next few years to have that imprinted somewhere. So, yeah, how does this translate into knowledge work and products? This is still something we figure out, and we try to take the approach that we use as few heuristics as possible. It's easier to run filters on the data and say like I'm going to keep this, discard that, train on this, train on that. But as humans, we watch TikTok and get exposed to a lot of garbage, and still the brain is able to learn and not completely go off the rails. And we think models should be the same as well.

是的。也许具体来说在短期内，我认为现在很多人担心的是连续运行这些智能体数天的巨大推理成本。

Yeah. Maybe concretely in the short term, I think a lot of what people are worried about these days is the huge inference costs of running these agents like for days on end.

高推理成本是好事。

High inference cost is a good thing.

我是说消耗 token 为了什么？

I mean consuming tokens for what?

Sonny 和 Fireworks 合作。他真的很喜欢 Zion。

Sonny works with Fireworks. He really loves Zion.

我喜欢推理。

I love inference.

我们也喜欢推理。

We love inference, too.

是的，所以我认为在短期内，那是直接的痛点。比如为什么你要一遍又一遍地读同样的文件，即使在同一个查询中，但更明显的是，同一公司不同的人，他们一遍又一遍地对同样的文档运行同样的查询。那应该是模型已经知道的事情。就像你问一个员工，他们不会在搜索框里输入“我昨天在做什么？”他们就是知道。

Yeah, so I think in the short term, that's the immediate pain point. Like why are you reading the same files over and over again, even in the same query, but definitely across people in the same company, they're running the same queries on the same documents over and over again. And that should be something the model just knows. In the same way you ask an employee, they don't type into the search box like, what was I working on yesterday? They just know.

但缓存不是能解决这个问题吗？

But doesn't caching kind of solve that?

我认为在某种程度上是的，但回到什么应该内化、什么应该在测试时检索的问题，我认为很多还是关于建立你的知识。所以如果你总是做 RAG，你就无法做出这样的关联：哦，我看到团队里有人在做这类研究，我在抽象层面上回忆起，哦，有件相关的事情你可能想知道。你甚至没有问过它，对吧？但我认为这种关联只能发生在权重中，因为它们并不是你让我搜索这个，我就去搜索这个。

I think to some extent, yeah, but I think going back to this question of what should be internalized versus what's something you retrieve at test time, I think again a lot of it is about building on your knowledge. So if you are always doing RAG, you can't make associations like, oh, I see somebody on the team is doing this kind of research, and I kind of recall at an abstract level, oh, there's this related thing that you might want to know about. You didn't even ask about it, right? But I think these kinds of associations can only happen in weights because they're not really about you asking me to search for this, I'm going to search for this.

另外，我认为检索系统（尤其是 AI 中的检索系统）的主要限制在于，问题不在于存储什么和放在哪里。问题在于如何寻址，如何查询。你甚至知道要找什么吗？是的。这涉及到某种直觉，有趣的是，有时模型并不具备这种直觉。它们不知道去哪里找。尤其是如果你局限于当前的做法，即关键词搜索，那么扩展强化学习更容易，而且在嵌入等基础设施方面涉及更少。所以，是的，知道搜索什么是一种直觉，可以发生在权重中。

Also, I think the main limitation with retrieval systems in general and in AI specifically is that the problem is not so much what to store and where to put it. The problem is how to address it, how to query the thing. Do you know what to look for even? Yeah. And this involves some sort of intuition that sometimes the models don't have, interestingly enough. They don't know where to look. And especially if you're limited to the current way of doing things, which is keyword search, it is just easier to scale RL and less involved in terms of infrastructure for embeddings and stuff. So, yeah, knowing what to search is something that's intuitive and can happen in the weights.

还有关于缓存和推理。我们公司很大程度上是从深入研究 KV 缓存开始的。这是一个很迷人的东西，对吧？KV 缓存是当前做法的一个庞然大物。想想看：一篇关于泰勒·斯威夫特的维基百科文章，它的 KV 缓存大约需要 80 GB 的 GPU HBM 内存。而整个 Llama 70B 模型，所有权重加起来大约 100 GB，而且某种程度上，它们记住了整个互联网。为什么一个东西可以如此高效？我们有证据表明，创建这种“气味”可以在极少的数字中打包大量信息。而 KV 缓存呢，你拿几十 KB 的文章，它就变成了 80 GB 的“大脑状态”。所以你可以缓存它，加载它，但会遇到从磁盘到 HBM 的问题。人们在研究这个，挺有意思的。但如果我们能拿这 80 GB，花一些离线算力，也许也在 Fireworks 上处理，然后压缩它，让它变得非常小，这样我们加载和缓存的东西就小了一千倍呢？那将对加载方式、速度以及表示的保真度产生巨大影响。

And also about caching and inference. Much of this company started with us taking a deep dive into KV caches and caching. This is a fascinating thing, right? KV cache is a monstrosity of the current way of doing things. Think about it: a KV cache for a single Wikipedia article about Taylor Swift or something like that would be about 80 GB of HBM memory on the GPU. An entire Llama 70B model, the entire weights of the model, would be about 100 GB, and with some distortion, they remember the entire internet. How come one thing is so bit-efficient? We have this proof of existence that creating this scent can pack a lot of information in very few numbers. Whereas this KV cache thing, you take a few tens of kilobytes of article and it becomes 80 GB of brain state. So you can cache this, you can load this, you'll have issues with disk to HBM stuff. People are working on it, it's pretty interesting. But what if we can take those 80 GB, spend some compute offline, maybe also on Fireworks and file, but then compress it and make it really small so that the thing we load and cache is a thousand times smaller? That would have tremendous implications for how we load things, how fast we can do things, and what the fidelity of the representation is.

非常有趣。

Super interesting.

是的。

Yeah.

未来一两年内，有哪些事情可能成为记忆领域的 ChatGPT 时刻？还是你认为事情不会那样发展？

What are some of the things that could happen in the next year or two that would be like the ChatGPT moment of memory? Or do you think that that's not how things will play out?

这是个好问题。我不知道。我认为人们一直在谈论的持续学习的第一个概念验证，就是你有一个实习生，你可以随时间教它东西，它真的会变好。我觉得大家都在等着看这个。不管现在的上下文工程技术有多复杂，它们都还没达到那个水平。所以我认为你需要所有这些工具才能实现。但我觉得会是类似的东西，模型真的变聪明了。比如，“哇，它和昨天不一样了。”

It's a good question. I don't know. I think the first proof of concept of the thing that people keep talking about with continual learning, which is you have an intern that you can teach things over time and it actually gets better. I think everybody's waiting to see that. No matter how sophisticated the context engineering approaches are these days, they're not getting there. So I think you need all of these tools at your disposal to make that happen. But I think it will be something like that where it's like the model's actually getting smarter. Like, "Whoa, it's different from yesterday."

是的。而且重要的是要说，ChatGPT 模型是未曾预料到的。我们之前读到的某些人的产品方向，和 ChatGPT 出现后的情况是不同的。

Yeah. And it's important to say that the ChatGPT model was not anticipated. We were just, you know, read about all the product directions that certain people had before ChatGPT was different.

这就是为什么我问你，不是为了……

That's why I asked you, not for...

对我来说，例子是这样的：“看，如果你今天辞职，你的唯一使命是做一个对你更好的模型。你会用 OpenAI、Anthropic 和所有前沿模型，然后 24/7 地工程化上下文技能，但作为个人，你能推动的进展非常有限。你最好等下一个版本的模型，然后从那里开始。我们希望看到一个未来，你花在事情上的时间越多，实际上能转化为性能质量的提升，至少在你关心的事情和领域上。”这很难实现，我们认为唯一可能实现的方式是开始规模化算力并在这些数据上训练，同时不破坏它们，这非常难。

I feel like to me the example is like, "Look, if you resign from your job today and your sole mission was to make a model that's better for you. And you would use OpenAI and Anthropic and all these frontier models and you would just 24/7 engineer the context right skills, your way to move the needle is very limited as an individual. You will just be better off waiting for the next version of the model and you will take it from there. And we would like to see a future where actually the more time you spend on the thing actually translates to the quality of performance, at least in the things and domains you care about." This is pretty hard to achieve, and the only reason we think it could be achieved is if you start scaling compute and training on these data without destroying them all importantly, which is pretty hard.

就几个，纯属好玩，快速问答，凭记忆回答。你上一次对 AI 的某件事感到惊讶是什么时候？任何领域。

Just a couple, just for fun, rapid-fire questions going off just memory. When's the last time you felt surprised about something in AI? In any area.

读到融资新闻的时候。每天都有很多惊喜。

When reading about fundraising. A lot of surprises every day.

我会说我们都感觉到了编码智能体能力的一些变化。

I would say all of us felt a little bit of a change around the capabilities of the coding agents.

确实。

That's true.

但我们之前一直在尝试这些东西，试图用更费力的方式让它们工作，所以并不完全意外。但对我来说，主要事件是 GitHub Copilot。那对我来说就是主要事件，然后是 ChatGPT，再然后是看到智能体式的东西。我想我们都预料到了，不同的人对它能走多远、能有多长的视野有不同的预期。但我觉得，我们还没看到根本不同的东西，人们现在正在研究全新的做事方式。但对我来说，模型实际上以无害的方式改变，并即时学习新东西，个人和经济上都可行，这很有趣。

But we've been dabbling with these things and trying to make them work in more effortful ways before, so it didn't come as a complete surprise. But yeah, I think to me the main events were GitHub Copilot. That for me was just the main event and ChatGPT and then seeing the agentic stuff. We all anticipated, I think, and different people had different expectations on how far it can go and how long horizon it can go. But I feel yeah, we're yet to see something fundamentally different and people are working on completely new ways of doing things now. But yeah, to me it's models actually changing in a way that's not harmful and learning new things on the fly, personally and economically viable. That's interesting.

现在有一种想法，我们每个人都可以有一个代币钱包，随身带到公司、不同的应用、不同的工作空间。你认为我们最终会有一个记忆银行、记忆钱包，随着我们在数字世界中移动而随身携带吗？

Right now there's this idea of like we each can have a token wallet that we're going to bring around to companies. Or to different apps. Different workspaces. Do you think that we're going to end up with like a memory bank, a memory wallet that we're going to move around across the digital world as we go?

我觉得这是个有趣的问题。我不知道我们是否已经完全想清楚了这个意义上正确的产品形态。在某种程度上，即使有 ChatGPT 的记忆，我也不太希望它跨我的个人和工作上下文记忆。

I think it's an interesting question. I don't know if we've fully figured out what the right kind of product form factor is in this sense. In a way, even with ChatGPT memory, let's say, I kind of don't want it to remember across my personal and work context.

就像，“哦，你可能喜欢这些床单，因为你上周在 GPU 上训练了一个模型。”这完全无关。在某种程度上，因为记忆有缺陷，但我也认为你确实希望工具和产品中的记忆是分离的，你能控制它。所以我个人认为需要某种分离，但具体是什么样子还有待确定。

Like it's like, "Oh, you know, you might like these sheets because you trained a model on a GPU last week." It's like that's totally irrelevant. And to some extent, it's like because the memory is flawed, but also I think you do want memory in your tools and the products that you use to be separated, to have control over that. So I personally think there needs to be some separation there, but I guess to be determined what that might look like.

是的。我认为一个圣杯是，你去上班，消耗所有代币，创造所有价值，然后所有知识产权等东西留在公司，但你学到的技能、发明的东西、做事的方式，其中一些你可以以经过 sanitize 且不损害其他公司知识产权的方式带到下一份工作。所以我确实认为携带一套技能会很有趣。我们现在在生物学上就是这样做的，我们签 NDA，有道德规则，但我认为在数字世界中这样做会非常有趣且回报丰厚，因为它会迫使每个人推动前沿，更深入地在公司和个人生活中实施 AI，然后获得回报。

Yeah. And I think a holy grail is like you go to work and you just burn through all these tokens and you create all this value and somehow, all the IP and stuff stays with the company, but somehow the skills you learned, the things you invented, your ways of doing things, some of them you can take with you as well to your next job in a way that's sanitized and not harmful to any other company's IP. So I do think carrying a set of skills will be interesting. We do it in our biology right now and we just sign NDAs and have ethical rules around it, but I think doing it in the digital world would be pretty interesting and pretty rewarding because it will force each of us to push the frontier and implement AI more deeply in our companies, in our individual life, and then be rewarded for it.

我 2007 年在斯坦福统计公司开始读博，当时 AI 无聊透顶。全是统计学习。基本上只有两个领域：计算机视觉和 NLP。所以视觉和语言是两大领域，我认为现在仍然如此。2012 年，AlexNet 出现了。视觉主导了大概 6 年。

I started a PhD in the stats firm in 2007 at Stanford and AI was boring as hell at the time. It was all statistical learning. And there's basically two areas: computer vision and NLP. So vision and language were kind of the two areas and I think that's still true. In 2012, AlexNet happened. Vision was dominating for 6 years or whatever.

你们是否对语言似乎在进展中主导了视觉感到惊讶？你认为视觉还有机会回归吗？你怎么看？

Are you guys surprised that language seems to be dominating over vision in progress? Do you think vision has any chance of coming back? How do you think about this?

是的，我觉得这相当令人惊讶。有些人可能预见到了，但我一直对语言作为沟通媒介感兴趣，语言中能完成许多复杂的抽象事物。我认为长期来看，语言和视觉会结合成一个更统一的系统，我们从所有不同模态获取输入，并以抽象方式理解它们。

Yeah, I think it is pretty surprising to me. I mean, some people maybe saw it coming, but I think I've always been interested in language as a medium for communication, and so many complex abstract things can be done in language. I do think in the longer term, language and vision will combine into a more unified system where we take inputs from all these different modalities and understand them in an abstract way.

对我来说，我从未对语言感兴趣。它看起来是一种非常高级的能力，整个动物王国都有非常不同的语言形式，与我们用文字自我交流的方式不同。我和许多 AI 领袖一样，一直认为自然的方式是体验世界、在其中行动，视觉和行动才是关键。但后来我看到了 ChatGPT 时刻，并在 Mosaic 做了一些工作，了解了 NLP 方面的内幕。令人震惊的是，语言应该很难。每个词都有一个独热嵌入向量，与其他词完全不同，是一个完全高维的空间，在某种意义上非常人工。我们用比最好的视觉模型大几个数量级的模型来学习它，但效果仍然很好。我确实认为图像和视频还有很多潜力可挖，你们在这个领域的投资是好的，但两者会以不同方式保持有趣。

Yeah, to me, I've never been interested in language. It seemed such an advanced capability that the entire animal kingdom has very different forms of speech and language than how we communicate with ourselves and in writing. I was always, as many other leaders in AI, had this thought that the natural thing is you have to experience the world, act in it, and vision and action will be the key. But then I saw the ChatGPT moment and went to do some work at Mosaic to learn how the sausage is made on the NLP side. The striking thing is that language should be pretty hard. Each word has a one-hot embedding vector that's completely dissimilar to any other word, a completely high-dimensional space, and it's really artificial in a sense. We learn it with models that are orders of magnitude bigger than the best vision models, and still things work pretty well. I do think there's a lot of juice to be squeezed in image and video, and I think you guys are making good investments in this space, but the two will keep being interesting in different ways.

我要告诉你我的疯狂理论。我很震惊语言超越了视觉，我低估了 2018、2019、2020 年 LLM 的发展，因为我对视觉有偏见。现在回想起来，我认为基本上在生物学中，视觉相对于语言有巨大的根本优势。大脑通过眼睛处理光学数据的比特率高出几个数量级，而且在到达电子之前有很多光学处理。因此，处理并到达大脑的训练数据总比特率似乎比音频数据高出几个数量级，因为声波本质上比光慢得多。然后音频从声学到电子有一个上采样，而视觉从光子到电子有一个下采样。在今天的计算机中，一切都是电子的，所以你削弱了视觉，提升了语言，所有处理都在同一水平上。我认为这可能部分解释了为什么从信息论角度看，语言和视觉在 LLM 层面处于相似水平，而 LLM 是一种非常智能的架构，更适合语言而非视觉。这听起来有多蠢，尤其是对你这位神经科学家来说？

I'm going to tell you my crackpot theory. I was pretty shocked that language surpassed vision, and I underestimated what was happening with LLMs in 2018, 2019, 2020 because I had this bias towards vision. When I look back, I think what's basically happening is that in biology, vision has a massive fundamental advantage over language. The bit rate that your brain can process optical data through the eye is many orders of magnitude greater, and there's a lot of optical processing before it reaches electrons. So the total bit rate of training data processed and making it to your brain seems many orders of magnitude greater than audio data, where sound waves are fundamentally much slower bit rate than light. Then there's an upscaling from acoustics to electronics in audio, whereas in vision there's a downscaling from photons to electrons. In computers today, everything is electronic, so you nerfed vision and promoted language, where all processing is on the same playing field. I think this might be part of why, from an information theory perspective, language and vision are on a similar playing field by the time you get to LLMs, and LLMs are a really smart architecture better suited for language than for vision. How dumb does this sound, especially to you, Dan, the neuroscientist?

Jesse 也有认知计算科学的背景，对吧？所以我的观点是，我们在知识工作中所做的很多事情，我们并没有进化去做。我们坐在电脑前写备忘录等等，我们并非为此进化而来。我们的大脑并不擅长这个。尽管如此，让 LLM 为我们做这些事还是有用的。作为人类，我们严重偏向视觉；其他啮齿动物更偏向嗅觉。我自己研究过这些。大脑中分配给视觉（枕叶）和语言区域（颞叶）的“地产”哪个更多？可能是视觉更多。我得用 ChatGPT 查一下，但我觉得情况就是这样。

And Jesse also has some background in cognitive computational science, right? So, I would say my point here is, look, much of what we're doing in knowledge work, we haven't evolved to do. We're sitting on these computers, writing memos, etc. We are not evolved to do this. Our brains are not wired for it. Still, it's useful to have LLMs to do this for us. As humans, we're heavily vision-biased; other rodents are more olfactory biased. I've worked on these things myself. What's the real estate in the brain allocated to vision, occipital lobes, versus language areas, temporal lobe? Probably more vision. I'll have to check with ChatGPT, but I think that's the situation.

你凭记忆知道吗？

Do you know from memory?

不，老兄，我在外化。我个人生活中是 RAG 的坚定信徒。但我认为……

No, man, I'm externalizing. I'm a big RAG believer in my personal lifestyle. But I think...

在极限情况下，我们都是 RAG。

In the limit, we're all RAG.

我只内化重要的事情，比如我对你的情感。开个玩笑。抱歉。总之，视觉占主导。当人们训练视觉语言模型时，语言最终主导了视觉内容。但很难说因为某个大脑更偏向某种模态，就意味着我们会更高效地处理它。我确实认为脑机接口的努力应该考虑到这一点。如何将其传回大脑？这才是真正重要的，要考虑我们那里有什么“地产”。但对于知识工作，我认为文本也同样适用。

I internalize just important things like my emotions to you. I was just kidding. Sorry. Anyways, vision is dominant. When people train vision-language models, language ends up dominating the vision content there. But it's hard to say that because a certain brain is more biased towards a certain modality, it doesn't mean we're going to more efficiently do it. I do think efforts on brain-computer interfaces should take this into account. How do you relay it back to the brain? That's where it's really important to think about what real estate we have there. But for knowledge work, it's equally fine if it's text, I think.

最后一个问题。如果一切顺利，5 到 10 年后的世界会是什么样子？Ngrams 在其中扮演什么角色？

Last question. If everything goes right, what does the world look like in 5, 10 years? And what is Ngrams role in it?

我想象一个世界，每个人都有自己的模型，这个模型与别人的模型和前沿模型截然不同。所有这些模型服务于不同目的。拥有一个真正了解你、并以对你个人有意义的方式帮助你的模型，无论是个人还是团队。我认为存在一种元素，即到处都有不同种类的智能。

I think I'm imagining a world where everyone has their own model that is really different from the other person's model and from the frontier model. All of these serve different purposes. To have a model that really knows you, but also helps you in ways that make sense to you personally, whether it's an individual or a team. I think there's an element of having different kinds of intelligence everywhere.

是的，对我来说，这是故事的一个变体：在神经科学中，我们知道记忆和导航密切相关。

Yeah, and to me, it's a variant of the story where in neuroscience, we know that memory and navigation are pretty closely related.

大脑中负责表征空间地标的同一回路，也掌管着情景记忆的某些元素。对我来说，我认为这家公司可以成为每个人通往数据平面的实际 LLM 接口。因此，它与 Databricks 和 Oracle 这样的伟大公司有一些相似之处——我们形成这些恰好是神经记忆的记忆，这些记忆由恰好是个性化的模型承载，而且有数亿个这样的模型，但它们本质上是一种数据平面的神经接口，其方式与我们已知的截然不同，更高效、更联想化。它不是按原样表征文件系统，而是表征该文件系统的大脑状态。这就是我的愿景。

Same circuits in the brain that represent landmarks in space are in charge of some elements of episodic memory and things like this. For me, I think the company can be the actual LLM interface to the data plane for everyone. So, sharing some similarities to great companies like Databricks and Oracle, where we form these memories that happen to be neural memories with models that happen to be personalized, and there are hundreds of millions of them, but they're basically a neural interface to the data plane in a way that's very different from what we know, and it's more efficient, it's more associative. It's not representing the file system as it is, it's representing a brain state of that file system. So, that's for me a vision.

以这个美好的愿景作为结尾。非常感谢你们过来分享正在构建的东西。

Beautiful vision to end on. Thank you guys so much for coming by to share what you're building.

太棒了。很喜欢。谢谢大家。

Awesome. Love it. Thank you guys.