为什么我们需要持续学习

这个模型基本上是冻结的，但新的经验和知识依然能保留下来。

The model is basically frozen, but the new experiences, new knowledge still persists.

人类不是通用人工智能，但我们边干边学。

Humans are not AGI, but we still learn on the job.

我们从经验中学习，这才是人类独特的地方。

We learn from experience, and that's what makes humans unique.

没错。

Yeah.

所以这才是终极考验。

And so that's the ultimate test.

我们怎么定义‘持续学习’真的实现了呢？

How do we define that we got to continual learning?

有没有哪种系统，能像人一样，在实践中不断学习、不断进步？

Is there a system that is able to learn on the job and get better through use, just like humans?

我们聊过的所有实验室，连他们自己都不会只盯着一种方法，而是有多个团队，用不同的方式探索持续学习。

In all of the labs we talk to, even the labs don't just tackle one approach, they actually have multiple teams that tackle continual learning through the different kind of paradigms.

任何关于持续学习的真诚讨论，都得从上下文学习说起，因为这玩意儿确实管用。

Any honest argument about continual learning pretty much has to start with in context learning because it genuinely works.

要是现在的AI模型根本学不会呢？

What if today's AI models can't actually learn?

现在大多数系统都是训一次，上线完就定型了。

Right now, most systems are trained once, deployed, and then frozen in time.

它们能推理、能检索、能生成，但就是没法真正从经验里更新自己。

They can reason, retrieve, and generate, generate, but they don't truly update from experience.

为了解决这问题，我们给它们搭了层层外挂：上下文窗口、检索系统、智能体框架。

To compensate, we've built layers around them: context windows, retrieval systems, agent scaffolding.

这些办法确实有效，但也引出一个更深层的问题：我们是在绕开局限，还是已经撞上了这种架构的天花板？

These approaches work, but they also raise a deeper question: are we just working around a limitation, or have we reached the ceiling of what this paradigm can do?

还有一条路，就是让模型不只是回应，还能进步——能持续学习、适应新信息，像人一样慢慢进化。

There's another path, one where models don't just respond but improve, where they learn continuously, adapt to new information, and evolve over time, more like humans do.

本期节目中，埃琳娜·伯格采访了a16z人工智能基础设施团队的合伙人玛莉卡·奥巴基罗娃，聊聊为什么持续学习这么重要，眼下缺了什么，以及要怎么才能造出真正能从经验中学习的系统。

In this episode, Elena Burger speaks with Malika Aubakirova, partner on the AI infrastructure team at a16z, about why continual learning matters, what's missing today, and what it would take to build systems that actually learn from experience.

大家下午好，现在是太平洋时间下午2点33分，正在关注这个情况的各位。

Good afternoon, everyone, who is currently monitoring the situation at 02:33PM Pacific time.

我是Elena。

I'm Elena.

我负责新媒体团队，今天和Malika一起在这里。

I work on the new media team, and I'm here with Malika.

Malika，你来简单介绍一下自己吧？

Malika, do you wanna introduce yourself?

是的。

Yeah.

非常感谢你们邀请我来参加。

Thank you so much for having me here.

我是AI基础设施团队的合伙人。

I'm a partner on AI infrastructure team.

很高兴能继续聊聊持续学习这个话题。

Excited to chat more about continual learning.

对。

Yes.

今天马利卡发表了一篇题为《为什么我们需要持续学习》的文章。

So today Malika published a piece called Why We Need Continual Learning.

我觉得在深入这篇文章之前，马利卡，你先聊聊写作的过程吧，因为你简直把全宇宙的AI研究员都聊了个遍。

And I think even before we go into this piece, I think like first Malika, you should just talk about your process writing it because, you know, it seems like you spoke to every single AI researcher under the sun.

特别想听听你是怎么开展这个过程的，然后再一起深入中间的内容。

So would love to just kind of like hear what your process was and then we can kind of get into the middle of

过程本身。

the process itself.

当然。

Absolutely.

其实我们没把所有参与的人列出来，就是因为有机会、也有条件跟一大批顶尖的研究员、创始人和博士生聊过。

Like, in fact, that's actually the reason why we didn't name all of the individuals involved because we had the opportunity and luxury to talk to a number of just incredible top researchers, founders, PhD students.

我们还办了持续学习主题的晚餐会。

We organized continual learning dinners.

说实话，这篇文章的成型很大程度上得益于他们的见解和经验，让内容比我们自己埋头写出来的要深刻得多、扎实得多。

And so honestly, this piece was shaped largely by their insights and learnings and made this piece much more sharper and grounded than anything else we could have written on our own.

真的非常感谢大家提供这么精彩的洞见。

So definitely thank you all for just incredible insights.

是的。

Yeah.

文章一开始呢，

And so the piece opens.

首先，大家都能看到这里引用了克里斯托弗·诺兰那部经典电影《记忆碎片》。

So first, everybody can now see this reference to the great Christopher Nolan movie Memento.

文章一开始，

The piece opens

对机器做了一个巧妙的反转。

With a twist on machines.

对机器做了一个巧妙的反转。

With a twist on machines.

那我们为什么用这个隐喻来开篇呢？

So why did we open the piece with this metaphor?

你知道吗，现在的AI模型，那种像《记忆碎片》一样的体验到底是什么？我们这里用的框架又是啥？

You know, what is the kind of like memento like experience of you know, AI models today and kind of like what what is the frame that that we used here?

是的。

Yeah.

没错。

Absolutely.

我不确定在座的各位有没有看过《记忆碎片》这部电影。

So I guess I'm not sure if everyone in the audience has seen Memento movie.

如果没看过，我简单说一下。

If you haven't, here is kind of like a quick blurb.

片子里的主角，莱昂纳多·迪卡普里奥，患有一种失忆症，没法形成新的记忆。

But, basically, the main protagonist, Leonardo Chelby, has a form of this amnesia where he cannot form new memories.

所以他活着就像卡在一个时间点上，之前的事还记得，但之后发生的一切，全记不住。

So he goes about his life with kind of like this cutoff date after which point he has kind of these long term memories, but really cannot retain anything new that he experiences.

所以他用便利贴给自己写提醒。

And so what he does is he uses the sticky notes where he writes some of the notes to himself.

他随身带着拍立得相机，抓拍生活中的各种瞬间。

He pulls out his Polaroid camera to capture the moments as he goes on about his life.

甚至在整部电影里，他还把一些重要的记忆纹在了身上，好让自己别忘掉。

And, I mean, he even tattoos some of the memories that he wants to imprint in his memory basically throughout kind of the movie.

那这到底有什么意义呢？

And so why why why does why why does this matter?

对吧？

Right?

这听起来像是个电影情节，但其实它和现在的AI模型运作方式一模一样。

Like, this is just kind of like an explanation, but why does this matter in like AI models, like, honestly, it kind of maps one to one to AI how AI models work today.

我们先用海量数据训练模型，把整个世界的知识都装进去，这叫预训练。

So we have the training phase where we basically encompass all of the world knowledge, and that that part is what we call pre training.

训练完之后，就到了一个分水岭，模型被部署到真实世界里使用，这叫推理。

And after the training phase, we basically have the cutoff date after which point we deploy the models into the world, and we call that inference.

所以问题来了，模型基本上是冻结的，但新的经验和知识依然在持续积累。

And so the question is, honestly, the model is basically frozen and but the new experiences, new knowledge still persists.

那我们该怎么为它搭建支撑体系呢？

So how do we go about scaffolding it?

现在我们通常用智能代理来增强它。

Today, we augment it with agent harnesses.

我们会用检索机制，比如知识库。

We use retrieval mechanisms like racks.

我还常举一个例子：系统提示词其实就像纹身一样。

We also even like the example that I like to give is we have the system prompt that essentially serves as a tattoo.

它其实就是我们围绕模型搭建的一整套支撑系统，帮助它通过反馈和使用不断学习。

It's basically kind of like all of the scaffolding that we build around the models to really learn through feedback and through use.

所以这篇内容其实是个好机会，让我们深入探讨一下，为什么持续学习这么重要。

So this post is honestly kind of like an opportunity to dive deeper into the topic, really discuss why continual learning matters.

对。

Yeah.

我们现在所处的，正是这种上下文学习的模式。

So we're living right now in this paradigm of in context learning.

这篇文章的核心观点就是：咱们得转向持续学习的模式了。

And basically the thrust of this piece is like, okay, we need to move into kind of like a continual learning paradigm.

你提到好几家公司在做这方面的尝试。

But you do you call out like a couple of different companies.

你提到了 Cursor。

You call out Cursor.

你还提到了 OpenClaw。

You call it OpenClaw.

这些团队都在做上下文学习相关的事，效果看起来挺不错的。

These teams that are kind of doing in context learning things, things in that modality that seem to be pretty effective.

你觉得他们现在还差在哪儿？

So where do you feel like they still fall short?

咱们在哪些地方还得继续改进？

And where do you feel like we still need improvements?

任何关于持续学习的真诚讨论，都得从上下文学习说起，因为它确实管用。

Any honest argument about continual learning pretty much has to start with in context learning because it genuinely works.

我们看到像Carpathi的自动研究项目这样的例子，还有文章里提到的OpenCLO，都是这样。

We see that with examples like Carpathi's Auto Research Project, kind of like the other examples we give in the article is OpenCLO.

底层模型谁都能用，但真正让它变得神奇的，是上下文的巧妙组织，对吧？

Like the underlying model was available to anyone, but what's really made it a special magical moment is this kind of like orchestration of the context, right?

谁都看得见上下文，但OpenCLR真正厉害的地方，是它能充分利用你的文件系统。

Everyone has the access to the context, but here OpenCLR really utilizes your file system.

它能帮你建立记忆。

It creates kind of memories.

对吧？

Right?

它甚至还能直接调用bash命令。

And it even has like a special bash access.

说实话，那些怀疑的人可能会问：干嘛搞得这么复杂？

So honestly, like the skeptics would say like, why complicate things?

对吧？

Right?

这些看起来很粗糙、不精致的界面，反而特别管用，而且会一直赢，因为它们太基础了。

Like these naive, but janky interfaces really tend to work and they will continue to win because they're so fundamental.

所以我和马特花了很多时间琢磨这个事儿。

And so that's where we spend a lot of time thinking about it with Matt.

我特别喜欢尤桑举的一个例子，他现在在斯坦福。

And the example that I really like is given by Yousan, who is currently at Stanford.

这是一个数学领域的例子。

It's an example from mathematics.

你可能听说过费马大定理，这个问题三百五十年来一直没人能解决。

You probably have heard about Fairmont's last theorem that was pretty much an open problem for over three fifty years.

其实有很多数学家都尝试过攻克它。

And the thing is, like, were a lot of mathematicians that tried to tackle it.

所有的文献、研究论文，谁都能看到。

All of the literature, researchers, like research papers were available to anyone.

但安德鲁·怀尔斯的做法是，他几乎闭关七年，自己发明了一套新方法，把两个数学分支——椭圆曲线和模形式——连了起来。

But what Andrew Wiles did, he basically went into near isolation for seven years and had to invent new techniques to bridge basically two fields of branches of mathematics, elliptic curves and modular forms.

你可别问我细节了，这超出了我的能力范围。

And please don't ask me about that because, that would be above my pay grade.

对。

Yeah.

咱们这儿就不深入讲密码学了。

We're not gonna get into cryptography here.

别担心。

Don't worry.

但关键在于，这真是一次纯粹的发现，哪怕把人类之前所有已知的信息、预训练数据全翻个底朝天，也根本想不到这个。

But basically the learning here, it's like a true genuine discovery that you really could not have even learned from all of the information and like whatever pre training data that humans kind of had before.

对。

Yeah.

没错。

Yeah.

所以，为了更具体一点，我想举几个实际的例子。

So I guess just also to ground it a little bit more, I want to give a few concrete practical examples.

第一个例子跟对抗性安全有关。

And the first one is essentially in adversarial security.

想象一下，突然出现了一种新的越狱攻击。

Like imagine there is a new jailbreak attack.

你的模型已经上线运行，正在被用户使用。

You have your model deployed in the wild and it's being used.

你试着更新系统提示，告诉它‘别这么做’。

Imagine you try to update your system prompt to say like, don't do this.

但根本没用。

Like it's not going to work.

对吧？

Right?

因为模型的所有参数都已经学会要尽量满足用户的需求。

Because all of the parameters in the model have learned to be helpful to the users.

所以你得把这种知识藏在权重里，让攻击者根本碰不到。

So you really have to encompass that kind of knowledge in the weights where the attackers don't have access to.

对吧？

Right?

攻击者能接触到的上下文，跟你普通用户看到的一模一样。

The attackers have access to your context just like any other user.

所以你得靠权重这种别的东西来真正解决这个问题。

And so you have to use something else like the weights to really tackle it.

第二个例子，想想你最爱的 JavaScript 库，比如 React，对吧？

And the second example is like imagine your favorite JavaScript library, like let's say React, right?

你通过预训练数据早就熟悉了一个叫 x 的函数。

You learn through all of your pre training data that there is a function called x.

但某天 React 更新了，结果发现这是个破坏性变更，x 函数突然就没了。

But at some point, a new version of React comes out and turns out that it's a breaking change and all of a sudden x function doesn't exist.

现在改成了 y 函数。

It's now a y function.

不管你再怎么在上下文里强调，都没法直接盖过模型参数里最自然、最根深蒂固的那套东西——比如它就是认准了 x。

No matter how much you say it in the context, you cannot just override what's the most intuitive throughout all of the model parameters to basically say x.

这些就是大概的例子。

And so those are kind of like the examples.

所以问题根本不是上下文学习有没有用。

And so the question is not whether in context learning works.

问题是，这会不会就是它的极限了。

The question is whether that's kind of the ceiling.

有意思。

Interesting.

你在这篇东西里做了一件特别棒的事，就是把学习到底发生在哪儿——或者说你管它叫‘压缩’——给说清楚了，能再解释下什么叫‘压缩’吗？

So you kind of you do something really, really helpful in this piece, which is you lay out where exactly learning happens or or you call it actually compaction in in the piece, and and maybe you can explain what compaction is.

不过我觉得现在最好先顺着这张图聊一聊，顺便也说说现在有哪些团队在做这事儿，他们从非参数到参数这一整条线上，各自都是怎么搞的。

But I think maybe a good thing to do now would be to kind of like talk through this chart and then also talk through just, you know, who are some of the teams that are really building in the space and what their different approaches are across the entire, you know, like, nonparametric to parametric spectrum.

没错。

Totally.

对。

Yeah.

没错。

Yeah.

我觉得问题不在于我们有没有记忆功能，而在于压缩到底发生在哪个环节。

I guess like the question is whether it's not that we have the memory features or not, it's more about where does the compaction happen to a point.

所以我们提出了一个高层次的框架，把整个机制分成三块：上下文、模块和权重。

So we make this like very high level framework in terms of just the three buckets of the context, the modules, and also the weights.

这里特别想强调的是，这三者本质上都是学习机制。

And the distinction that like, the one call out that I think is important is all of these are learning method mechanisms.

就连上下文学习，其实也属于持续学习的一种。

And even in context learning, it's still a form of continual learning.

只是我们在文章里没把重点放在这上面。

It's just not where we focus the most energy in the article.

但上下文本质上就是我们说的非参数学习——权重根本不会更新。

But context is essentially what we call nonparametric learning where we don't actually update the weights.

而这就是现在大家最熟悉的方式。

And that's what people are most familiar today with.

所以你想到的那些公司，比如Pinecone，还有那些做智能体框架、记忆结构的，像Leta、Mzero，都是这类。

So that's exactly what you think about Rack, companies like Pinecone, companies that built agent harnesses, memory scaffolding, like Leta, Mzero.

这方面的活动确实挺多的。

And there's definitely a activity lot here.

目前这个领域里，大家思考问题的核心关键，其实就在于限制。

The main defining factor in how things are currently thought about in this specific field is the limitation.

这个限制就是，你的上下文长度是有限的。

And the limitation here is like you have a limited amount of context length.

所以怎么最高效地利用这一点，就成了核心问题。

And so how do you utilize it in the most efficient way is kind of the main question here.

然后还有一种折中的情况，就是你希望一定程度上能访问权重。

And then we have kind of like the middle ground of you want some access to weights.

你想更新它，但有没有一种折中办法，能以更灵活、更小规模的方式学习和更新权重，而不必改动整个模型？

You want to update it, but is there kind of a middle ground where you can learn and update your weights in some adaptable, smaller manner where you don't update the entire model?

所以这算是中间路线。

And so this is kind of the middle ground.

斯坦福有一篇很棒的论文叫《Cartridges》，讲的就是怎么更新KBCache。

And there is a great Stanford paper on this called Cartridges that kind of explains how you can update KBCaches.

这篇文章主要花精力讨论的还是权重和参数化学习这部分。

And again, like the energy where we spend in the article for the most part is on the weights and kind of like the parametric learning.

这个领域说实话还处于早期阶段。

That field is honestly still kind of in early states.

现在有不少公司都在做强化学习、数据和系统这块，也有些公司直接说Transformer架构才是瓶颈，得搞新架构。

So we have a bunch of companies that are doing kind of the RL, data and systems, and some that are basically saying that transformer architectures are the bottlenecks and we need the novel architectures.

我们列了几个例子，但重点是，我们接触的所有实验室，其实都没只盯着一种方法。

And so we kind of list a few, but it's just worth calling out that like in all of the labs that we talk to, even the labs don't just tackle one approach.

他们通常有好几个团队，分别用不同的方式研究持续学习。

They actually have multiple teams that tackle continual learning through the different kind of paradigms.

既然咱们还剩五分钟，不如聊聊你在这篇文章最后得出的那些结论？

So I think given that we have about five minutes left here, I think it might be good to kind of chat through some of the conclusions that you come to at the end of this piece.

所以你其实是觉得，我们甚至可能需要重新定义一下‘模型’到底是什么。

So you sort of suggest that we might even need to redefine what a model even is.

我很想听听你亲自怎么说，这到底意味着什么，我们该怎么理解这件事？

So would love to kind of hear in your own words, you know, like what does that mean and how should we think about that?

完全同意。

Totally.

现在的模型在训练截止日期后就固定了，就像我们之前说的，但我最近又想到 Ilya 说过的话。

I mean, the models are currently frozen after the cutoff date, like what we mentioned before, but I think I come back to just what Ilya talked about just recently.

他说，我们在 AGI 上其实有点跑过头了。

And what he said was basically like, with AGI, we almost overshot the target.

我们人类根本不是 AGI，但我们在工作中一直在学。

We humans are not AGI, but we still learn on the job.

我们靠经验学习，这才是人类真正独特的地方。

We learn from experience and that's what makes kind of humans kind of unique.

没错。

Yeah.

所以这简直就是终极考验。

And so that's kind of like the ultimate test.

我们该怎么定义持续学习呢？有没有一个系统能像人一样，在实践中边做边学、越用越强？

Like how do we define that, that we got to continual learning is like, well, is there a system that is able to learn on the job and get better through use just like humans?

我觉得这大概就是核心问题了。

I think that would be kind of the question.

顺便说一下，伯克利的一些研究者，还有其他实验室的人，正在开发一些评估标准，希望能更清晰地界定什么是持续学习。

And by the way, there are researchers from Berkeley and some of the other labs that are actually working on benchmarks that will hopefully help us define what is continual learning in a better form.

你刚才其实已经提到这一点了，就是说模型开始有那么一点点‘经验’的苗头了。

And you also, I mean, you kind of get at the you just got at this already, but you talk about, you know, models getting to this place of having, you know, glimmers of experience.

对你来说，有没有一个具体的里程碑，能让你觉得：哎，这次真不一样了，AI已经远远超越了现在的水平？

And like, is there some kind of concrete milestone for you that would tell you like, you know, this we have, you know, like seen something that is just so far beyond what current AI models are capable of?

也许是一个全新的发现，也可能只是它能持续工作更久，还不出错。

Maybe it's, you know, a novel discovery or maybe it's just, you know, the ability to work for longer and longer and longer amounts of time without, you know, messing up in some way.

所以我很想知道，你心里有没有什么判断标准？

Like, so would be curious to kind of hear your kind of heuristic is there.

我刚才简单提了一下，就是实际工作中的学习，也就是部署后通过经验不断进步的能力。

Something that I briefly mentioned just now is honestly kind of on the job learning, learning from experience when you are deployed, whether you can get better.

现在有些人用的测试方法其实挺简单的。

And the test that some people use currently is pretty simple.

你先用 XYZ 数据训练一个模型。

You basically have you train a model that is learned on XYZ data.

部署之后，就看看它能不能从没见过的新数据里学到东西。

And once you deploy, you just want to check whether it learns something out of distribution, something that it hasn't seen before.

我们现在已经开始看到一些例子了，比如余三做的测试时训练，还有那篇《发现》论文，都算得上是些新颖的突破。

And we are starting to see some examples like the test time training done by Yu San, with the discover paper that kind of makes some of the novel inventions.

这真的改变了模型的形态，让它们能边干活边适应、边学习，专门解决具体问题。

It really changes kind of like the shape of the models to adapt and learn on the job to tackle specific problems.

这篇就是那篇论文，对吧？

This is the paper, yeah?

对。

Yes.

没错。

Exactly.

这次试着去发现它。

Learn to discover it this time.

对。

Yes.

好的。

Alright.

在家看的朋友们，都去读一读这篇论文吧。

Homework homework for everyone who's who's watching at home is is to read this paper.

明白吧？

You know?

Malika，你有什么想说的收尾话吗？

So do you have do you have any, you know, closing words, Malika?

你觉得现在谁应该来找你，进一步聊聊这个话题？

Do you, you know, like, should who should be reaching out to you right now to talk more about this?

你希望听到哪些人的声音？

What kinds of people do you wanna hear from?

任何对持续学习感兴趣的人，尤其是创业者，我们都很欢迎。

Anyone who is interested in continual learning, especially founders, we

特别想跟你聊聊。

would love to just chat with you.

不然我们就会困在自己的记忆里出不来。

Otherwise, we will be stuck in our own memento.

困在当下的循环里。

In our own perpetual present.

是的。

Yeah.

没错。

Exactly.

我们可不想接下来几年光顾着写便利贴、在身上纹字，或者拍一堆录像带——就像《记忆碎片》里那些疯狂的事。

We do not want to spend the next couple years of our lives writing on sticky notes, tattooing ourselves, you know, making video tapes, all of the crazy things that happen in Memento.

Malika，真的太感谢你来参加我们的节目了。

Well, thank you so much, Malika, for hopping on the show.

今天聊得特别开心。

This has been a ton of fun.

希望咱们以后能多写几篇，你得经常来啊。

I hope we got to write a ton more pieces and, you know, like, you got to come on all the time.

今天真的太愉快了。

This was a ton of fun.

继续往前走吧。

Well, keep forward.

谢谢你。

Thank you.

再次感谢你。

Thank you.

感谢你收听这期的a16z播客。

Thanks for listening to this episode of the a 16 z podcast.

如果你喜欢这期节目，别忘了点赞、评论、订阅，给个评分或写个评价，再分享给你的朋友和家人。

If you like this episode, be sure to like, comment, subscribe, leave us a rating or a review, and share it with your friends and family.

想听更多节目，去YouTube、Apple Podcast或者Spotify找我们吧。

For more episodes, go to YouTube, Apple Podcast, and Spotify.

在X上关注我们@a16z，再订阅我们的Substack：a16z.substack.com。

Follow us on x at a sixteen z, and subscribe to our Substack at a16z.substack.com.

再次感谢收听，咱们下期节目见。

Thanks again for listening, and I'll see you in the next episode.

本内容仅作教育用途，不构成任何投资或金融产品的买卖或持有建议。

This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product.

本播客由第三方制作，可能包含付费推广广告、其他公司提及以及与a16z无关的个人内容。

This podcast has been produced by a third party and may include paid promotional advertisements, other company references, and individuals unaffiliated with a16z.

这些广告、公司和个人均不代表Ah Capital Management LLC、a16z或其任何关联方的立场。

Such advertisements, companies, and individuals are not endorsed by Ah Capital Management LLC, a16z, or any of its affiliates.

所有信息均基于发布时的可靠来源，但a16z不保证其绝对准确。

Information is from sources deemed reliable on the date of publication, but a16z does not guarantee its accuracy.