Jerry Tworek：打造随用户实时学习的 AI，挑战 Transformer 架构

我其实不知道我们看起来是什么样，不过没关系。嗨，Jerry，欢迎来到 Venture with Grace。

I actually don't know what we look like, but okay. Hi, Jerry. Welcome to Venture with Grace.

非常感谢，很高兴来到这里。

Thank you very much. Very happy to be here.

那也许我们可以聊聊我们的万圣节特别装扮。今天是万圣节。

So maybe we can talk about our Halloween special costume. Today is Halloween.

今天是万圣节。好的。

Today's Halloween. Okay.

你想讲讲为什么选《猎魔人》吗？

You want to give the rundown of why you picked the Witcher?

我是猎魔人。利维亚的杰洛特，怪物与 Transformer 的猎杀者。

I'm a Witcher. Geralt of Rivia, Slayer of Monsters and Transformers.

对，这就是我现在的身份。

Yeah, that's the person I am right now.

有一种特定的神经网络架构，我认为它已经进入了每个人的家里和 AI 助手中，我们需要猎杀它，用更好的东西取代它。

There's a specific architecture of neural network that I think got its place in everyone's homes and AI assistants, and we need to slay it and replace it with something better.

我真的很喜欢现在任何事都能和 AI 扯上关系。

I really like that you can tie anything into AI these days.

我想从 Core Automation 开始聊。显然，过去大家都在谈 Transformer，而你们算是 Transformer 猎手。多讲讲吧。

I want to start with core automation. Obviously the big thing is, in the past everybody's talking about Transformers, and then you guys are kind of the Transformer slayer. Tell us more.

Transformer 太成功了，以至于 AI 研究者们有点自满，开始相信 Transformer 就是路的尽头。我们能够把它扩展好几个数量级，看到了性能提升，也看到了新能力。我想稍微反其道而行，证明其实还有不同的方式来排列神经元、不同的方式来在 GPU 上安排内存，从而带来更聪明的模型和更好、更稳健的未来学习系统。这就是我们要做的事。我在某些方面对 Transformer 已经交付的成果非常满意，对今天已有的 AI 助手也非常满意，但还有更好的技术值得去构建，这就是我和我的公司着手要做的事。

Transformers have been so successful that AI researchers kind of became complacent and started believing Transformer is the end of the road. We've been able to scale those multiple orders of magnitude and we've seen improvements and we've seen new skills. I would like to be a little bit counterculture to that and show there are actually different ways how you can arrange your neurons, different ways how you can arrange your memory on a GPU, that will yield itself to smarter models and better, more robust learning systems of the future. And this is what we are to do. I'm very happy in some ways what Transformers have already delivered, I'm very happy of the AI assistants that are there today, but there is a better technology to be built and this is what I together with my company have set out to do.

我有很多问题。第一个是，我觉得 Transformer 的特长在于识别上下文或内容中哪部分更重要，这是它的专长；然后它也需要大量训练数据之类的东西。所以对你们来说，听起来你们不需要很多数据？

I have so many questions. Number one is, I feel like Transformers' specialization is kind of for identifying which part of the context or content is more important. So that's kind of the specialization, and then it also takes up a lot of training data and stuff like that. So for you guys, it sounds like you guys don't require a lot of data.

我们确实需要数据。数据是必要的。我们认为我们可以做到更高的数据效率，会有一些倍数上的优势，但数据是需要的。我认为我们最想摆脱的，某种程度上是产品特定的、但更重要的是客户特定的微调。目前大实验室在购买海量数据，用来让模型擅长医疗、擅长服务律师、擅长服务程序员、擅长做公司估值。

We do require data. Data is necessary. We think we can be more data efficient. There's some multiplier, but data is needed. And I think the biggest thing what we want to get rid of is, in some way product specific but even more importantly customer specific fine-tuning. Currently the big labs are purchasing huge amounts of data for how to make models good at health, how to make them good for lawyers, how to make them good for programmers, how to make them good for making valuation of companies.

而且我认为，让模型在一个地方掌握所有职业的知识、掌握我们想让它做的一切，是不现实的。我们在 Core Automation 想做的是，构建一个能从用户数据中学习、并在与用户协作的过程中即时学习的模型。所以如果你是律所，你不必非得买一个对你公司案件无所不知的现成模型，而是它可以和你一起学习——你带着它、给它看你真正关心的数据，它从中改进。也就是把更多学习从实验室转移到真正的客户手中。

And I think it's unrealistic for the models to have knowledge of every profession in one place and everything what we want them to do. What we at Core Automation want to do, we want to build a model that learns from the user data and on the fly, working with the user. So if you are a law firm, you don't need to necessarily buy an off-the-shelf model that knows everything about your company's cases, but can learn together with you as you bring it along and show it to the data that you actually care about, and it improves from that. So like move more of the learning from the lab into actual customers' hands.

再多说说。基本上我觉得现在上我们播客的大多数人都在谈开放权重模型。本质上就是你下载一些东西，然后在你自己的数据库里运行。但你们和那有什么不同？

Say more. So basically I feel like nowadays most people coming out of our podcast talking about open weight models. So essentially you download stuff and then run stuff within your own database. But how do you guys differ from that?

首先，我们用的是不同的算法。开放权重模型的问题在于微调很难、效率很低。它需要很多人力，结果也没有真正保证，而且并不太好。

First and foremost, we use different algorithms. The problem with open weight models is that fine-tuning is hard, inefficient. It requires a lot of people and the results are not really guaranteed and not really very good.

所以你们保证结果？

So you guarantee results.

是的。我们想通过不同的架构，提供很多用微调 Transformer 很难做到的东西。就像一边是神经网络，另一边是数据，问题是怎么把这两者结合起来。微调是一种方式，但可以说它并不太好。

Yes. And we want to, through a different architecture, provide a lot of the things that are just hard to do with fine-tuning Transformers. Like on one hand you have a neural network, on the other you have data, and there's a question how do you combine those things together. Fine-tuning is one way to do it, which arguably is just not very good.

我们正在构建它的替代方案，我们认为那会好得多得多。

And we are building an alternative to that that we think will be much, much better.

好，那我想聊聊你更想谈哪部分，是人还是商业逻辑？因为显然我们先聊人吧，因为你们显然招了首席人力官，OpenAI 的前首席人力官，她显然在 Hudson Trading 待了超过 15 年，我相信她招募了大量优秀人才。上次我追踪你们团队，大概是 16、20 人左右。

Okay, so I want to talk about which part do you want to talk about more, the people or the business logic? Because obviously let's talk about the people, because you guys obviously hire the chief people officer, the former chief people officer of OpenAI, who obviously has been at Hudson Trading for over 15 years, who I'm sure has recruited a ton of amazing talents. And last time I tracked back on your team, it's like 16, 20 people, something like that.

目前是 27 人。

27 at this point.

好。哇，互联网上次更新之后已经 27 人了。我在 AI 人才领域有很多问题。显然其中一点是，如今人才比过去值钱得多。但如今这么多决策其实是由 AI 做出的，为什么有些人比别人更有价值？你们在做人才规划时，什么样的人算好人才？你来自交易时代——不是时代，是交易团队。那为什么那会是比学计算机科学的人显著更好的 AI 人才来源？

Okay. Wow. 27 since the last time the internet has updated. I have so many questions in the AI talent space. So obviously one of the things is talent nowadays is worth a lot more than the past. But nowadays so many decisions are actually made by AI. Why is some people more valuable than others? And when you guys are thinking about talent planning, what is considered a good talent? So you come from the training era, or not era, but the trading squad. So why is that a significantly better source of AI people than people who are studying computer science?

我们在很多方面确实非常看重人才，对吧？当你创办一个 AI 实验室时，人才是你最重要的资源之一。我们想打造的团队类型，通常是一支能快速行动、能构建一些前沿系统的小团队。我们用自动化、用智能体深度赋能每个人，所以我们可以用相当小的团队做出很多东西。但仍然需要一定数量的人。有时我开玩笑说，对今天的模型而言，人就像是它们之上持续学习的“肉身代理”。

We definitely care a lot about talent in many ways, right? Like when you are starting an AI lab, talent is one of your most important resources. And the type of team we want to build is generally a small team of people that can move very quickly and build some of the frontier systems. We enable people very deeply with automation, with agents, so we can build a lot with a pretty small team. But there's still a number of people that you need. And sometimes I am joking that for today's models, people serve as continually learning meat proxies on top of them.

显然，Julia 和我已经是好朋友一段时间了，我们基本上决定一起创办这家公司，我很在意我们如何建设团队、文化和工作方式。这就是我们在做的事。我认为我们挑选每一个人都非常有针对性，去承担特定角色。我们不是泛泛地招那些我们觉得人好又聪明的人。我们有一个使命要完成，我们组建团队几乎就像在选秀一支橄榄球队：这个人在这个位置，那个人在那个位置，我们就是这样运作的。

And we obviously, Julia has been a close friend for a while, and we basically decided to start this company together, and I cared about how do we build a team and a culture and how we work. This is what we are doing, and I think we are picking every person very specifically to fulfill a specific role. We are not just blanket hiring people that we think are nice and smart. We have a mission to do, and we are building a team almost like you're drafting a football team. This person in this role, this person in this role, and that's how we are operating.

我们在思考，未来的公司总体上会和今天的公司非常不同，随着 AI 自动化程度提高，它能承担大量琐碎的工作，以及过去相当耗费人力的苦活累活。问题在于，你需要多少被最大化赋能的人，以及如何把能比以往快得多的人类工作，组织进一家真正能交付相当大胆、相当艰难使命的公司里。

We are thinking that companies of the future generally will work very differently from companies of today, with increased automation of AI being able to do a lot of mundane work and a lot of grinding that was pretty labor intensive in the past. The question is how many maximally enabled humans do you need, and how to structure the work of humans that can work much faster than ever before into a company that actually delivers a pretty audacious and a pretty difficult mission.

你提到人才或者公司结构会不一样。那到底有多不一样？还有，你背后那两把剑是不是背累了？

You mentioned like the talent or like the company structure will be different. How different also like are you tired with like the two swords in the back?

你得有两把剑。一把是留给怪物的，以防万一。

You need to have two swords. One is for monsters, just in case.

好。那么，未来的公司和今天有什么不同？

Yeah. Okay. So how is a company in the future different from today?

我觉得未来的公司肯定在很大程度上是 AI 运营的。AI 会处理流经公司的所有信息。AI 会提出所有决策。AI 会操作大量软件和机器。问题在于，人类在这个过程中处在什么位置。我最喜欢的未来是，人类总体上给出目标——我们想用公司的资源实现什么。我们应该听取它的建议。但我喜欢这样一个世界：我们明确处于控制之中，握着越来越自动化的系统的方向盘，而这些系统总体上能完成我们想做的事。这有点像——我看到一种可能的未来：整个公司就是一场董事会，会上决定如何分配公司资源、我们想用这些资源做什么，然后 AI 去执行，下一次会议上我们看看进展如何、下一步做什么。要达到那一步还需要几轮迭代，但我确实看到我们正缓慢而确定地朝那个方向挪动。

Yeah, I think the company of the future surely is largely AI operated. AIs will be processing all the information flowing through the company. AI will be proposing all the decision making. AI will be operating lots of software and machinery. There are questions of where do the humans come in in that process. And the future that I like the most is where the humans generally give the objectives of what do we want to achieve with the company's resources. We should take advice from it. But I like the world where we are clearly in control and at the steering wheel of the increasingly automated systems that can generally fulfill what we want to do. And that's kind of like — I see there's a potential future where the whole company is just the board meeting where there is a decision of how do we allocate company resources, what do we want to do with those resources, and then AI goes and does it, and there's a next meeting where we see how did that go and what's the next thing to do. That's still going to take a few iterations to get there, but I do see how slowly but surely we are inching towards that.

好，我还有一个问题。现在我个人在用——你随便说一家实验室——我大概都在用他们的产品。所以除了——显然你们比较低调——你们……产品发布了吗？

Okay, so another question I had was like, you know, nowadays I'm personally using — you name it, a lab — I'm probably using their product. So besides — obviously you guys, since you guys are kind of stealthy — what do you — product out yet?

好。那你们内部用什么？你们有自己的——

Okay. So what do you guys use in-house? Like do you have your own like —

什么都用——我们越来越多地自己造东西。我们有内部的代码——或者说没有——但我们开始用 Codex 了，已经用了一段时间。某种程度上，我们正在慢慢迁移到自己的方案上。

Everything — like building more and more stuff ourselves. So we have internal code — or no — but we have started using Codex and we've been using it for a while. We are slowly migrating to our own solutions in some ways.

最先进的模型非常好，但模型之上的那一层还有很多可改进的地方，还有很多关于我们如何使用 AI 的问题——我认为会演进，而且确实在演进。Codex 有点像是一个非常通用的工具，对每个人都有用，但实际上在模型之上并没有增加那么多价值。

Like the state-of-the-art models are very good, but the layer on top of models still has a lot to improve, and a lot of like how do we use AIs — I think will evolve, and surely does evolve. Codex is kind of like a very general tool that can be useful for everyone, but actually doesn't add that much value on top of just the model.

我们造了一些东西，我们认为它们既更灵活、给个人更多能力，也更贴合我们想实现的目标以及我们希望如何协作。但我们用的模型在很大程度上仍然是 OpenAI 的模型。

We've built things that we think are both more flexible, give individuals more power, but also tailor slightly more to what we are trying to achieve and how we want to work together for us. But the models we are using largely are still OpenAI models.

你觉得有哪些东西你会带走——不管是文化，还是 OpenAI 做事的方式——带到你自己的公司，然后你们又在推行哪些新东西？

What are something that you feel like you would take away — whether it's like culture or how things have been done at OpenAI — that you would take to your own company, and then what are some of the new things that you guys have been implementing?

我最在意、并且在创办这家新公司时带过来的东西，是做一个创新者、敢于冒险。OpenAI 有一段时间是一家怀有非常大胆梦想的公司——要造出通用人工智能。它当时并不太清楚该怎么做，但对投身这个使命有很强的信念和信心。很多年里，OpenAI 年复一年地做突破性研究，朝那个目标前进，并且在如何为目标构建算力、如何为目标组建团队上非常有策略。它一直在朝那里走——如果没有 OpenAI，就不会有我们今天拥有的 AI。我认为，所有认为 OpenAI 是世界上一股正向力量的人，都应该为此极其感激。

The thing I care about the most, and I brought with myself when I started this new company, is being the innovator and daring. OpenAI for a while was a company that had very audacious dreams of building artificial general intelligence. It didn't have that much idea of how to do it, but it had a lot of conviction and a lot of confidence in going in that mission. And for many years OpenAI did breakthrough research, year after year, going towards that goal, and being very strategic about how to build compute towards that goal, how to build a team towards that goal. And was going there — we wouldn't have the AI that we have today without OpenAI being there. And I think all of us who think OpenAI is a positive force in the world should be extremely grateful for that.

所以，敢于冒险、高抱负，以及真正把力量集中在创新研究上，是我想要带进来的。而且正因为时代不同了、打法不同了——我认为像 OpenAI 那样的公司即使在今天也不可能再出现。它在创立那会儿是可能的，但格局已经变了一些。我试图汲取 OpenAI 研究文化中最好的部分，再加入一点我们自己的东西，比如我们所处的位置——这里事情变化非常快，所以如今我们需要更强的紧迫感。我们更像是一种反主流文化和弱者，因为我们在押注对抗现有范式，而且我们显然小得多、资源少得多。所以我们需要比过去的公司更加拼命、更加灵活。

So like daring, high ambition, and actually focusing your forces on innovative research is what I want to really take into that account. And just because times are different and plays are different — I don't think a company like OpenAI is possible even today. It was possible in the day it started, but the landscape has shifted a little bit. I'm trying to take the best of the research culture of OpenAI and trying to add a little bit more of our own, like the place where we are, where things move very quickly, so we need more sense of urgency these days. We are a bit more of a counterculture and underdog, because we are betting against the existing paradigm, and we obviously are much smaller and have less resources. So we need to be much scrappier than companies of the past.

我是说，你们没有——5.32 亿美元的种子轮。跟预算相比这感觉非常小——大家的预算都花在哪儿了？而且我记得英伟达也是你们的投资方。所以钱是不是直接又流回去了——

I mean, you have no — $532 million for seed. That feels very small compared to the budget of — where do people spend the budget? And also I believe Nvidia is your investor as well. So does it just go straight back to —

直接流回英伟达。我们想买的就是市面上最好的 GPU，而且要买很多。

It goes straight back to Nvidia. All we want to buy is the best GPUs on the market, and we want to buy a lot of them.

好。所以就是直接买 GPU。

Okay. So just straight up GPUs.

是的。

Yes.

钱还花在别的什么上吗？

Spending money on anything else?

我们在办公室上花了一点，但不多。

We spend a little on the office, but not that much.

好。

Okay.

还有零食。

And snacks.

还有零食。那——

And snacks. And the —

我们的零食挺古怪、挺讲究的。

We have pretty weird and fancy snacks.

好。比如什么？

Okay. Like what?

咸三文鱼皮。

Salty salmon skin.

咸三文鱼皮。在某些——

Salty salmon skin. At some —

这在亚洲很受欢迎。我以前没吃过，但我们有个智能体把订零食自动化了，结果零食的分布越来越古怪，到处都是。

It's popular in Asia. I haven't had it before, but we had an agent that automated ordering snacks, and that distribution was getting weirder and weirder all over.

哇。把这事交给智能体可是个非常重要的任务。

Wow. It's a very important task to hand it to the agents.

最重要的是，这可不是你真希望人类花大量时间去做的事。

Most importantly, it's not something you really want humans to be spending a lot of time on.

那商业模式呢？我感觉，你知道，OpenAI 以及——我想有两件事。一个是公司的战略，另一部分是商业模式之类的。我感觉现在——

What about like the business model? So like I feel like, you know, OpenAI and I guess like there's two things. One is like the strategy of the company, and then the other part is like the business models and things. I feel like nowadays like um —

天哪。真抱歉。显然每次我们录音，就有人开始煮咖啡。这对大家来说非常重要。

Oh god. So sorry. Apparently whenever we're recording, people start making coffee. It's very important for people.

是啊。很多人都很在意咖啡。

Yeah. A lot of people care about coffee.

我想，当你考虑商业模式时，或者看看过去 OpenAI 和 Anthropic 的情况——我猜 OpenAI 是从消费者驱动的产品开始的。从新闻来看，它似乎本来不会那么出名或那么受欢迎，但它成了一个非常受欢迎的消费产品。另一方面，Anthropic 则去打造了一个非常强大的工作工具。我只是好奇，从商业角度，你如何考虑在商业领域定位自己？

I guess when you're thinking about the business model, or like the past OpenAI and Anthropic situations—I guess OpenAI started with a consumer-driven product. From the news, it seems like it wasn't supposed to be as famous or as popular as it gets, but it became a very popular consumer product. On the other hand, Anthropic went to building a very strong work tool. I'm just curious, from the business perspective, how do you think about positioning yourself in the business arena?

而且既然你们是——我和许多不同 AI 公司的高管聊过，人们说这像是基础模型获得融资的最后一批。你对此怎么看？还会有更多 AI 实验室出现吗？为什么或为什么不？

And since you guys are—I was chatting with a lot of executives from different AI companies, and people were saying this is like the last badge of foundational model getting funding. What's your opinion on that? Are there more AI labs gonna come around, and why or why not?

是的，我确实认为我们可能是最后一批成立的公司之一，主要是因为 AI 的失控效应和模型变得如此之好，以至于初创公司很难赶上并进入擂台。我确实认为，初创公司存在并能获胜的许多原因在于大公司效率低下、行动太慢且无法适应，而灵活玩家能在快速变化的环境中有所作为。AI 在未来几年将做的是，大公司将能够非常快速地行动，因为它们不必处理周围所有的政治问题。它们只需用 AI 来调配资源。所以我确实认为那会——

Yeah, I do think we may be one of the last batch of companies getting founded, mostly because of the runaway effects of AI and models being so good that it will be hard for startups to catch up and enter the ring. I do think that a lot of reasons why startups exist and can win is because larger companies are inefficient and moving too slow and cannot adapt, in a place where a nimble player can in a rapidly changing landscape. What AI will do in the next few years is that big companies will be able to move very quickly because they will not have to have all the political issues around them. They will just be able to move AI to direct resources. So I do think that will—

我想我们可以继续。抱歉，我——

I think we can keep going. So sorry, I do—

随时，是的，就像这样——

Whenever, yeah, like this—

好的，所以我们在谈论——

Okay, so we're talking about like—

这需要很多资源来——

There it needs a lot of resources for—

继续，抱歉。

Keep going, sorry.

是的，所以故事是这样的：我认为 AI 驱动的效率很可能让大公司能够更快地竞争，而且在某种程度上，如果有一个没有工作、没有劳动力的未来，很难想象初创公司究竟会是什么样子。可能会有一些风险承担发生，但我认为它会不同，很难准确预测未来会是什么样子。但我们主要关注的是我们的模型和我们正在构建的东西，是为你工作的模型。Core Automation 对打造你的朋友或你的治疗师不感兴趣。它是那种你告诉它你想让它做什么，它就会为你完成的模型，在你关心的任务和你想要完成的所有工作上创造价值。我们甚至不想构建那种开箱即用、知道你想要的一切的模型。你教它们，就像入职培训员工一样。你拿那个模型，向它展示：这是我的房子长什么样，这是我的电脑长什么样，这是我的数据长什么样，这些是我想要你做的事情。我会给你演示几次，然后我希望你为我做这些事。

Yeah, so the story was: I think AI-driven efficiency will likely make big companies able to compete much faster, and in some way, if there is a future without work and without labor, it's hard to imagine how starting up companies exactly will look like. There's probably some amount of risk-taking that may take place, but I think it will be different, and it's hard to predict exactly how the future will look like there. But what we are largely focusing on for our models and what we are building is models that do work for you. Core Automation is not interested in building your friends or your therapist. It's the model that is meant to—you tell it what you want it to do and it gets it done for you, generating value on the tasks that you care about and all the work you want to be done. And we don't even want to build the models that have that ability off the shelf, that know everything you want. You teach them, like onboarding an employee. You take that model and show them: this is how my house looks like, this is how my computer looks like, this is how my data looks like, these are the things I want you to do. I will show you a few times, and then I expect you to be doing them for me.

你会说 Core Automation 会成为一个像智能体替代品那样的智能体吗？

Would you say Core Automation is going to become like an agent that's like an agent replacement?

这本来就是一个智能体，但我认为它作为智能体的体验会略有不同。如果智能体能随着时间的推移有意义地适应你，我认为它会更加个性化。

This is meant to be an agent, but I think it will be a slightly different experience with an agent. I think it'll be a bit more personal if the agent meaningfully adapts to you over time.

你喜欢我本能发送的邮件吗？

Do you like my email that I'm sent by instinct?

不。

No.

它够个性化吗，还是不够？

Was it personal enough or no?

我没注意。

I haven't paid attention.

好的。我只是好奇，现在当前的智能体格局有什么不同——就像每个人都有五个智能体，无论是 Rockbot 还是你最喜欢的智能体——它们和 Core Automation 的核心区别是什么？

Okay. So I'm just curious, what is the difference between the current agent landscape nowadays—like everyone has five agents, whether it's Rockbot or name your favorite agent—what's the core difference between them and Core Automation?

是的，我认为一点是当前智能体的所有个性化都是在上下文中发生的。所以你通常有一个系统消息，你可以稍微指定一下，然后你可以放入一些模型有时可以使用的记忆文件。这很好；这就是当今系统的工作方式。但一件事是，智能体并没有以任何有意义的方式真正学习任何新技能。就像如果智能体现在有某件事很难做，它不会自我改进，直到 Anthropic 或 OpenAI 为你发布一个新模型，或者为了完整起见，Kimi。而且它很难处理那些不太容易口头表达的事情。每个曾经提示过模型的人都知道并理解，并非所有事情都很容易描述。就像系统消息——你可以尝试以特定方式引导模型，但往往更令人沮丧。有时你想，哦也许我可以展示给他们看，也许我会告诉他们，嘿，少一点这个，多一点那个。但这样的更新应该发生——而且再次，在一个向你学习的模型中，它就会发生。模型会与你合作，在接下来的一天内，会在你的上下文中稍微更多、稍微更好地得到你需要的东西，并稍微更好地理解你在问什么，即使你模糊和不完整，也比当前模型好，因为当前模型自我更新的能力非常有限。

Yeah, I think one thing is how all of the personalization of the current agents happens in context. So you usually have a system message which you can specify a little bit, and then you can put in some memory files that the model sometimes can use. This is great; this is how today's systems work. But one thing is the agents don't really learn any new skills in any meaningful way. Like if there is something that an agent has a hard time doing now, it will not improve itself until Anthropic or OpenAI releases a new model for you, or Kimi, for the sake of completeness. And it has a hard time getting things that are not very easy to verbalize. Everyone who has ever prompted a model knows and understands that not everything is very easy to describe. Like system messages—you can try to steer the model in a particular way, but it's much more often frustrating than not. Sometimes you think, oh maybe I could show them all, maybe I will tell them, hey a little bit less of this and a little bit more of that. But such an update should be happening—and again, in a model that learns from you, it will just happen. The model will work with you, and within the next day will get what you need in your context slightly more and slightly better, and slightly understand what you are asking even if you are vague and incomplete, better than current models that just have very limited ability to update themselves.

所以现在模型更新是基于它输入的数据，而如今每家公司都在将自己的数据输入模型。所以我猜格式会变成什么样——在 Core Automation 智能体时代会有显著变化吗？你有没有想象过任何特定的用例会因此产生显著差异？

So when the model is updating right now, it's based off the data it feeds into it, and nowadays each company is feeding their own data into the model. So I guess what is the format going to be like—will it significantly change in the Core Automation agent era? And is there any particular use case that you're imagining this is going to create a significant difference?

我思考最多的是思考公司如何运营。我们如何将公司带到这样一个地方：人类进来做他们的工作——也许这里那里有一点智能体的帮助？但你越来越多地听到，许多公司甚至很难找到智能体有意义有用的地方，以及如何将这些转化为生产力，转化为 largely 由 AI 工作和智能体式劳动力驱动的地方，以及我们将其视为 less 劳动力场所、更多资源分配场所的地方。

The thing I think about the most is thinking about how companies operate. How do we take companies to a place where humans come in and do their work—maybe with a little bit of help of agents here and there? But you are hearing more and more how a lot of companies are having a hard time even finding the places where the agents are meaningfully useful and how to turn those into productivity, into places that are largely driven by AI work and agentic workforce, and places where we kind of think of it as less as places of labor and more like places of resource allocation.

我猜当涉及到收费时，如今人们在做基于使用量的收费等等，然后有些人有点在替代不同种类的人类劳动力。

I guess when it comes to charging, nowadays people are doing usage-based charging and things, and then some people are kind of replacing the human workforce of different kinds.

你觉得这方面会发生变化吗？对于收听我们节目的听众，或者我们很多 LinkedIn 上的观众来说，有时候人们在 AI 行业找工作。这对他们意味着什么？人们会去做一些明显不同的事情吗？显然人们可以去做饭或者做别的事情。

Do you feel like that aspect will change? For people listening to our show, or a lot of our LinkedIn audience, sometimes people are looking for jobs in the AI industry. What has that meant for them? Are people going to be doing something significantly different? Obviously people could be cooking or doing other things.

我确实相信我们正在走向我所说的“后劳动社会”。

I do believe we are heading towards what I call a post-labor society.

好。

Okay.

未来有一种可能，我们不一定需要工作。比如，我有时会把生活想象成高中，高中是人生中一段没人真的需要工作的时期。你不需要创造价值，你大部分时间都在学习和和朋友玩。如果我设想一种完全没有义务的生活会是什么样，那会比我经历过的高中好得多。我当时在做作业。

There is a future where we don't necessarily work. For example, I sometimes imagine life like high school, where high school is a moment in life where no one has to really work. You don't have to generate value; you mostly spend your time learning and hanging out with friends. If I imagine how a life could look if I was completely free of any obligations, it would be a much better high school experience than mine. I was doing homework.

如果你愿意，你还是可以做作业。

You can still do homework if you would like that.

我绝对算是个有点反作业的人。

I was definitely somewhat of an anti-homework person.

你是个反作业的人。

You're an anti-homework person.

对，我不太喜欢作业。

Yeah, I don't like homework that much.

是啊，你看起来像是两分钟就能写完作业，然后别人会抄你的作业。

Yeah, you look like you would finish homework within two minutes and then people would copy your homework.

我通常会早上做，不过是的，我不喜欢在家把时间花在这上面。

I would usually do it in the morning, but yeah, I don't like spending my time at home on it.

那么哪种工作会最快被取代？

So what kind of job will be replaced the fastest?

真的很难说。我觉得更像是各种类型的工作会逐渐被取代。看看今天的就业市场，程序员的工作真的被取代了吗？我不这么认为，但它已经被彻底改变了。每个工程师写代码都比以前快得多，方式也完全不同。有很多人几乎完全不再读代码了，在很多方面都是如此。所以这个职业变了，它的运作方式变了，我认为这也影响了就业市场，但我不认为可以说有什么被取代了。我认为如果要发生取代，更像是所有工作同时发生，而不是某个工作比另一个更快。当你只有部分自动化时，你通常仍然需要人在岗位上，仍然需要只有人类才能提供的这种监督水平和上下文。但一旦我们实现了能替你完成工作的技术，我认为它会在所有地方同时发生。我觉得这在某种程度上对社会稍微更好，因为我们不会处在这种奇怪的世界里：一个工作被自动化了，另一个没有，我们到底该怎么应对。

It's hard to say, really. I think it will more be like various types of jobs will be replaced gradually. When we look at the job market today, was the job of programmers really replaced? I don't think so, but it was already radically transformed. Every single engineer can write code much faster and much differently than before. There are whole groups of people that stopped reading code completely, almost, in many ways. So the profession has changed, the dynamics of it have changed, and I think that also affects the job market, but I don't think we can say anything got replaced. I think the replacement will happen, if anything, all at the same time rather than one job faster than the other. Once you have partial automation, you still usually need a human in the job, and you still need to be able to have this level of oversight and a level of context that only humans can. But once we achieve technology that can do the work for you, I think it will happen everywhere at once. And I think that's in a way slightly better for society, because we are not in this weird world where one job got automated and the other one didn't, and how do we really navigate that.

基本上所有工作会同时被自动化。这算是结论。不是自动化，而是某种程度上。

All the jobs will be automated at the same time, basically. That's kind of the conclusion. Not automated, but like somewhere.

是的，物理层面是一个很大的分界：可以在电脑上完成的工作，和真正需要你出力、搬东西的工作。这大概是我设想的一大分界。所有做硬件的人都知道，机器人在这方面还需要更长时间，但同时这个领域有那么多投资、那么多工作在做，所以最终也会发生。

Yeah, the physical aspect, that's one kind of big divide: jobs that can be done at the computer and jobs where you actually need your muscles and move stuff around. That's probably one big divide I am imagining. And everyone who works with hardware knows robots will take a while longer for it, but at the same time there is so much investment, so much work done in the space that also will happen eventually.

当你考虑商业层面时，你有没有发现——现在我感觉每次和创始人聊天，通常大家的观点都是你得先和客户聊，了解客户是什么样的、你的客户是谁。是企业客户，还是其他研究人员？你怎么考虑创收，因为显然你们是一家实验室，真正产生收入大概需要很长时间。你构建公司的思路是什么？你在追逐什么样的势头，比如为了融下一轮或者做些什么？

When you're thinking about the business side of things, do you find—nowadays I feel like whenever I chat with a founder, typically people's opinion is you have to chat with the customer first, hear about what the customers are like and who your customers are. Are they enterprise, or are they other researchers? And how do you think about generating revenue, because obviously you guys are a lab and it probably takes a long time to actually generate revenue. What's your thought process on building the company? What's the momentum that you're chasing to maybe raise the next round or do stuff?

首先，我们是一家研究公司。所以我们在考虑商业化之前，主要是在交付突破性技术。我不想参与现有的智能体市场——训练 Transformer 然后部署它们。那不是我们当初要做的事。所以首先我们需要构建我们想要的东西，然后才能把它部署到世界上。但我主要考虑的是企业客户，以及那些有兴趣、愿意经历内部转型、把业务转向越来越自动化方式的公司。再说一次，这会是渐进的，会一步一步来。但企业已经在各种职能中使用智能体了，我想通过我们不同的技术，为他们提供比现有方案更适合他们用例的东西。我理想的收费方式是按机架收费。我觉得很多企业喜欢本地部署硬件，我想说这就像是英伟达的 GB300 和 VL70 双机架，你可以放进自己的数据中心。我们放入核心自动化软件，那个软件会从你的数据中学习，并能执行你公司需要的所有业务功能。这就是我们想做的。我不认为按 token 收费在某些情况下说得通，但很多时候用户对用量很敏感，按 token 收费突然让你觉得——

First and foremost, we are a research company. So we are largely working on delivering the breakthrough technology even before thinking about commercialization. I don't want to participate in the existing agents market of training transformers and deploying them. It's not what we set out to do. So first we need to build what we want, and then we can deploy it to the world. But I am largely thinking about enterprise customers and the companies that are interested and willing to undergo the internal transformation of moving the business towards more and more automated ways of doing things. And again it will be gradual. It will be step by step. But businesses already are playing with using agents in various role functions, and I would like to offer them through our different technology something that is just much better for their use cases than the existing setup. And my ideal way of charging would be per rack. I think a lot of businesses like having hardware on-prem, and I would like to say this is like a GB300 and VL70 two-rack from Nvidia that you can put in your data center. We put their core automation software, and that software will learn from your data and can perform all the business functions that you need in your company. This is what we would like to do. And I don't think charging per token probably makes sense in some cases, but also very often users are usage conscious, and charging per token suddenly makes you feel—

分享。

Share.

拜托，我们在直播，抱歉。天哪，我的错，我不知道这个。我是说，我们只是来——我通常只租这个地方吃晚饭，所以通常这类事情不会打扰我们，但今天什么事都发生了，这让节目非常多样。

Please, we're streaming, sorry. Oh my god, my bad. I didn't know this. I mean, we only came—I typically only rent this place for dinner, so typically these kinds of things don't bother us, but today everything has been happening, so which makes the show very diverse.

我们只是偶尔被打断。

We only got interrupted sometimes.

抱歉，好。

Sorry, okay.

没关系。

That's okay.

对，按机架，机架。

Yeah, the per rack, the rack.

像是按机架，对。

It was like per rack, yeah.

按机架的构建模式，你基本上购买自己的硬件，用自己的容量，只为你服务。因为模型是定制的，那个机架有记忆，那个记忆整合你业务的数据，并用于你的使用。而且你知道你的数据不会离开你的系统。它不会流向某个你从未听说过、很难信任的其他公司。

Per rack building model where you basically buy your own hardware with your own capacity that is only serving for you. And because the model is customized, that rack has a memory, and that memory integrates the data from your business and is used for your usage. And you know that your data doesn't leave your systems. It is not something that goes to some other company that you never heard of and it's hard to trust.

所以我认为这是未来一种很好的商业模式。最终，算力驱动一切，我们希望成为 Nvidia 芯片之上的额外一层，帮助你实现业务目标。

So I think it's a good business model of the future. In the end, compute drives it, and we want to be kind of like an extra layer on top of Nvidia chips that helps you achieve your business goals.

我们来谈谈 Nvidia。显然你们和 Nvidia 关系很好。有些人既是投资者又是客户。那么你如何看待 Nvidia 这一层？显然我们播客上也有几个人在构建 Nvidia 层。Nvidia 仍然是市场上最大的供应商之一。那么你对行业内其他玩家在 GPU 方面的做法有什么想法？你如何考虑将解决方案扩展到不同领域？

Let's talk about Nvidia. You guys have a great relationship with Nvidia apparently. So some people are both investor and client at the same time. So how do you kind of like think about the Nvidia layer? So obviously there are like a few people on our podcast also like are building Nvidia Slayer. I obviously Nvidia is still like one of the like not well like the biggest provider in the market. So like what's your thought process on like these other kind of like players in the industry their approach to GPUs and how do you think about like maybe expanding the solution into different things as well?

是的，是的。这是个好问题。首先，至少在今天，Nvidia 构建了市场上最好的训练芯片，这就是为什么我们非常乐意与他们合作。有许多公司试图在硬件市场竞争，我很难说这会如何发展。我只是不是硬件市场的专家。我总是很高兴使用当天最好的硬件，并确保我们的模型在其中运行得非常好。我们选择不构建自己的硬件是有原因的。我们愿意设计和开发所有的深度学习栈以及所有训练模型的方式，但我们喜欢硬件层是我们起步的地方。所以我们采用现有的芯片，我们的目标是尝试弄清楚如何在现有硬件上开发具有最大学习能力的最佳模型，而现有硬件主要是为了非常快速地执行矩阵乘法而构建的。

Yeah. Yeah. That's a great question. First and foremost, at least today, Nvidia builds just the best training chips in the market, and that's why we are very happy to be working with them. There are many companies trying to compete in a hardware market, and it's hard for me to say much about how that will play out. I'm just not an expert in a hardware market. I'm always very happy to be using the best hardware of the day and make sure our models will be running very well in there. There is a reason why we didn't choose to build our own hardware. We are willing to devise and develop all of the deep learning stack and all of the ways of training models, but we like the hardware layer is where we are starting. So we take the existing chips and our goal is try to figure out how to develop the best model with the maximum amount of learning capabilities on the existing hardware which largely is built to execute matrix multiplications very very fast.

想想你作为 CEO 的日常，自己当研究员和经营公司有什么区别？因为我猜你似乎会喜欢研究。

Thinking about I guess like your day-to-day as a CEO, what's the difference between like being the researcher yourself to like running a company? Because I would assume like you seems like you would enjoy research.

我确实非常喜欢研究。这是我最喜欢做的事情之一。显然，作为 CEO 并建立自己的公司确实需要很多其他事情，比如你需要为公司确保资金。你需要面对一些外部事务，你需要能够讲述你的故事并向所有人解释，然后当今的 AI 公司并不真正存在于真空中，你如何融入生态系统以及你与其他公司的关系也非常重要，还有很多运营上的事情你需要以某种方式去做，当你是一家公司的 CEO 时，一切都是你的问题，一切都是你的责任，不仅仅是研究，研究是我最关心的事情之一。但我每天有很多其他头疼的事情和很多其他需要担心的事情，那些都超出了研究范围。但在某种程度上，尝试设计你想要构建的东西——这种新型模型——然后尝试将其带到世界，是一件极具赋能作用的事情。我们生活的现实是，创办新公司是实现某些技术目标的最佳方式之一，然后公司不仅仅是模型，不仅仅是训练，而是所有需要结合在一起的东西，它是一个需要很好运作的有机体，这就是我们，这就是我们正在做的事情。

I do enjoy research a lot. It's one of my favorite things to do. Obviously being a CEO and building your own company does require a lot of other things like you know one thing is like you need to secure funding for your company. You need to like be some of some external facing and you need to be able to tell your story and explain to all of other people then AI companies of the day that don't really exist in a vacuum and how do you fit into the ecosystem and how you relate to other companies is also very important and there are also a lot of operational things that you need to do in some way when you are a CEO of a company everything is your problem and everything is your responsibility not just research research is the only one of the things I care about the most. But I have a lot of other headaches and a lot of other things I need to worry about every day that are just beyond research. But in some way it is an extremely enabling thing to do to try to design something that you want to be built which is this new type of model and then try to bring it to the world. The reality we live in is a starting a new company is one of the best ways to achieve certain technological goals and then companies are not just models not just training but everything that needs to come together it's an organism that's that needs to function very well and that's what we are that's what we are doing.

是什么激励着你？我不知道，我听了很多你的播客，但我不确定你是否分享过来自波兰的故事，你知道你是如何长大的，然后你是否一直都知道你想成为一名企业家？显然你似乎有一个很棒的交易生涯，并且乘着 AI 的浪潮，但你喜欢做什么？你认为自己作为创始人的不公平优势是什么？

What motivates you since like you know you want to like I don't know like I don't think I mean I listened to a bunch of your podcasts but like I didn't know if you ever shared the story of like coming from Poland and is there you know how you kind of like grew up and then like how you kind of like do you always know that you want to become an entrepreneur or obviously it seems like you know you have a you know great like trading career and to like you know riding the wave of AI but what kind of like you know is something that you enjoy are doing and then what do you identify as your like unfair advantage as a as a founder?

是的，是的。我认为在很多方面主要驱动我的是我只是喜欢构建技术。这对我来说是让我起床并想做事的原因。我认为特别是构建新的世界问题，不是 AI，但我非常喜欢用科学和工程来构建事物和解决问题。这就是我喜欢做的事情，也是我想花时间做的事情。其余的问题是我以什么配置为世界其他地方做这件事，我一生中时不时地想过成为企业家。通常是一个问题：时机对吗？我有正确的想法吗？世界准备好接受我想做的事情了吗？

Yeah. Yeah. Like I think like what mostly drives me in many ways is I just like building technology. This is to me what makes me get out of bed and what makes me want to do things. I think building specifically new world problems not AI but I very much enjoy building things and solving problems using science and engineering. And that's kind of what I just like doing and what I want to be spending my days. And the rest of the questions how in which configuration for the rest of the world I want to be doing that like I have thought about being an entrepreneur on and off throughout my life. It's usually a question. Is it the right time? Do I have the right idea? Is the world ready for what I'm trying to do?

世界还没准备好，但很快就会准备好。

The word is not ready but it's it will be ready soon.

没错，没错。我大学时并没有真正有一个公司的想法，我认为那不是一家很好的公司，因为它主要是为了制作个性化的互动数学学习材料，那是我第一个互动数学，听起来像芯片。

Exactly. Exactly. There was I didn't really have an idea for a company starting in college which I don't think it was a very good company because it was mostly meant to be a company making like personalized and interactive like study materials for mathematics that was my first interactive mathematical that sounds like chip

在某种程度上，从事 AI 和 chd 帮助我意识到这一点，我认为那很好，我想过去并没有真正的好时机，也没有真正非常好的时机来实现这个想法。如今可能几个小时的 vibe coding，你已经可以有一个原型了。

Exactly in some way working on AI and chd helped me realize that and I think that is pretty nice I think I think back in the day there wasn't really a good time and there wasn't really a very good very good moment for this idea. These days probably like a few hours of vibe coding and you already can have a prototype of exactly

你刚才就像 tenure 一样。

You were just like tenure earlier.

没错。时机非常难把握，是创办公司最重要的事情之一。问题是，现在是否是建立我想建立的公司的正确时机，这极其关键，你不想太早，也不想太晚。

Exactly. Timing timing is very hard and one of the most important things in starting a company. The question is it is this the right time to build the company I want to build is extremely critical and you don't want to be too early and you don't want to be too late

你是在正确的时间吗？

Like are you in the right time?

我想是的。是的，这就是为什么我决定这么做。我认为创办我想做的那种公司的时间窗口实际上非常短。所以我没有太多时间犹豫。

I assume so. Yeah, that's why that's why I decided to do it. I was thinking that the window is actually very short for starting a type of company I want to do. So I didn't have much time to hesitate.

首要原因是，你需要智能体足够好来启动你，没有智能体，启动一个 AI 实验室是一个非常缓慢的过程，需要大量时间。如果你看看一年前、两年前开始的 neolabs，他们最初以冰川般的速度前进。你需要做的系统数量以及你需要构建的工作量，要达到任何目标都是非常多的。有了现代 AI 解决方案，我们可能能够将两年的工作压缩到 3 个月。令人惊讶的是，你现在可以比以往任何时候都移动得快得多。

The first and foremost reason is you need agents to be good enough to kickstart you without agents. Starting an AI lab is a very slow process and requires a lot of time. If you see the neolabs that started a year ago, two years ago, they were moving in a glacial pace initially. The amount of systems you need to do and the amount of work that you need to build out to get anywhere is just a lot. With like modern AI solutions, we're able to probably compress two years of work in 3 months. It's amazing how much quicker you can move right now than ever before.

第二点是，你不想太晚，以免别人在某些我们正在做的研究上已经相当领先。我认为我们正在引领找到一种与 Transformer 非常不同、在智能体式工作负载上比 Transformer 更好的东西。但最终其他公司也会尝试问同样的问题。我认为他们现在仍然忙于训练 Transformer，但这肯定会改变。

And the second thing is you don't want to be too late from a perspective of already someone being pretty far advanced into some types of research that we are doing. I think we are leading the guard in finding something that is very different to transformer, better than transformer in agentic workloads. But eventually other companies will be trying to ask the same questions. They are, I think, right now today still busy with training transformers, but that will definitely change.

你是怎么遇到 Rohan 的？

How do you meet Rohan?

给我们节目打个广告。其实 Rohan，看看我们的主题，我的联合创始人，伟大的科学家，对深度学习充满热情。当我试图为那家公司组建团队时，你可以想象，你给大致认识、也大致了解其研究的人打电话：“嘿，我正在组建团队，这就是我们想做的事，你加入吗？”Rohan 就是接电话并说“你知道，Jerry，我加入”的人之一。

Plug from our show. Like actually Rohan, let's look at our subject, my co-founder, great scientist and great passionate of deep learning. When I was trying to assemble a crew for that company, you can imagine you give people phone calls that you roughly know and roughly know their research: “Hey, I'm assembling a crew, this is what we are trying to do, are you in or not?” And Rohan was one of the people who answered that call and said, “You know, Jerry, I'm in.”

哇。我是说，有人拒绝吗？

Oh wow. I mean, is there anyone who said no?

肯定有些人拒绝了。我们决定加入的人的比例随时间波动，但尤其是早期，我们的成交率相当不错。

There are definitely some people that said no. And our ratio of people who decide to join is fluctuating over time, but especially early on we had a pretty good close rate.

你认为是什么让人们加入你们？仅仅是因为你们在解决世界级的、第一世界的问题吗？

What do you think makes people join you guys? Is it just because you guys are solving the world-class, the first-world problem right there?

是的。至少有一点：我们在招聘优秀研究员方面很顺利，因为我认为我们是既有雄心勃勃的研究计划解决新的有趣问题、又能向那些在我们旅程中同行足够久的人很好地阐述它的公司之一。对于其他职位，我们与其他公司稍微更一致。如果你在系统或数据核心自动化方面工作，故事的主要部分是你将尝试以尽可能 AI 原生的方式开发这些东西，使用尽可能多的自动化。你将努力成为一个人完成其他地方一个团队的工作量。这就是我们的优势。

Yeah. At least one thing: we have a very good time with hiring excellent researchers because I think we are one of the companies with both an ambitious research plan solving new interesting problems and actually we are able to pretty well articulate it for those who come along enough in our journey. And definitely for other roles we are slightly more eye to eye with other companies. If you work on systems or on data at core automation, the main part of the story is you will try to develop those things in as AI-native a way as possible using as much automation. You will be trying to become a person that does the equivalent of a work of a team in other places. And that's our advantage.

根据 Twitter，我觉得你最近不喜欢评估。

According to Twitter, I don't think you like evals these days.

是的。我认为评估大多会导致古德哈特定律和过于狭隘地思考问题。

Yes. I think evals mostly lead you to goodharting and to thinking about things too narrowly.

评估这个词有点像广告，对吧？就像你只挑你擅长的，然后也挑合适的评估东西。

The word eval is kind of like advertising, right? So it's like you just pick whatever you're good at, then pick the right evaluation things too.

那些东西里信号很少。每个发布模型的人都说“哦，我们在那些评估上很好”，他们 cherry-pick 并把模型呈现得很好，而那些模型的实际市场价值与基准所说的非常不同。因为那些是公开的，每个人都只想看起来好，而模型之间有更深层的差异。不幸的是，因为真的没有好方法，人们尝试所有不同的模型并试图弄清楚——至少那些愿意花精力的人。而不想花精力的人就选最流行的模型然后跑掉。但我认为这甚至不是思考模型、思考评估的好方法。这有点不合时宜，那是模型冻结、部署后没有真正有意义学习的时代。所以如果你拿一个现成的 GPT，它有一些评估分数，因为那是它训练的内容，但如果模型能学习，那么那些评估分数明天就不同了，甚至每分钟都可能不同。

There's very little signal in those things. Everyone who ever releases a model says, “Oh, we are good on those evals,” and they cherry-pick and present their models as good, and the actual market value of those models is very different from what those benchmarks say. And because those are public, everyone just tries to look good, and there are much deeper differences between models. Unfortunately, because there's just no really good way, people are trying all those different models and trying to figure out—at least those that want to spend energy on it. And people who don't want to spend energy just take the most popular model and run away with it. But I don't think it's even a very good way of thinking about models, about evals. It's a bit of an anachronism of times when the models were frozen and were not really meaningfully learning after they are deployed. So if you take a GPT off the shelf, it has some evals because that's what it was trained on, but if the model can learn, then those eval scores are different tomorrow, even can be different from minute to minute.

你怎么用——我是说，你怎么知道你的东西运行良好？

How do you use—I mean, how do you know if your stuff works well?

嗯，我知道，我测试它。我尝试它。

Well, I know I test it. I try it.

你尝试它。

You try it.

理想情况下是在我们正在做的工作上。如果模型为我做到了我想做的事，那么我是个快乐的人。如果它没有，那很模型。

Ideally on the work that we are doing. If the model does for me what I wanted to do, then I am a happy person. If it doesn't, that's very model.

我开玩笑的。

I'm just kidding.

什么是——我猜当你在思考建立一个成功的实验室和成功的公司时，让我们谈谈研究员。谁是好的研究员？我是说，显然我自己做过一些研究员活动，我认为在市场或创业圈，人们试图从 OpenAI 或 Anthropic 招聘人。但甚至每个人的头衔都是技术员工，现在还有非技术员工。那么你怎么——谁是合适的人？我是说，显然人们发表研究论文，你如何识别谁合适？然后你提到你组建自己的团队，人们有不同的职能，哪些职能通常最难招聘，哪些最容易满足？

What is—I guess when you're thinking about building a successful lab and building a successful company, let's talk about the researchers. Who are the good researchers? I mean, obviously I've done some researcher events myself and I think in the market or in the startup land people are trying to hire people from OpenAI or Anthropic. But even everyone's title is like member of technical staff, and nowadays there's the member of nontechnical staff. So how do you—who is the right person? I mean, obviously people publish research papers, and how do you identify who is the right fit? And then you mentioned about you assembling your own squad, like people have different functions, and what are the functions that are typically the most difficult to hire versus the most easier to kind of fulfill?

我不认为有那么难。这主要是特定时机和你的网络中谁在的交叉点，对于更容易招聘的角色，或者谁更热情。但一般来说，实验室需要填补多个角色，世界上显然有人在做这些。问题是：是否有合适的人拥有合适的技能，特别是现在，对你公司正在做的事情感兴趣？我认为优秀研究员的一个特点是：他们理解自己在做什么、在做什么工作。这可能是初级研究员和高级研究员之间最大的区别——你花时间深入思考你工作的主题，随着时间的推移，你对神经网络及其训练动态有越来越好的把握和心理理解。你通常想招聘避免群体思维、避免只是跟随趋势的人。

I don't think there are that much. It's mostly an intersection of a specific timing and who is in your network for the roles that are easier to hire or who is more enthusiastic. But in general there are multiple roles you need to fill in the lab, and there are obviously in the world people doing that. The question is: is there a right person with the right skill set, particularly right now, interested in what you are doing in your company? I think what makes good researchers is one thing: they understand what they are doing, what they are working on. This is probably the biggest difference between, let's say, junior researchers and senior researchers—the time you spend thinking deeply about the topics you work with, you have somewhat a better and better grasp over time and mental understanding of neural networks and their dynamics of training. You generally want to hire people who avoid groupthink and who avoid just following the trends.

太难了。

So difficult.

确实难，但这就是你想要的，尤其是对研究员来说。只思考别人在想什么的研究员并不是非常有用的研究员，因为研究是关于解决未解决的问题并尝试超越当前方法。这在当前世界有点棘手，因为风险很高，公司之间的动态就是这样，但在我看来，这是成功的必要条件。

It is difficult, but that's what you want, specifically for researchers the most. The researcher that just thinks what everyone else is thinking is not really a very useful researcher, because research is about solving things that are unsolved and trying to go beyond the current methods. Which is kind of tricky in the current world where the stakes are very high and the dynamics between companies is what it is, but is necessary to succeed in my opinion.

最后但同样重要的是，最成功的研究者通常对系统和 GPU 都相当了解，因为这归根结底是我们工作的底层基础，也是整个深度学习所依托的层面。我们需要能够把关于正在训练的神经元的洞见，与正在运行的数学模型的洞见结合起来。那我们怎么找到好的中间地带——我想，当你评估研究者时，一个研究者日常的工作大概是什么样子？然后你如何为每个人的研究设定——我不想说基准，但类似里程碑的东西？因为我觉得研究者听起来像是在为问题创造解决方案，然后做一堆测试，对吧？一个人设计新东西，多久算太久？你会根据某些特定功能来评判吗，还是你如何——不是基准，而是你如何知道人们确实在把事情做好？

And last but not least, the most successful researchers generally understand systems and understand GPUs pretty well, because that's in the end the substrate we are working with, and that's just the layer from which all of deep learning comes from. We need to be able to combine the insights we have about the neurons that we are training with the math models that we are executing. And how do we find the good middle ground between—I guess when you're evaluating the researchers, what does a researcher's day-to-day kind of look like? And then how do you set up—I wouldn't say benchmark, but like milestones for everybody's research? Since I feel like researchers, it sounds like people are creating the solution for things and doing a bunch of tests, right? How long would it be too long for someone to design something new? And would you judge based off certain features, or how do you—not benchmark, but how do you know people are actually performing on things?

是的，是的。某种程度上，一个研究者在一个实验室或像我们这样的公司里表现如何，最重要的一点就是：他们会不会卡在事情上？这大概是研究者头号的失败情形。他们看着一个问题，并不真正知道怎么解决，然后一直做那些不能推动他们前进的事。如果他们在一个地方卡得太久，没有获得任何新知识、任何新洞见，那就是毫无进展。我见过一些团队、一群人就是卡在原地，像是在跑步机上拼命跑却无法前进。好的研究者会很快想明白：我有没有办法至少推进我对某件事的理解？有没有我能知道的问题？即使问题非常难、感觉很远，通常你也得有某种办法走到下一步，再走到再下一步，在那里你确定自己在取得进展。这我认为对研究者非常重要。对我来说，问题在于：做这项工作的人是否周期性地加深了对所研究对象的理解？这已经相当不错了。另一点就是不要想得太小。这里有一些判断空间：今天我们往哪个方向走，明天往哪个方向走。他们是否擅长解读数据，是否足够有怀疑精神、不被愚弄？随机性也愚弄过很多人。这来自经验，是一项你需要练就的技能。通常如果人们把这些结合起来，最终会相当成功。机器学习现在是极其多产的领域。如果你聪明，身边有几个好同事，还有一定数量的 GPU，你几乎肯定能发现一些前人没发现过、且完全有价值的东西。

Yeah. Yeah. In some way, the most important thing about a researcher performing in a lab or in a company like ours is just: do they get stuck on things or not? This is probably the number one failure case of researchers. They just look at a problem, they don't really know how to solve it, and they keep doing things that don't move them forward. If they are stuck in a place too long without gaining any new knowledge, any new insights, this just is not going anywhere. And I've seen teams, groups of people just stuck in place, kind of like being on a treadmill trying to run and not being able to move forward. A good researcher tries to understand pretty quickly: do I have a way to at least advance my understanding of something? Are there questions that I can know? And even if the problem is very hard and feels distant, usually you need to have some way to get to the next step and to get to the step afterwards, where you're sure you are making progress. That's something that I think is very important for researchers. And to me, the question is: does the person doing their work increase their understanding of what they are working with periodically? That's already pretty great. And the other one is just not thinking too small. There's some discretion of which direction do we go today, which direction do we go tomorrow. Are they good at interpreting data, and are they good at being in some way skeptical enough, not being fooled? Randomness has fooled many people as well. It comes from experience, and it's a skill that you need to get to. Usually if people combine those things, they end up being pretty successful. Machine learning is an extremely fruitful place right now. If you are smart, if you have a few good colleagues around you and a number of GPUs, you almost surely can discover something that no one has discovered before that is fully valuable.

我们能聊聊人才格局吗？最优秀的研究者是在这些非常大的实验室里，还是有某些人们有点忽视的地方？还是你们直接从所有顶尖学校招人？以及你如何看待经验与——我想是学术经验，技术学术经验以及工作经验——相比一个完全没有经验但代码能力极强的人？

Can we talk about the talent landscape? So are the best researchers at these really big labs, or is there something that people are kind of overlooking? Or do you guys straight up hire people from all the top schools? And how do you think about experience versus—I guess academic experience, like a tech academic experience as well as work experience—versus someone who has absolutely zero experience but is a really crack coder?

是的，是的。我不认为任何人对人才有垄断。有一个有趣的——

Yeah. Yeah. I don't think anyone has a monopoly on talent. There's an interesting—

没有垄断。

Don't have a monopoly.

我不认为我们对人才有垄断。我不认为任何人有。我认为人才有自然分散的能力。前沿实验室确实在某个时刻是最了不起的人们的绝对温床。有时你甚至不知道是那些人很了不起，还是只是在对的时间处在对的地方。但毫无疑问，在那里的人确实做出了绝对了不起的东西，取得了伟大成就。但很多人会同意我的说法：随着成长和岁月流逝，至少非常大的公司的人才密度随着时间略有下降。但我认为更有破坏性的是，它们在各种方面变得稍微更官僚了，这降低了能动性。公司越小，个人越容易取得巨大成果。公司越大，团队或结构越多，就越难做到这一点。这就是人才稍微分散的原因，有很多人创办初创公司、加入初创公司，尝试做这类事情。我不认为有一两家初创公司拥有最好的人。我觉得恰恰相反：几乎每家初创公司都有几个真正了不起的人，他们之前取得过成就，显然非常优秀，他们在建立自己的团队和小组，押注新一代研究者来驱动这家初创公司的发展。所以某种程度上，在基础层面上，人才在系统中有些均匀分布，当然也有各种——有些公司人才密度稍低，有些稍高。但极其有才华的人会寻找那些能用自己在地球上的时间和脑力取得最大成就的地方，这差不多是很多人的优化函数。

I don't think we have a monopoly on talent. I don't think anyone has. I think talent has a natural ability to disperse. Frontier labs definitely at some moment were the absolute hotbed of the most amazing people. Sometimes you don't even know if those people are amazing or it's just a good place at a good time. But definitely people being there have built absolutely amazing things and had great achievements. But many people will agree with me when they say that throughout growth and throughout years, the talent density of at least very big companies has slightly lowered over time. But I think the more damaging thing is that they've become slightly more bureaucratic in various ways, which just lowers the agency. When a company is smaller, it is easier for an individual to achieve tremendous outcomes. The bigger the company is, the more teams or structure there is, it's kind of harder to achieve that. And that's why talent disperses a little bit, and there are many people starting startups and joining startups and trying to do things like that. I don't think there are one or two startups that have the best people. I think it's kind of the opposite: almost every startup has a few really great people that have achieved things before and are clearly excellent, and they're building their own teams and their own groups of people, trying to bet on the new generation of researchers that will drive what this startup will do. So in some way, on base levels, the talent is somewhat equally spread in the system, with obviously various—there are some companies that maybe are slightly less talent dense, some slightly more talent dense. But extremely talented people seek the places where they can achieve the most with their time on earth and with the brain power that they have, and that's kind of the optimization function for many.

你们实验室有没有设定某种不同的里程碑？因为我觉得，Anthropic 和 OpenAI 算是第一波实验室，我们能看出它们有 IPO 的潜力，而且显然它们比其他很多公司花的时间短得多。

Is there a different milestone that you kind of set your mind to for your lab? Because I think the thing is, Anthropic and OpenAI are like the first wave of labs where we can see there's a potential for them to IPO, and obviously it takes them a lot shorter than a lot of other companies.

我不会说更短，因为我觉得——你知道，我是七八年前起步的，而现在有像 Stripe 这样的公司，上市要等很久。显然，人们达到那个地位的时间跨度已经完全不同了。而且现在我记得大概要 3 亿美元 ARR 之类的，才能真正达到那个地位。我相信未来 10 年这还会变。在你心里，有没有一种不同的目标时间区间——比如我们会花三年在 X 上，也就是开发，我不知道，最好的模型和实验室，打磨技术，然后也许接下来三年做 GTM，不管是做产品的消费级版本还是别的什么？所以你是有一个总体、非常具体的计划，还是觉得更多是你现在把日常落地做好，然后走一步看一步？我敢肯定也有公司大概已经开始，想在某个时候收购你们——或者就是昨天，我不知道。但是啊。

I wouldn't say shorter, since I feel like — I started, you know, seven, eight years ago, and nowadays we have companies like Stripe that take forever to IPO. Obviously it's just a really different range of timelines for people to achieve that status. And nowadays I believe it was like 300 million ARR, or something like that, to actually achieve that status. I'm sure in the next 10 years it will shift. In your mind, is there any different kind of target time zone that you think about — like, oh, we're going to spend three years on X, which is developing, I don't know, the best model and lab, and refining the technology, and maybe the next three years on GTM, whether it's creating a consumer version of the product or stuff like that? So do you have a general, really concrete plan, or do you feel like it's more towards you implementing the day-to-day now and then you will see what's happening? I'm sure there's also a company probably starting, trying to buy you guys at some point — slash maybe yesterday, I don't know. But yeah.

是啊是啊，不不。我觉得，你知道，我们有一份内部路线图，相当精确，但同时这不是我们会公开谈的东西——我们现在在做什么，以及我们想在年底前达成什么。但这些东西我们有，我们已经定好了，我们知道我们要往哪走。就创办公司而言，我觉得 Anthropic 取得了不可思议、不可思议的成就——五年内建成一家万亿美元公司。现在很多人看他们的投资人材料，说他们大概是资本主义史上效率最高的 Scaling（规模扩张）公司，这很了不起，绝对是一项成就。但我认为我们处在一条指数级加速的曲线上。所以我确实相信会有下一家公司，效率更高、扩张更快。我相信会有下一家公司——大概已经起步了——下一家会在三年左右成为万亿美元公司。

Yeah, yeah, no, no. I think, you know, we have an internal road map, pretty precise, but at the same time it's not something we really talk about publicly — what we are doing right now and what we want to achieve by the end of this year. But those things we have, we have set, and we know where we are going. In terms of building companies, I think Anthropic has achieved this incredible, incredible achievement of building a trillion-dollar company in five years. And many are saying, right now, looking at their investor materials, that they are probably the most efficient scaling company in the history of capitalism, which is something — definitely an achievement. But I think we live in an exponential curve of acceleration. So I do believe there will be a next company that will be even more efficient and that will scale even faster. I believe that there will be a next company — probably already has been started — the next company that will become a trillion-dollar company in like three years.

你知道这些人是谁吗？

Do you know who these people are?

哪——

Which —

我是说，下一代这样的公司。

I mean, like the next generation of these companies.

一些核心自动化。

Some core automation.

我们知道一些，但看看未来会怎样吧。不过我觉得至少其中一家会。

We have some, but let's see how the future will go. But I think at least one of those companies will.

我想注意一下时间。也许我们问一个正经问题。一个正经问题是：以今天——你对 AI 了解多少——你觉得 AI 领域里最被低估的技术是什么？

So I want to be mindful of the time. Maybe we do like one real question. So one real question would be: given today's — how much do you know about AI — what is the most undervalued technology that you think in the AI arena?

总体上是网络架构以及相关研究。我觉得它被严重低估了。人们有一阵子把它忘了，它会回归的。

Generally the network architectures and researching that. I think it's severely undervalued. People forgot about it for a while, and it will have a comeback.

会像加密货币一样回归。你不信加密货币？开玩笑的。好，快问快答。你最喜欢的书是什么？

It will have a comeback like crypto. You don't believe in crypto? I'm just kidding. Okay, fire round for you. What's your favorite book?

《猎魔人》。

Witcher.

不会吧。什么？好吧。

No way. What? Okay.

你刚说快问快答。所以我就说了第一个想到的。

You just said fire round. So I said the first thing that came to my mind.

真不敢相信你读过这本书。

I can't believe you read the book.

我读过这本书。显然，我读过书。

I read the book. Obviously, I read a book.

你会邀请谁参加晚宴？

Who would you invite to a dinner party?

巴拉克·奥巴马。

Barack Obama.

谁对你的职业生涯影响最大？

Who made the biggest impact in your career?

好问题。大概是某位老前辈吧。

Good question. Probably some old one.

工作之外在哪里能找到你？

Where can we find you outside of work?

在家。我不做太多事。我就是工作，陪家人。

At home. I don't do many things. I just work and spend time with my family.

很喜欢。好，Jerry，非常感谢你今天来做客。这次对话太有意思了。

Love that. Well, Jerry, thank you so much for coming on the show today. It was such a fun conversation.

谢谢。很高兴能来。

Thank you. Happy to be here.

让我——我想我们的——要来替我们结束对话了，因为——

Let me — I think our — is going to end the conversation for us, because —

等等，我正好在——结束

Wait, I end at the exact —