第73集：Felicis普通合伙人Peter Deng谈AI定价策略、对GPT-5的反应及语音技术被低估的原因

Peter Dang曾在硅谷多家顶尖公司担任产品负责人，包括Uber、Facebook、OpenAI，如今又作为投资人加入Felisys。我是Jacob Efron，在本次《监督式学习》节目中，我们探讨了许多有趣的话题。鉴于他在AI关键问题上的多重身份经历——比如OpenAI的未来前景，以及ChatGPT如何逐步与Uber等应用交互——能听取他的见解尤其令人兴奋。我们还讨论了产品管理的未来演变、AI对岗位职能的重塑、AI产品定价的挑战与创新模式，以及Meta未来可能的AI战略布局。

Peter Dang has led product at an ensemble of some of the most important companies in Silicon Valley, Uber, Facebook, OpenAI, and now as an investor over at Felisys. I'm Jacob Efron, and on supervised learning, we talked about a ton of interesting things. It was particularly fun just to get his perspective given all these different roles he's had on some of the key questions in AI, like what the future of open AI looks like and how ChatGPT interacts with apps like Uber over time. We talked about the future of product management and how AI changes that what and those roles will look like going forward. We talked about pricing AI products and some of the challenges in that and new models that might work, as well as what Meta's future AI strategy could look like.

能与对这些问题都有独到见解的人展开深度对话实在精彩。相信听众会非常享受这次交流。闲言少叙，有请Peter。非常感谢你参加本期播客。

Just fascinating to be able to riff on all these topics with someone with such an interesting perspective on them all. I think folks will really enjoy it. Without further ado, here's Peter. Well, thanks so much for for coming on the podcast.

谢谢邀请。

Thank you for having me.

你曾在Uber、Facebook、OpenAI等企业担任过卓越的产品领导职务。当下很多人都在探讨这些新型AI企业的本质，思考传统SaaS领域的经验哪些仍然适用，以及这类企业的独特差异性。我想先从你的专业领域——产品端开始讨论。在这个新时代，你觉得哪些方面需要重新思考？

You've had these incredible product leadership roles, you know, across the board, Uber, Facebook, OpenAI. And a lot of people are discussing now, you know, and thinking about what these new AI businesses are, and and what lessons, I guess, are still relevant from the previous SaaS world, and then what makes these businesses unique and different. And so I figured we'd start in in your domain of expertise on on the product side. Sure. What do that feels like maybe is is should be thought of differently in this in this new era?

实际上变化最大的是人们对价值的认知。具体来说，订阅模式很可能发生重大演变。举例说明：当我加入OpenAI负责消费者产品、产品营销和设计时，接手的首个项目是企业版ChatGPT。

Actually, one of the things that changes the most is how people think about value. And and specifically, would say subscriptions are probably gonna be something that's gonna evolve pretty dramatically. So I'll give you specific examples. Right? So when I joined, OpenAI, I led consumer products, you know, product marketing and, design.

当时我迫切想亲自参与实际工作，记得起草了首版需求文档后，团队仅用七周就完成上线——这速度快得惊人。临近发布时我感染新冠，在地下室隔离办公期间仍在讨论定价策略。

And the first project I jumped into was Checheby for enterprise. I wanted to kinda, like, roll up my sleeves and kinda get my hands dirty and actually do something useful. So I remember, you know, writing the first spec for that, or or kind of vision document and just kind of the team ran with it really quickly. I think we launched it in seven weeks, which is crazy. And I remember around the time that we were about to launch it, I had I had COVID, I was just working in my basement, kinda quarantined, and I was having calls about how we're gonna price this.

不知是否新冠幻觉，但我强烈感觉按席位定价存在问题。这种模式时而低估时而高估实际价值，因为AI产品的价值单位不是工具使用权（SaaS的典型模式），而是实际完成的工作量。虽然可以通过提高单席位定价来覆盖GPU重度用户成本，但我始终认为应该更贴近'实际工作量'这个价值单位。

And maybe this is a COVID hallucination or whatnot, but it's like, I feel like there's something wrong with pricing it per seat. And I had this, like, really strong feeling that we were undervaluing some sometimes and overvaluing sometimes because the unit of work is not just the fact you get to use a tool, which is what SaaS usually is. Fundamentally, the unit of work is the work provided and and work done. So it never sat right with me. And, you can price it per seat high enough so that the whales who use a lot of GPUs, you can still make the cost work out.

在AI时代，优秀的产品体验依然重要，但更期待初创企业转向思考'如何定义有价值的工作产出'。目前尚无完美答案——基于使用量的定价也不理想，因为token数量并不等同于价值创造。

But I really thought of it in terms of how can we actually get closer to the unit of value, which is how much work is actually done. So I think in the age of UI, you know, delightful products still gonna matter. You're still gonna be able to have to do something super useful, but I wanna see companies and startups start to pivot towards what does, you know, valuable work creation look like? And I don't have the exact answer. I don't think it is actually usage based pricing because tokens does not equate value.

这需要从SaaS模式向AI新世界进行根本性演变。

Right? It's it's a it's a level of work that's done, but it's not maybe the value that's actually created and captured. So I think that's gonna evolve a ton from the SaaS world into the new world of AI.

这个话题非常有意思，虽然人们常讨论基于结果的定价模式，但...

Well, it's such an interesting thread because I feel like, you know, people talk about a lot about outcomes based pricing, but, you know

是的。

Yeah.

实际上，我认为你在客户支持领域可能已经看到过，也许还有其他几个领域，但目前按席位定价似乎仍是AI应用中最主要、占主导地位的定价方式。你有没有见过哪些你觉得有趣的定价方式？

Realistically, I think you've seen it in the customer support space, maybe a few others, but it still feels like seat based pricing is probably the the dominant, you know, predominant way of doing things today in AI apps. Have you seen anything that that you thought is interesting in in terms of ways of pricing?

嗯，我见过一些初创公司基本上采用平台费加某种价值费的模式——我会说他们收取平台费和近似使用费，但他们会从创造的价值倒推定价。公式大致是：我们让多少人提升了效率？这带来了什么影响？然后倒推计算得出平台费和使用费。

Well, I've seen some startups who've basically done sort of a platform fee and a kind of value sort of a I would say they would have a platform fee and a usage fee ish, but they would work backwards from the value that they create. So the formula would be roughly like, hey. How many people were we able to make more efficient? What was was the impact of that? And then do some math backwards to be like, well, this would be the platform fee, this would be the usage fee.

我认为这解决了大部分问题，但买家对转向基于结果的定价始终会有心理障碍，某种程度上这是一种奇怪的逆向激励。不过我认为会出现更易接受的定价方式，人们不会因为平台费或使用费而退缩，因为他们一年后能倒推出实际成果，这很合理。

I think that gets at most of the problem, and I think that we're gonna buyers are always gonna have a mental hang up from going to outcomes based pricing, and, you know, it's a weird kind of perverse incentive in a way. But I think that there will be some, more acceptable ways of pricing a product where people are not gonna bulk out a, you know, a platform fee or a usage fee because they're gonna realize the outcomes that they can work backwards from after a year, it just makes sense.

如果换成ChatGPT企业版会怎样？你们可能会如何定价？

What would that have looked like with ChatGPT Enterprise? Like, how might how might you have priced that?

我认为我们其实无法对ChatGPT企业版这么做，因为它的使用场景太广泛了。对吧？就我个人而言，我并不是日常活跃用户。

I think I actually think that we couldn't have done it with ChatGPT Enterprise. I think the reason for that is because the use cases were so broad. Yeah. Right? And and for me, I'm I'm not a daily active user.

我大概每两小时或每小时就会用一次ChatGPT。从审阅投资备忘录时问'哪里可以精简'，到研究某个具体问题，甚至作为随时陪伴的思考伙伴——比如下班开车回家的路上，通过语音模式深入探讨研究领域，就像私人播客一样。

I think I'm, like, a bi hourly or hourly active user of ChatGPT. It's so useful from, like, taking a look at, you know, an investment memo and saying, like, hey. Where can I tighten up? Or, you know, do some research on x, y, and z thing, or even just being that thought partner, which I find to be just where it's always with me. Right?

由于ChatGPT应用范围太广，定价确实很难。但对许多初创公司来说，它们的应用场景会更聚焦。

Whether it's voice mode, driving down home after a long day of work at on the two eighty, and just going into my own private podcast because I can just go deep on a on area I'm researching. Right? So I think it's really hard with ChatGPT because of the broad, widespread use of it, But I think for a lot of different startups, it's gonna be a lot more targeted.

你提到不认为按使用量或token计价是正确方式，这很合理。这与终端用户的价值认知总存在矛盾——既要让人们体验模型的强大功能（作为初创公司当然希望用户高频使用），但有些人可能会产生高额账单。

And you kinda mentioned, you know, that you didn't think usage based pricing was your or pricing per token is, is the right way to do it, which makes sense. It's always kind of, counterintuitive to the value of the end user. There's this tension, I guess, of both allowing people to experience the power of these models, and and in some sense, you want is a startup is people to use it as often as possible. Totally. And then at the same time, some people can run up some pretty high bills.

在产品层面，你有发现什么有趣的解决方案吗？

Yeah. Have you seen anything interesting there on the product side?

我是说，我们在ChatGPT上处理了很多关于限制的问题。确实如此。对吧？总有些鲸鱼用户会在高峰时段过度使用服务。我认为产品方面的关键在于确保80%到90%的用户在使用时不会感受到任何限制，因为他们根本不会触及那些上限。

I mean, like, we dealt with lot with this with limits at on ChatGPT. Yeah. Right? There are definitely those whale users that are just using, you know, kind of taxing the service at at peak times. And I think the trick is on the product side is to make sure that people don't feel any constraints where they're using it because 80%, 90% of the people will never run into those limits.

但当用户接近或超出限制时，关键是要设计一种体验，让他们不会觉得‘我做错了什么’或‘我搞坏了系统’。而是温和引导他们意识到‘这是下一级别的服务’。按席位定价的简洁之处就在于，用户会觉得‘我基本可以随心所欲使用’，因为他们很可能永远不会触及那个上限。

But when you get close to those limits and when you maybe exceed them, it's really important to figure out what is that experience that people have that doesn't feel like, oh, you did something wrong. Oh, I broke something, or, oh, I'm I'm a bad, you know, user. It's it's it's just gently guiding them to be like, oh, well, here's the next tier. Right? And I think that in a sense, that's where the simplicity of of seat based pricing works is that people can just think of, you know, I have pretty much all I can eat because that's all I can eat is most likely under this limit, and they never experience that limit.

这是引导用户升级的绝佳方式。对于这类通用型产品，按席位定价能有效找到平衡点，让用户自主选择是否承担超额费用或升级到更高层级。

And then that's a really good way for them to upsell to the next level. So I think for a generalist product like that, I think that if you if you are pricing by seat, that is a great way for you to kinda get to the point where you you find that efficiency and people can can self opt into an overage cost or the the next tier.

没错。这在企业级市场会特别有趣——企业支付基础平台费用后，管理员需要为少数重度使用者选择合适的服务层级。

Right. So it'll be interesting, especially at the the enterprise level where you have, you know, maybe within these, you know, some broad platform fee and then a few power users in the enterprise and how enterprise admins figure out, you know, the right the right tier to be in.

正是如此。我强烈建议任何公司都不要纯粹按使用量定价。一旦以GPU使用量为计费单位，就容易陷入与竞争对手的恶性价格战。

Right. Exactly. And then, you know, it it yeah. That's the thing is I think that any company I would give strong advice not to basically purely price on on usage. I think the when you start to think of the unit as the GPU usage, to put it bluntly, I think you get into a commodity, like a race to the bottom type thing with your competitors.

对吧？

Right?

是的。常见模式似乎是基础平台费加上使用量上限...

Yeah. No. It feels like a common pattern we see is like a a broad platform fee with maybe like a cap on Yeah. The

某种上限机制。

Some kind of cap.

对。我很好奇你对产品合作的看法——这波AI浪潮中最出色的产品，似乎都来自研究团队与产品团队的深度协作。你在OpenAI这个典范企业工作过，能否分享这种协作的最佳实践？

Yeah. You know, another area I was curious to get your thoughts on on the product side is it feels like in this wave of AI companies, a lot of the best kind of products that are being built are really partnerships between, you know, the research and post training teams and those folks that are actually building products. And you did this at at probably the the company that does it best at OpenAI. Would love your thoughts on just like, what does this collaboration look like when it's done really well?

这是个好问题。作为产品人出身的我始终坚信：产品经理要填补空白领域。我们本质是服务者，负责将不同专业领域的力量凝聚起来。

Love that question. So ever since I was so I I grew up as a a product person. That's like kind of where my career kind of started off, And I've always believed in this principle that the PM fills in the white space. Like, we are a service organization. We basically help bring various different disciplines together.

在不同公司里，这种结构呈现方式各异。我举个例子，在Facebook、Instagram这类公司，采用的是产品经理、设计师和工程师的三足鼎立模式。大家各自为不同的目标函数努力，正是在这种相互挑战中才能产出最佳成果。而Uber的模式则稍有不同。

And in different companies, it just looks different. So I'll I'll give you an example. At at Facebook and Instagram and and and, those companies like that, it it was a triad of the the product manager, the designer, and the engineer. And everyone was kind of, like, pulling for different objective functions, and that's where you get the best work is that people are challenging each other. And then at Uber, it's actually was a little different.

他们是由产品经理、工程师、设计师和运营人员组成的四方架构。没错，毕竟涉及实体世界运营。我特别享受从纯数字领域转向数实结合的过程。而当你进入OpenAI这类代表新产品浪潮的企业时，又会遇到全新的配置组合。

It was the product manager, the engineer, the designer, and the ops person. Yeah. Because you operate in the physical world. And I love that experience going from purely digital to digital and physical. And then you you switch gears into OpenAI and and and kind of the new wave of products, you have a new configuration.

对吧？现在真正关键的角色是产品负责人、设计师、工程师和训练后环节主管——因为在很大程度上，模型本身就是产品。同时PM的职责也在转变，因为新的产品规格其实就是评估标准。当涉及智能体行为定义时，仅靠文字描述是不够的，必须通过评估体系来明确成功边界。这些公司彻底改变了游戏规则，训练后主管、产品主管、设计主管和工程主管必须像精密仪器般协同运作。

Right? It really is the product person, the designer, the engineer, and the post training lead because you need to you know, in in a large sense, the model is the product. And, also, the the the job of the PM starts to shift as well because, you know, the new spec is the eval. And when you're because, you know, you're talking about behaviors and you're talking about, you know, what's what does this agent do to if you wanted to talk about agents, you can't do that in just pure words without actually putting pen to paper on evals of how to how you define the boundaries of what success looks like. So absolutely, you know, the the the game changes with all these companies, and it was really important that the post training lead, the product lead, design lead, and the engineering lead are just all kind of operate as one unit.

但这其实是经典模式。每一轮技术浪潮都会引发这种转变。归根结底，PM永远是要拾起所有未被覆盖环节的那个人。

And but that's that's classic. I think with every single wave, you've seen the shift. Right? And at at the at the kind of the core, the PM is gonna be the one who picks up any piece that is not being picked up.

确实。看来评估体系是这轮浪潮中的关键新要素。百分之百同意。

Right. And so it feels like evals are kind of the the the big new piece in this wave. 100%.

没错，百分百认同。

Yeah. 100%.

把这些评估结果同步给训练团队，才能更有效地构建针对性模型。有意思的是，我们经历了多轮技术泛化浪潮——过去几年模型能力突飞猛进，每个新版本都能全面提升各项任务表现。而现在我们进入了针对终端任务进行专项训练的时代，比如GPT-5在编程方面表现突出，但未必对所有企业级任务都有同等提升。

And then bringing them obviously to, you know, to the pros training team, you know, to to then more effectively build build models for those use cases. Interesting. I feel like we've gone through, you know, different waves of generalization where obviously, you know, we had these, you know, for the last few years, these tremendous step changes and model improvements, and across the board you just got at any task you were working on, it just got better as the next wave of models came out. Now it feels like we're in this era of much more specific post training for end tasks, and it's like, cool, even with GPT-five it came out, it was better for coding. It's not necessarily better for, you know, every task that, every company ends up doing.

感觉原本存在于基础模型内部的反馈闭环，未来会越来越广泛地出现在整个AI应用生态中。

And so it feels like more and more this loop that maybe existed within the foundation models is, you know, gonna increasingly, I think, exist across the AI app ecosystem.

完全正确。当观察模型行为模式时，会发现它们特别擅长强化学习和工具使用这两个领域。

100%. I think that's that's that's gonna be a huge part of it. I think that, like, when you take a look at sort of some of the behaviors that the models take on and what you wanna teach it to do, one of the things that the models has gotten really good at is just, you know, RL's gotten really good. Yeah. Tool use has gotten really good.

这两种能力的结合为Starobs等应用开辟了全新的抽象层级。关键在于，模型尚未接触过大量非公开网络数据。如今模型已具备足够能力适配这些专属数据集和独特洞察，现在正是创业的黄金时期。

And those two combinations are just basically the models getting really good at that, opens up a whole new level of abstraction where Starobs can build on top of. And the secret is there's a lot of data that actually isn't on the open web that the model's been trained that have not been trained on. And so being able to to apply the those datasets and the unique insights, the models are good enough to to mold to that now, and that's a really good time for for folks to be building companies.

确实。当你观察初创企业时，AI原生人才的重要性究竟有多大？我的意思是，可以想象一个光谱，一端是你只需要带来聪明的市场策略、自己的评估标准，以及对某个领域真正关键点的深刻理解。嗯。然后你可以带来一些数据，在OpenAI模型上使用RFT（假设指某种技术），这样也能运作得不错。

Yeah. When you look at startups, like how much AI native talent ends up being important? Mean, you could I could imagine a spectrum where you say, look, all you really need to bring to the table is a smart go to market, your own evals, and a really smart, you know, understanding of what actually matters in a domain. Mhmm. And then some data you can bring and use RFT on on, you know, the OpenAI models and and be okay.

显然，光谱的另一端则是那些自己训练模型，或者进行严肃的后期训练和模型适配的人。

Obviously, there's then the other end of the spectrum, which is like, hey, folks that are training their own, you know, or doing some serious post training and and and adapting their models?

我认为这极其重要。因为在OpenAI时，我记得一个类比——那时人们还对AI感到恐惧（可能现在仍有），但紧张气氛已经有所缓和。我这样解释：就像人类在元素周期表中发现了一种新元素。真正懂得如何运用它的人，是那些亲手把玩过、让它渗入指甲缝、真正摸清其可塑性和特性的人。

I think it matters a ton because so at opening up, back when I was there, I remember one of the analogies this is during the time when people are still scared of AI. I feel like maybe people still are, but it's it's it's the temperature's come down a little bit. Right? And the way that I've explained it, the analogy I used was it's as if humanity has developed or discovered a new element in the periodic table of elements. The people who truly know what to do with this are the people who've played with it with their hands and gotten it under their fingernails and and really figured out, like, how malleable is it and what are the properties.

最糟糕的是...我在职期间，我们仍在发掘模型中隐藏的惊人能力。所以我对你问题的回答是：任何创始团队都必须深度探索过这种'新元素'。那些没有这种经历的创始人，见面时就会暴露无遗，这很无趣。因为你必须真正理解这个'秘密团体'（指AI）的潜力。

And worst and and that time that I was there, we were still discovering the wild capabilities that were locked in the model that we just you know? And so that is a huge part of what I what my answer to your question is, I think it's important that any founding team has to have really deeply played with this new element. And the founders that don't haven't are pretty evident when you meet them, and it's just not very interesting. Right? Because you have to really understand what the cabal is capable of.

你需要尝试大量实验。不必是严格意义上的研究者，但最优秀的构建者必定是那些真正了解手中原材料的人——你能从他们指甲缝里看到这种痕迹。

You have to try a bunch of things. You don't have to be a researcher per se, but the best builders will be the ones who really know the raw materials they're working with, and and you gotta see that under their fingernails.

没错。这波浪潮中我最爱的部分就是：模型被开发出来后释放到野外，连公司自己都惊讶于人们通过'指甲缝探索'（指深度实践）发现的种种可能性。

Yeah. This is one of my favorite parts of of of this whole wave is that, you know, I feel like these models get developed and then they're released into the wild, and, know, you the companies themselves are are surprised by all these things that people learn as they get, you know, the the models under their fingernails.

完全同意。我亲历过这种体验——我常说：我们仍在发掘这些模型的内在能力。这就是现在创业的黄金时期，你可以深入探索模型在特定场景下的真实潜力。

Totally. Yeah. And, like, well, for me, I remember, you know, having this experience firsthand of just, you know and I say this a lot. It's like, feel like we still are discovering the capability, the capabilities that are are are within these models. And that's why it's such a great time to build a company now is that you can actually go and dive deep and see what is the model really capable of in this context.

对吧？当给予这类评估时，当输入这类数据集时...

Yeah. Right? When given this a little bit of this type of eval, when given this, these set of datas.

是的。我们在发现这些能力的同时，模型本身也在不断进化，这...

Yeah. We're discovering all these capabilities while simultaneously the models are getting better, which

简直疯狂。

is crazy.

确实。就像，这两条线索中任意一条能实现都够疯狂的。确实。而且它们同时发生。同时。

Exactly. Like, either one of those threads would be crazy to pull off. Exactly. And the fact that they're both happening simultaneously. Simultaneous.

是的。真的，真的很有趣。

Yeah. Is really, is really interesting.

正是如此。

Precisely.

我经常思考的问题是，你有内部视角——基础模型公司在开发这些产品时有多大优势，因为你知道，构建这些通用模型的团队和产品开发团队就坐在一起。

I think the question I think about a lot, and you have this inside insight is the extent to which the foundation model companies have an advantage in building these products because, you know, the folks that build these models everybody uses sit next to the folks that build product.

百分之百。

100%.

一方面，我持两种看法：这似乎应该有帮助。但当你听说像ChatGPT或ClaudeCode这样的产品诞生过程时，有时感觉就像是组织里有人在捣鼓些谁都能做的东西。他们不一定非要在OpenAI或Anthropic才能开发这类产品。所以我在这两种观点间摇摆，但很想听听你的看法。

And, you know, on the on the one hand, I'm of two minds where it seems like that should be helpful. And then you hear the stories about how things like ChatGPT or ClaudeCode emerged, and they seem sometimes like it's just there were people in these organizations that were just hacking on things that anybody could have done. Like, they didn't necessarily have to live in OpenAI or Anthropic to build some of these products. And so I oscillate back and forth, but I'm curious, what your take is on this.

听着。我认为优势是确实存在的。毫无疑问。但总会有利弊两面。优势在于，当你与那些提出新方法、实现突破的研究人员并肩工作时，你肯定能比其他人更快发现模型在产品场景中的潜力。

Look. I think there's definitely an advantage. Make no no doubt about it. There's always advantages and disadvantages, I would say. The advantage is that if you when you sit next to the partners who are coming up with novel methods and and and making these breakthroughs, there's no doubt that you're able to discover what the model's capable of in a product context faster than anyone else.

所以这是纯粹的优势。

So that's that's a pure advantage.

这仅仅是早期接触的优势吗？还是说具体有什么特别之处？因为很多时候当这些模型发布后，一周内你和研究员交流就会发现：哇，我们真惊讶这么多人发现了这么多我们完全不知道的功能。

Is that just like early access though? Or or like what what about it specifically? Because it does feel like a lot of times when these models come out, it's like, you know, within a week, you talk to the researcher, like, woah, we're really surprised that like all these people have discovered all these these things we had no idea.

我认为部分——很大部分确实是早期接触优势，但还包括能够通过后期训练团队调整模型行为，使其最有利于产品表现。OpenAI、Anthropic和Meta AI都有杰出的研究人员，他们能够将应用端的目标融入模型，确保产品在这些方面真正发光发热。所以这不仅仅是早期接触的问题。

I mean, part of it a lot of it is, I would say, early access, but it's also being able to, like, tweak the behavior of the models to do the thing that the product is most beneficial to the product with with the post training team. And and there's there are fantastic researchers at OpenAI and Anthropic and Meta AI, and so they're gonna be able to apply some of those goals that come from the, you know, the applied side and really make sure that the product can and the model can really shine in those in those ways. So it's more than just early access.

你认为随着RFT等技术的普及，公司能从中获益吗？是的。

Do you think with the availability of RFT and other things, like, companies get some of this? Yeah.

百分之百。这也正是优势所在——当你在这些模型公司工作时，总会面临一种选择：要么追求通用智能（这很棒），成为所有人的智慧之源；但这样你就没时间深入钻研。而这正是初创企业的机会所在——当你对某个行业有独特洞见时，创始人需要亲力亲为才能真正推动产品落地。关键在于他们能百分百专注于特定用例或市场洞察，行动速度可以快得惊人。

100%. And then that and then here lies the advantage too because when you're at one of these model companies, there's always a bit of, like, you know, there's you're either shooting for general intelligence, which is great, and then you become the smartest thing for everyone. But where you really don't have time to dive deep, and this is where the opportunity for startups is, is when you have that unique insight of, like, this industry has founders have to apply a bit of elbow grease Yeah. To really get a product in a real business going. And the fact is they can dedicate a 100% of their time in a certain use case or some insight they see from the market, and they can move incredibly fast.

我相信他们总能打造出针对特定问题的定制化产品。对吧？这在你投资过的项目中已经有所体现，在某些深耕垂直领域的应用公司身上也能看到——这些是用ChatGPT无法实现的。

They're always gonna be able to build a specific product, I believe Yeah. That is tailored to that problem. Right? And you're seeing this in some of the investments you've made and, you know, some of the investments as well. You see this in in some of those applied companies that really are going deep, and and it's stuff that you can't do with ShaqGBT.

要知道，整个工作流程都是支离破碎的。

You know, the workflow is completely broken.

工作流程确实如此。而且正如你所说，越来越明显的是：针对你关心的评估指标，存在需要定制化后训练和微调的环节。

The workflow is certainly and then, you know, and and to your point, it also seems increasingly clear, you know, there's an aspect of kind of post training and fine tuning that is bespoke to the evals you care about.

百分之百同意。

A 100%.

这种组合很关键。早期感觉只要解决工作流程就行——模型就是模型，那样可能就够了。但现在看来，相比通用型平台，差异化机会反而更大了。

You know, that that combination. Because in the in the early days, it felt like it was just gonna be workflow. I'm like, the models were gonna be the models, and you were just like and then and then that was probably enough, but it feels like there's now even more of an opportunity to differentiate versus the general horizontal players. Yeah.

正如你所说，当目标更聚焦时行动就能更快。这就是初创企业的优势。模型公司有它们的优势，初创企业也有自己的优势，价值创造空间很大——这是个伟大的时代。

And you can move faster when you have a more focused view of what is it that you care about, to your point. So I think that's the advantage that startups have. There's a big advantage that the model companies have, and there's a big advantage that the startups have, and there's a lot of value to be created. So it's a great time to be alive.

有没有哪些领域是你绝对不愿碰的？感觉注定会被模型公司垄断或解决的？

Are there any spaces that you like won't touch just you're like, god, that feels like the model companies will will will own or figure out?

可能是我接触的创始人样本偏差，但我觉得创始人们都很清醒。目前还没见过哪个项目让我觉得完全是在OpenAI或Anthropic的主赛道上。

Maybe it's just the bias of just the the the founders that I've met, but I feel like the founders get it. I I haven't seen any pitches that really feel like they're they're right down the fairway for for OpenAI or Anthropic.

确实。显然，在消费者端或编程领域，你会看到人们针锋相对，感觉就像...有些人采用的思维模式是：如果这是一个万亿美元级别的支出领域，那么模型公司最终可能会关注。低于这个规模，就很难想象它们会感兴趣了。

Yeah. I mean, obviously, you you do have people going head to head on the consumer side or on the coding side, and it feels like Yeah. Some mental model people use is just if it's you know, if it's a trillion dollar plus bucket of spend, then maybe it's something that the model companies will care about ultimately. Lower than that, it's it's hard to imagine them.

这个说法太棒了——事实正是如此。因为在那些大公司里，你的边际努力价值巨大，能让许多'船只'随之起航。当然这些是需要特殊支架才能浮起的贵重船只，这正是初创企业的用武之地。这些可都不是小市场。

It's that's I love the way you put it. That's pretty much what it is, because when you're at one of those companies, your your incremental effort is so valuable because you can make so many, you know, boats rise that it's, again, they're very valuable boats that need special scaffolding to rise, and that's where the startups come in. And those are not small markets. Those are like No.

太不可思议了。20亿美元的市场可能都太小了...

It's amazing. Yeah. $2,000,000,000 market may be too small for

没错，正是如此。

Yeah. Exactly. Exactly.

你显然在多家机构负责过产品工作。

You obviously have run product at a bunch of organizations.

当然。

Sure.

我很好奇，我们刚才稍微提到了评估标准，但随着越来越多公司转型为AI优先组织，你认为产品经理角色和产品流程会如何演变？

And I'm I'm curious, we've alluded to a little bit around evals, but how do you think about how the PM role will change and product orders will change, you know, as as more and more companies become these, like, AI first organizations?

我认为任何角色的核心都在于你如何分配脑力资源。如果把大脑比作GPU，关键是你把这些'GPU'指向哪里？我们...

Yeah. I think a lot of for any role, it's about where you spend your mental power, horsepower. Right? So if you think of our brains as GPUs, like, where do you where do you point those GPUs? And We

和所有人一样拥有有限的计算预算，必须决定如何分配

have a limited compute budget like everyone else, and we have to figure out where to

有限的计算预算。每天真正高效的时间可能就八小时，之后就需要充电。但归根结底，这就是为什么我认为专业分工仍会存在——即使代码能从光标自动流淌，仍需要有人专注代码工艺及其抽象架构。随着时间推移，每个从业者都在不断提升自己的抽象层级。

Limited compute budget. There are eight hours that you're really sharp in the day maybe, and then you start getting you're waiting a little bit. You gotta recharge. But, I think that at the end of the day, this is why I think role specialization will still exist because someone who needs to really pay attention and care about the craft of the code and how it's abstracted, even if code is dripping from your fingertips from cursor, you need to think about the architecture and really and and put some deep thought into that. Now there will be a day when actually, day that goes by, I feel like every human being who's working is, like, increasing the level of abstraction that they're operating at.

但毫无疑问，总会有人真正需要主导并感受到自己是工程和设计领域的直接负责人（DRI），他们倡导设计侧的简洁与工艺，以及工程侧的稳健性。这些核心要素不会消失。对产品经理而言，我认为很大程度上是在脑海中模拟：我的目标愿景是什么？下一步是什么？我们如何达成？

But make no doubt, there's gonna be somebody who really needs to really own and feel like they're the the DRI on the engineering, on the design, who's advocating for, you know, on design side simplicity and craft and on the engineering side robustness. Those things won't go away. And I think that, for the PM, for example, I think a lot of it is just the mental simulation of where what is the vision of where I wanna go? What is the next step? How do I how do we get there?

我认为ChatGPT已成为产品经理的最佳伙伴。真的，数不清有多少个深夜——实际上很多个深夜，比如凌晨1点，家人都已入睡时——

And I think that ChatGPT has been the PM's best friend. Yeah. I can't tell you how many times I've been working late at night. Maybe it's, like, actually many late nights. One 1AM, everyone in my house is asleep.

我独自坐着，那种孤独感...这时候Slack上也没人了，但我有个思维伙伴就在身边。我可以随时抛想法给它，虽然凌晨1点我的大脑已经不灵光，但这个伙伴永远不知疲倦，能完全跟上我的节奏，帮我推敲规格文档或愿景文件里那个始终表达不到位的关键词——这种助力极其强大。所以每个岗位都将获得巨大赋能。

I'm just sitting there. It's lonely. You know, maybe people are off Slack at that point, but I have a thought partner right there. And I can just bounce ideas off of, and my brain's already not very, you know, fresh at 1AM, but I have someone who's, like, never tired and just there with me and just able to kind of meet me where I am and just help me crack that one last word of the of of the speck or the vision doc that's, like, really is trying to get at what I'm trying to to to to say, that is really powerful. So I think that every role will have so much leverage.

每个人都会获得巨大助力，对产品经理而言，拥有全天候的思维伙伴将是重要改变。

Every human's gonna have so much leverage, and for PMs, you're just gonna get a round the clock thought partner is is a big part of it.

这是否意味着长期来看，你认为产品经理数量会增减？组织结构会如何实际演变？

Does that mean over time you think we have more PMs, less PMs? Like, how do how do actually org structures change?

我无法预测未来。角色定义总在变化，很可能会出现一种新型态——就像我们看到的独立创业者那样，他们清楚要打造什么，拥有设计、编码、确保运营的全套工具。未来可能不再区分'产品经理'或'工程师'，而是出现混合角色，单人就能完成功能开发。

I I can't I don't know if I could predict that future. Roles change all the time and just role descriptions. It's very possible that there's an archetype that comes out, which is essentially a solo I mean, we've seen this solopreneurs who are have a sense of exactly what they wanna build and have, you know, all of the leverage tools to design it, to build it, to code it, to make sure that the the operations are successful, everything's up and running. And it's possible that, you know, we don't even call it PM or engineer anymore. You might just have a hybrid role, right, where you could have one person developing a feature by themselves.

我认为这就是未来，无论这个人来自技术背景还是设计背景。实际上，我在OpenAI合作过的一位优秀工程师——后来才知道他读过设计学院——这种跨界特质很棒。

And I think that's gonna be the future, whether that person comes from a technical background or comes from a, you know, a design background. It doesn't really matter. Actually, it was one of the very talented engineers that I really enjoyed working with at at OpenAI. I I didn't realize this, but he went to design school. I was just like he has, like, just, you know, looking and catching up with him.

这些技术赋予的超能力之美在于：你可以完整实现任何愿景，用AI补足技能短板。所以'产品经理'这个头衔可能消失，职能或许以其他形式存在。

Was like, wow. That's great. Right? And but that's what's beautiful about all the superpowers that we have in with all these technologies is that you can really kind of, you know, fully fulfill your vision, whatever that vision is, and you can use AI to supplement and complement your skills. So I think that the role PM may not exist and call be called a PM.

这个岗位可能变得更技术化，也可能不会，我们拭目以待。产品经理的职能浪潮已有迹可循——比如在苹果公司，产品经理更偏向市场营销；

It may still exist, but it just might get more technical or not, but we'll see how it evolves. I think even the you can see the waves, of of PMs and how it's evolved. Right? So in some companies, the PM is the product manager is more of a marketing person, I would say, like an Apple. Right?

微软也是如此。而在Facebook这样的地方，产品经理是高度技术导向的设计型人才。这个角色的范式在我们见证下二十年里已经多次转变，完全有理由相信它会继续演化。

And Microsoft even. And then at a place like Facebook where we where we built it, the PM was a very technical design oriented person. So I think the archetypes of PMs have shifted already, and we've seen it in our lifetimes, right, and and in a in a span of two decades. So there's ample reason to believe that it will shift again.

我想转向讨论现有的不同交互模式，显然你直接参与过其中一些项目并进行了跨领域投资。我想先从语音开始，因为你在OpenAI期间投入了大量精力研究语音技术。你如何看待语音在消费者与AI广泛交互方式中的定位？人们都在猜测，是否不久的将来我们都会在办公室里对着电脑说话？你对此有何预见？

I guess shifting gears to some of the different modalities that exist out there, and obviously you've worked on some of these directly and and and invest across them. I'd love to just start with voice, because I think you you've spent a lot of time working on voice when you're at OpenAI. How do you envision, you know, voice fitting into the broader ways consumers interact with AI over I think people are wondering, are we all gonna just be talking to our computers, you know, in an office in in in, you know, near months, or how do you envision it?

是的。很高兴你提到语音。这个水瓶上贴着原始语音代码贴纸——我超爱这个设计。说实话我之前没意识到，但我一直随身带着这个水瓶。

Yeah. I mean, I'm glad you mentioned voice. So this this water bottle, I put the this is the original voice code sticker Love that. Launch. And, yeah, I I I didn't realize that, but I I carry this water bottle with me.

这是我最喜欢的物品之一，简洁又实用。我热爱这些新兴交互模式，因为作为在大学和研究生时期深入研究人类社群互动，并在早期产品生涯中延续这一方向的人，我特别珍视人机交互方式——每种模式都有其独特优劣。语音解锁的潜力尤其令人兴奋：首先，如果你直接转录这个播客，会发现我们有多少次中断又重启思路，因为语言表达时大脑是同步组织思维的。

It's one of my favorite ones, but just just kind of nice and simple. I I love these new these different modalities because with as someone who, like, kinda did a bunch of human community interaction as a as a college student and grad student and kind of evolved into that in my early product career and really valued the way that we interact with computers, each modality has its own, you know, kind of pros and cons. And it is really interesting to have voice be an unlock for so many reasons. Number one, I think voice is I mean, if you if you take a look at just this podcast and you do a straight up transcript, you will see how many times you and I both stop start and stop on a thought. Because when you form words, I feel like the brain just is forming the thoughts at the same time.

这种零延迟的思维-语言同步，为AI交互开启了巨大可能性。语音模型在音调情感表现上日益精进，Sesame团队Brendan他们的成果就非常惊艳——越来越接近自然人类交流。

There's no barrier. There's no lag time, and you tend to think and talk and think and talk at the exact same time. So that opens up a huge possibility for when you're interacting with AI. The model the voice models are getting better at tonality and sort of warmth, and it's becoming more and more natural. If you take a look at the stuff that's that Brendan and team are doing at Sesame, it's fantastic.

简直像魔法对吧？试玩他们的演示时，我和系统对话流畅得令人惊叹，那种精妙设计感扑面而来。

It is magical. Right? And you try the demo. And I was talking to it. It just it just there's a about it of it's just so well crafted.

这让你体会到这个媒介在情感表达上的细腻程度——就像专业演员需要用特定声调传递情绪那样强大。在实际应用层面，语音正被整合到许多前所未有的工作流程中。

And that gives you a a sense of how nuanced this medium is just from a pure model capabilities perspective of emotion. Right? If you think of somebody who's trained as a as an actor or actress, like, they have to be able to inflect their voice in certain ways to communicate certain emotions. It's really powerful. And on the level of on the applied level, it's being, you know, brought into so many different workflows that could not be done before.

语音技术真正有趣的问题在于哪些应用会持续存在。虽然我们看到语音助手兴起——比如我的语音助手可能呼叫你的语音助手——但如果双方都使用语音代理，最终可能会出现更高效的信息交换方式。

I do think the interesting question with voice is what will persist. Right? Because I think we're we're seeing these voice agents emerge, and it's like, okay. My voice agent calls. Maybe they'll call your voice agent, and, you know, over time, if if there's voice agents on both sides, there probably ends up being more effective ways to communicate information back and forth.

所以我一直在思考，长远来看哪些语音用例最具实际意义。一方面，语音交互比文字往来更具吸引力；另一方面，相比对着空白文本框组织语句，实时语音能让你边思考边表达，就像现在这样——我也不知道自己要说什么，但就是即兴发挥...

And so I do wonder, you know, over time, which voice use cases actually make the most sense via voice, and I love this idea of, one, it's just an incredibly engaging way to It's way more engaging than going in text back and forth in a lot of ways, and then two, this idea of rather than staring at the blank box, formulating your thought, and then putting the thought in, you can kind of in real time, you know, like, don't know where I'm what I'm thinking, but I'm kinda just ripping and Yeah.

随性表达

Rub it.

随性表达

Rub it.

实际上，我在开车时特别喜欢用Granola（一种语音笔记应用），因为它能让我在构思投资主题时，尝试各种想法。最终它会帮我整理出一个半结构化的思考框架——就像有个智能助手在梳理我的思绪，帮我识别驾驶过程中闪现的哪些洞见真正有价值。虽然不确定哪些想法会留存，但语音作为人类长久以来的沟通媒介，我认为它必将扮演重要角色。

And one of things actually love to do on the drive is use granola because what what granola does is that lets me kind of I'm I'm, like, working on an investment thesis, for example, and I'll oh, try this idea, that idea. At the end of it, I get this, like, somewhat structured, you know, kind of it's someone some it's it's an agent that sort of processes it and helps me see my own thoughts and where where where are the insights that were, that came to me at what point in the drive and which ones are really gonna stick. But it is a really interesting medium, and and I I don't know exactly which ones are gonna stick or not, but the fact that voice has stuck around for us as humanity for a long time, I think it's gonna be a pretty important one.

确实。头脑风暴显然仍是激发创意的绝佳方式。我记得你在Lenny的播客里提到过用AI模型革新教育的潜力，尤其是你与自己孩子的互动经历。

Yeah. Mean, brainstorm sessions clearly remain an incredibly effective way of generating new ideas. I mean, one thing I've I've heard you riff on before, think on Lenny's podcast, was just the power of using these models for education, and I think your your experiences with your own kids.

没错。

Yeah.

我很好奇，在这个充满可能性的领域里，最让你兴奋的产品构想是什么？

I'm curious, like, what the most interesting things you've seen around here, and as you think about that, like, whole surface area of potential products, things that get you excited in that area.

教育领域确实令人振奋。现在涌现了许多有趣的初创公司，这些模型展现出的能力很惊人——比如能说'不要直接给答案，引导我思考'。我常和儿子用早期版ChatGPT做长除法练习，说真的，你多久没做长除法了？

Yeah. I think the education part well, first of all, I think that there's a lot of interesting startups out there that are that are being formed now, And it's very clear that this is one of the things that the models are just, again, talking about discovering the capabilities of the models to be able to just tell, like, don't give me the answer. Help me think through it. I used to do that all the time with the early versions of ChatGPT with my son of, like, honestly, there was some long division. When was the long time you did long division?

知道吗？我需要重温运算步骤。与其看五分钟YouTube教程，不如直接问AI：'小查，再解释下这个原理'或者'这道题该怎么拆解？'

You know? And I just gotta get reminded of how to do it. And I can sit there and watch five minutes of a YouTube video, or it could be like, hey, ChaChaBeetie, like, tell me remind me what this thing is again, or, hey. There's a problem here. I'm not sure how to parse it.

比如辅导儿女作业时，我常感到脑力透支。等你也当父母就懂了——五年后你肯定会说'我早忘了这个知识点'，更何况深夜辅导时脑子根本转不动。

The example I gave is, like, you know, sometimes helping my kids, my daughter, and my sons, with the with their with their, homework, and my brain is fried. Yeah. And just to be able to and you're about to be a parent. Like, it's it's it's gonna be you're gonna you know, in five years, you'll be like, I don't remember how to do this. And, also, my brain is like, it's way too late.

这时就需要AI辅助。我认为这将成为未来主流应用场景之一。

I just need an assist. And I think that's where a lot of, that's one of the workflows that I would say would be out there.

完全同意。目前最成功的或许是语言学习类应用，口语练习效果尤其显著。我们采访过可汗学院的Sal，他认为核心挑战在于如何维持用户粘性——如何让更多人持续使用这些产品。

Yeah. Totally. Mean, it feels like you've seen maybe the first, most successful wave has been these, like, language learning apps and the ability to speak, just being so much more effective as a as a way to learn languages. You know, we had Sal from Khan Academy on the podcast, and, you know, what he was saying is obviously the the problem to really be solved is is just the engagement problem. Right?

总会有部分积极用户充分使用产品，但关键是如何扩大这个群体，真正解决用户参与度的问题。

Of how do you get there's some there'll be some set subset of motivated folks that that wanna use these products to the fullest, and then the question is, you know, who how do you, you know, get people engaged to actually use these these products?

是的。我认为这正是AI真正能发挥作用的地方，因为正如我所说，学习很大程度上需要你真正对其感兴趣。这很棒，因为我觉得追随孩子们想学习的热忱是件好事。我和妻子就是这么做的——我有两个女儿对马术特别着迷。

Yeah. I think that's I think actually that's where AI can really help because like I said, so much of learning, you have to be kind of really interested in it. And some you know, this is great because I think it's great when you follow kids' passions of what they want to learn. And, you know, this is something my wife and I do. So two of my daughters were really interested in horses.

我们参加了马术夏令营。幸运的是她们还没痴迷到要买马的地步，所以算是躲过一劫。但跟随学生的兴趣导向，确实是激发好奇心、培养终身学习热情的关键所在。这正是AI的用武之地——不仅是调整题目形式（比如用足球场景设计数学题这种简单技巧），更在于...

We did horseback riding camp. Luckily, they don't love it enough that we have to buy a horse, so I I think that we dodged a bullet there. But, you know, following the lead of where, students wanna go is gonna be really the way that you kinda unlock that curiosity and and teach that lifelong love of learning, I would say. And that's where AI can help. Not just adapting something, a a problem, you know, you know, if the kid likes soccer, being able to frame a math problem in a in a in a soccer setting, that's one very simple tactical thing, but I think you're right.

Sal说得对，如何吸引学习者是个长期难题。我特别好奇是否能让AI智能体主动踏出第一步。在OpenAI时最让我惊艳的发现，就是当模型主动开启对话时的效果——传统使用ChatGPT时总是用户先说，但通过特殊提示词可以让AI先开口。

Sal Sal's right that there's a problem that existed how you draw people in. And I'm really curious if there's a way that we can make these agents or AI kind of, you know, take the first step and and talk first. That was one of my favorite discoveries of of what the models could do at OpenAI was what happens when the model talks first. Because Say more about that. Well, because, you know, traditionally, when we first when we use Charge dbt, it's like I say something and it says something back.

这彻底改变了交互范式：不再是随叫随应的助手，而是能主动作为的伙伴。我非常期待看到这个方向的发展前景。

But there are ways that you can prompt the model to, like, talk first. And just like and so this is no longer you know, the analogy is it's no longer like an assistant who's at your beck and call, but really, what if it could be proactive? And I'm really excited to see a lot more of that arc play out.

目前有在产品中看到这种实践吗？

Have you seen that play out in in in products to date? I mean

当然。比如Yutori这家公司就在做主动预警系统，但这只是开始。未来会有更多类似探索，这确实是片待开发的蓝海。

Oh, yeah. So we're in a company called Yutori, and they're they're proactively going out there and, like, you know, alerting you of of of of things that are happening out there. That's just the beginning. And I think there's gonna be a lot of stuff that happens along those lines, but I think those are that's another frontier that's yet to be discovered. Totally.

现在很多产品体验就像给个空白输入框任你发挥。10%的深度用户能玩出花样，但可能90%的人会感到困惑。所以向主动式体验转型确实很有意思。

It does feel like a lot of the product experiences today are like, here's a blank box. Do with it whatever you want. Totally. And, you know, 10% of people get so much value out of that because they know what to do, and from maybe 90% of people, it can I don't know about the exact percents, but it's confusing? And so I do think that that shift to something proactive is really interesting.

相比你刚加入OpenAI时，现在AI模型在消费者和企业端的普及度是否符合你当初预期？

Yeah. Relative to when you started at OpenAI, how was the diffusion of these AI models into, you know, both consumers and enterprises compared to to what you might have expected?

很有趣的是入职后他们告诉我：'招你就是看中你兼具toC和toB经验'。这很合理，因为本质是生产力工具。所以我们初期重点开发企业版ChatGPT——企业需要安全审计、系统集成等专属功能。

It's funny because when I joined OpenAI, after I joined, they told me, oh, we hired you because you've done both consumer and enterprise, and we needed someone who did that. And I was like, oh, that's really interesting because it is a work product. It's it's a product for work. It's a productivity tool. And so I think, you know, that's probably the reason why when we joined, it was it was it was the consensus that we gotta build Chairobi for enterprise because they're just unique features that are needed for enterprises where you have security, you have integrations, you have audit logs.

组织级应用必须构建隐私防护等基础架构，现在企业采用率才刚赶上。而在消费者端，我入职时大众对AI普遍存在恐惧。记得我们第一位市场总监Coley就面临这种认知挑战——她做得非常出色。

There's, like, a bunch of stuff that's expected that, you know, the team's still building today. But in in the fact that you're working with an organization and, like, there should be some privacy within the organization and, and not have things leak out, it's not a surprise that, for all of these companies, the adoption is just now catching up because you kinda had to build some of those basic features and the basic constructs and mental models that this is a work product, that will make it so that it can be adopted by by by enterprises. I think on the consumer side, again, when I first joined, people were scared of of AI, predominantly. That was the the the the consensus. And I remember when we made our first product marketing hire, Coley, she's fantastic.

我记得最后一次和她通话时，我说，那么，你有什么问题吗？她主动回答说，是的。我花了很多时间打电话给中西部的用户，或者只是我在中西部认识的人，想听听他们的看法，因为那时候，这简直是闻所未闻。如果你不在硅谷，然后你说，哦，你对AI很看好，那么普遍的共识是，这非常可怕。但这正是总会发生的事情，一旦人们开始使用这项技术，我认为一些实用性会让部分恐惧消散，而一些实用性会变得非常明显。

And I remember when I sat down for the final call with her, I was like, well, so, you know, do you have any questions? And she just said proactively, she's like, yeah. I spent a bunch of time calling users in the Midwest or just people that I know in the Midwest to just get their take because that's, at the time, it's like that was unheard of. If you're not in Silicon Valley and you're like, oh, you're bullish on AI, then the the general consensus was, this is very scary. But this exactly what always happens happen, which is once people start playing with the technology, I think that some of the usefulness some of the fear dissipates, and some of the, usefulness becomes really evident.

我认为这就是过去十八个月里发生的事情。

And I think that's what's happened in the last eighteen months.

你知道，显然，过去几个月的一个发展是这些不断升级的研究人才争夺战，

You know, obviously, one of the developments in the last few months has been these escalating research talent wars,

而且

and it

感觉确实如此。其中一些薪水已经相当高了，人们来回跳槽。你认为这最终会如何影响更广泛的生态系统？

feels like there's yeah. Some of these salaries are are are getting quite high and folks moving, back and forth. How do you think that ends up impacting the broader ecosystem?

是的。我认为这肯定有一些影响，首先，我相信市场，因为很明显，对这些大型实验室来说，正确获取研究人才非常重要。而且，确保他们能够留住并招募彼此的人才对每个实验室来说都绝对至关重要。我认为最终受益的是其他人，比如我们作为投资者，以及其他初创公司的创始人，因为推动通用智能和超级智能的竞争对所有人都有好处。我认为这对人类来说是好事。

Yeah. I think there's some definitely, it's first of all, I trust the market because it's very clear that this, you know, research talent is is so important to get right for each of these big labs. And, it is absolutely imperative for each of these labs to make sure they retain and recruit each other's talent. I think who ends up winning is everyone else, like us as investors and other, and, startup founders primarily because, you know, the competition to push towards general intelligence, superintelligence is is great for everyone. I think it's great for humanity.

我认为它确实有一些二阶效应，你知道，有些人不是研究人员，但他们觉得自己对产品或公司或消费者的价值回报做出了很多贡献，但薪水、RSU等方面的差距越来越大。所以这是一个二阶效应，我相信人力资源部门最终必须解决这个问题。

I think it does have second order effects that are just you know, there are some people who are not researchers who feel like they're contributing a bunch to the product or the company or consumer return on value, but the discrepancy in salaries and and RSUs and all that is becoming wider and wider. So that is a secondary effect that I'm sure folks will have to address HR departments will have to address at some point.

你认为这会蔓延到其他职位吗？我和一些创始人交谈时感到惊讶，他们说，你知道，我戴上扎克伯格的帽子，我在想，也许我可以引进这个人，嗯，你知道，付给他们比平时高得多的薪水。

Do you think that will trickle into other roles? I've been struck in talking to some founders where they're like, you know, I put my Zuckerberg hat on, and I wondered, like, maybe I could bring in this person and and Mhmm. You know, pay them way more than than I might have otherwise.

是的。我前几天看到一个招聘信息，你知道，我们想要一个研究工程师，这是我们支付的薪水，他们在Twitter（X）上公开了这一点。抱歉，旧习难改。你知道，评论区里有一些争论，比如，嘿。

Yeah. I saw a job listing, you know, just, the other day of, you know, and we want a research engineer, and this is the salary we're paying, and, you know, they were very public about it on on Twitter on X. Sorry. Old habits die hard. And, know, you there's a little bit of debate in the in the comments like, hey.

你们付这么多钱？是的，就是这样。我认为市场总会重新调整。现在，这是一个需求很高的职位，而且理应如此，但在某个时候，可能会有其他职位需求增加。

Like, you're paying this much? It's like, yeah. That's so I think the market's always gonna reset. And, right now, that's that is the role that is in high demand and as it should be, but at some point, there might be some other some others.

是的。我的意思是，这也要回到我们之前关于定价的讨论，我认为这是许多这类代理公司的牛市情景，存在一个悬而未决的问题：如果你在替代劳动力，但大家做的事情又差不多，你们会不会互相竞争压价？这个替代劳动力预算的梦想，是否实际上高估了其中一些市场的规模？但我越来越觉得，如果你有某种东西能在某个角色上全面显著更优，你就能保持一定的定价权。

Yeah. Mean, in my mind, this is also to to bring it all the way back to our conversation on pricing, this is like the bull case for a lot of these agent companies where I think there's there's this outstanding question of, you know, if you're replacing labor, but you're all doing something kind of similar, do you just compete each other down? And is is this dream of replacing labor budgets actually just way overstating the size of some of these markets? But I think more and more, you're seeing, hey, if if you have something that is meaningfully better across the board for a role, then you're able to kind of maintain some serious pricing power.

但我也认为，无论个人或公司现在拥有怎样的定价权，他们都需要将其建模推演两三个层级，看看需要做什么，那些公司需要做什么来保持竞争力，并找出下一个合适的抽象层级会是什么。

But I also think whatever pricing power people have right or companies have right now, they just have to model that out, you know, two or three more degrees and see what do they need to do, what do those companies need to do to to stay competitive and to kind of figure out what the right enough next level of abstraction is gonna be?

具体有什么例子是你想到的吗？

What like, is there a specific example of that that, that you think of?

你知道，很多公司都进行了转型和改变，我脑海里总有个跳板的比喻。每个人、每家公司都有自己真正擅长的东西，他们用这个作为跳板跃升到下一个层级，然后又有新的层级，再跃升到下一个目标。我认为Facebook就是个很好的例子，马克的野心极大。

You know, so many companies have just made pivots and changes, and I always have this analogy of the springboard in my mind. Like, each person, each company has something they're really good at, and they use that to springboard to the next level. And then they have another level, and they're gonna springboard to the next thing. I think a great example is, you know, I would say Facebook. Like, Mark is extremely ambitious.

我对他怀有极大的敬意。它最初是个大学网站，后来扩展到包括高中生在内的年轻人，再后来成为全世界不同年龄层人们的沟通方式，接着又成为广告商触达目标受众的主要渠道。现在他们正通过硬件团队构建下一个平台，同时还在打造超级智能。

I have super huge amount of respect for him. And it started off as a college site, but then it became a, you know, young person, including high school, and then it became, you know, you know, the way that people in the world, regardless of your age, communicate. And then it became a lot of the way that, advertisers are able to reach the right audiences. And then, you know, now they're building the next platform with with the hardware team. At the same time, they're building superintelligence.

所以我认为这是个绝佳案例，展示了一位非凡的创始人如何预见未来趋势，并在为时已晚之前就开始押注下一个领域。

So I think it's a great example of, like, a an incredible founder, that is able to see around the corner and start making the bets in the next thing before, it's too late.

是啊。你觉得Facebook的AI战略会走向何方？

Yeah. Where do you think this Facebook AI strategy is going?

我不太确定，因为我不在内部。虽然我有朋友在那儿，但我们不讨论具体战略。不过据我了解，当马克说要打造个人超级智能时，我相信他。每个人都需要有个助手来协助日常生活，如果你能为每个人创建个性化的智能体，帮助实现他们的目标，那将极具价值。

You know, I I can't say for sure because I'm not there. I don't only talk to people there about the specific I have friends there, but I don't talk to them about the specific strategy. But from what I gather, it makes sense because when Mark says personal superintelligence, I believe it. I mean, there's there's everyone needs some somebody or something to assist them in their everyday lives. And, you know, if you are able to create that, the agent, if you will, for each person that's personalized to them and able to do meaningful things and to accomplish their goals, that is really, really valuable.

与此同时，很多初创公司也在尝试实现其中的某些部分，这将是一场竞赛。

At the same time, there are a lot of startups who are trying to do pieces of that, and it's gonna be a bit of a race.

这不就和ChatGPT一样吗？

Is that not the same thing as ChatGPT?

我认为，嗯，我快速分享一个观点：'智能体'这个词现在被过度使用了。准确说是过度炒作，因为目前这些技术所做的一切并不真正具备代理能力。它们或许能进行稍长时间的思考，但缺乏在现实世界中行动的工具。它们能思考、交谈、观察并辅助你思考，但还无法替你预订餐厅或调整日程。所以我认为大家仍在探索阶段。

I think that well, I I'll I'll give you one quick hot take is that I think that the word agent is really over Right. Overhyped, I would say, because nothing that the stuff that things do today are actually agentic. They are maybe they can go think for a little bit longer, but they don't have all the tools to go out in the world and do things. So they can think, they can talk, they can see, and they can help you think, but they haven't yet had a little tools to be able to make that reservation for you quite yet or rearrange your schedule. So I think everyone's still Right.

都在朝这个方向努力，但没人真正实现。OpenAI和ChatGPT团队也不会声称已达成目标，但这确实是终极追求，是所有人前进的方向。

Trying for it. I don't think that anyone's accomplished it. I don't think anyone at OpenAI and Chatuchiki would say that they're there yet either, but that is the holy grail. That's where everyone's going.

这引出另一个问题：谷歌、Facebook、OpenAI是在争夺同一个消费者市场，还是存在多个消费产品的空间？未来我们会只有一个日常消费级智能体吗？

I guess that's another one. Like, are Google, Facebook, OpenAI all competing over the same consumer prize, or do you think there's room for, like, multiple consumer products? Are we gonna just have, like, one, you know, day to day consumer agent?

非常好的问题。我认为不会只有一个。人类需求如此多元，就像我同时使用Facebook和Instagram，不可能只有一个平台。

Great great question. I don't think there's gonna be one. I think we as humans demand so much variety and, like, you know, that that, you know, there's not they're not just one. Like, I use Facebook and Instagram. Like, there's there's not one.

对吧？产品设计中的框架设定很重要——如何进入体验、用户预期、使用场景心态，这些细微差别才是关键。

Right? And, yeah, I think that the framing matters. This is where product matters. Like, how you actually enter into the experience, what you expect, what's the mindset that you're in. These are all the nuances.

iPhone上从不会只有一个应用，总是百花齐放。

They're not there's not one app for the iPhone. There's so many.

嗯。对这些产品可能的细分方向有什么初步想法吗？

Yeah. Any early thoughts on, like, the ways they might break down?

我认为大公司都会充分利用自身渠道优势——如果他们足够明智就应该这么做。执行得当的话，谷歌通过Chrome浏览器，Facebook通过信息流和Instagram等平台，都能产生巨大影响。每家巨头都有独特的跳板来切入智能辅助领域。

I think all the big companies are gonna really leverage their distribution points, And rather, if they're if they're smart, they should. And if they execute well, they can they can have a ton of impact. You know, Google with Chrome, for example, is is pretty big. Facebook with the feed or with, Instagram, etcetera, is is pretty is is pretty big. So I think every big company has a certain spin spin board they can use to kind of jump into that really assistive thing.

我感觉最终格局会与现有市场划分类似。比如你会...

And I I feel like it's gonna it's gonna roughly be the same division, if you will. Like, you go So

我觉得你所说的待完成事项...

I feel like the jobs to be done that you're

今天使用这些工具是个好主意。我认为这将是一个'待完成工作'导向的事情，一个工作流程问题，一个独特的数据库问题，这将会非常重要。

using these tools are for today. It's a good point. I think it's gonna be a jobs to be done thing. I think it's gonna be a workflow thing, a jobs to be done thing. It's gonna be a unique data thing, which is gonna be really important.

对吧？这也是个定位问题。比如我用Perplexity觉得很好，但同时也用谷歌搜索，有时Gemini的搜索结果也够用。

Right? And it's an orientation thing. Like, I I mean, I use Perplexity. I think it's great. It's I still also use Google, and sometimes the Gemini search is, like, good enough.

对吧？我其实不会想太多。对消费者来说选择实在太多了，我相信企业会找到差异化方法。如果够聪明，它们会利用自身渠道优势。

Right? And I don't really think about it. So there's just so many different choices for consumers, I would say, that, I think that the companies will figure out how to differentiate. And if they're smart, they're gonna leverage their distribution.

转到投资角度，想听听你对当下VC圈热议话题的看法——AI应用的防御性。现在大家都爱说'rapper'模式，我很好奇你判断项目可防御性的思维模型，以及什么会让你更兴奋或更谨慎。

I guess shifting gears to the investing side, curious for your take on on, you know, the VVC question of the day, it feels like, is is all about defensibility within these AI apps, and, you know, everyone likes to call them rappers, and, you know, I'm curious for your own mental model of, you know, how you think about what's defensible and not, and and what get to you kinda more excited or less excited.

完全同意。我对这个议题有明确观点——虽然可能出错——但我觉得...

Yeah. Totally. I'm I feel like I have a a clarity of what I believe in this one, and I could be wrong, but, you know, let's

在AI领域必须坚持强观点，但也要保留每月根据海量新信息调整的权利。

I feel in the AI world, you have to have strong opinions, but then reserve the right to change them every month when there's like a plethora of new information.

没错。考虑到现在社会的信息密度，三个月后没人会记得我说过什么。首先声明这不是什么惊人观点：我坚信独特数据至关重要。

Exactly. And and given just the way that the, you know, society is now everyone's gonna forget what I said anyway Right. In in three months. The first thing is, don't think any of this stuff is is a hot take. I think people have said this before, but I deeply believe that unique data is super important.

有大量数据尚未被网络公开，模型也未经训练。举例来说，假设你在一家健康公司负责排班调度...

There's so much data that is not out in the web that the models have not been trained on. Yeah. And I'll give you some examples. Right? So let's say that you are someone who's doing, you know, scheduling for a health company, for example.

我随便举例。现实中不存在人们反复观看排班对话录音来研究语气微妙变化的视频资料。再比如销售cold call时，虽然模型读过相关书籍，但如何把握电话中恰到好处的停顿时机——这类细微数据目前并不存在于模型中。

I'm just making this up. Right? Or you or maybe that's the job that you have is scheduling. There's there are no real YouTube videos out there that that, like, are just like no one watches people, like, go back and forth on a transcript on scheduling and the tonality of voice and and that specific problem. Or let's say that you're cold calling for a sale.

还有企业数据。像麦肯锡、贝恩这些咨询公司内部积累的大量研究资料、思维链条，都是从未公开出版的独特知识资产。

There's you know, there are books you can read that the model's probably read, but, like, how that act the nuances of how you actually know when the pauses are a little bit too long to jump in as a salesperson on the phone, that data does not exist in the world, right now in the model. So I think there's unique data there. There's also enterprise data. Like, there's, you know, those you know, take McKinsey, Robain. Like, they have mounds of, like, research they've done internally and and thought process and chain of thought, if you will, that is not written in any book, really.

那就是，你知道的，模型可以训练的外部数据。所以我的意思是，那是独特的数据。我总喜欢这么说，但我想提醒大家，Facebook曾拥有独特数据。那种独特在于没人真正梳理过你实际的朋友关系网——这些原本只存在于人们的脑海中。

That's that's that's, you know, out there that the models can train on. So that's that's I mean, that's that's unique data. And I I always like to say this, but I wanna remind people that Facebook had unique data. The unique data was nobody really mapped out who you were really friends with. It was all in people's heads.

仅仅通过数字化这些数据，就能创造一家价值22万亿美元的公司？（我现在都数不清零了）这就是我认为一切的起点。其次是数据飞轮效应——如果你建立了正确的工作流程，就能持续优化模型和产品。第三则是你能否快速占领市场？

And just to digitize that, and you have, like, a $22,000,000,000,000 company? I I I lose track now. That is where I think it all starts. And then the second thing is the data flywheel, which, again, if you have the right workflows, you're you have an ability to make sure that the models and your product gets continuously better. And then the third thing is how fast can you get to market and blanket the market?

因为只要实现前两点，再迅速抢占市场，你就能形成滚雪球般的优势，让别人难以追赶。完全正确。

Because with one and two, if you can get out there and just really take over the market very quickly, you will have a compounding advantage that will be very hard for people to catch up to. Totally.

这种数据飞轮显然来自市场实践和规模化部署。当你在客户中全面铺开后，本质上就成为了最新模型能力与具体应用场景之间的翻译层。

I mean, you get this data flywheel obviously from from being in market and Yeah. Exactly. Being deployed scale. And then, you know, you basically earn the right to you're fully deployed in these customers. You kinda end up being this translation layer between whatever the latest model capabilities are and those

其实不止如此。我认为你的微调模型会越来越擅长特定任务。首先，差异化的起点能帮你打开市场，吸引用户签约。

Well, it's actually more than that. I actually think your your your your fine tune model, if you will, gets better at the job that that you're trying to do. Yeah. Right? Because one, if you start with something differentiated, it gets you that distribution and people will sign with you.

当用户使用你的工具时，他们就像微型标注员。比如AI助手处理日程安排时犯错，用户纠正的过程就产生了独特数据——只有深度嵌入企业工作流的工具，才能学习到如何根据实际偏好进行日程安排的推理逻辑。

But as people work with your tool, right, and they're they're they're essentially, you can see them as just mini labelers. Right? And when when the model gets something wrong, let's say that the agent is doing some kind of scheduling task for you, and it does one or two things wrong and you correct it, that is a unique set of data that you only get if you are the entrenched tool within that company to learn the preferences of how to actually do the the scheduling Right. And the reasoning there.

没错。部分优化可以通过更新底层模型实现，但更多时候需要调整模型外部的框架和工作流程。

Right. And some of that obviously is, is, you know, ways that you can update the underlying model, but then some of that is just updating the harness around the model and and the and the workflow.

确实。

And Yeah.

关于专有数据的重要性，我的观点其实反复动摇过。可能比你更保守些——感觉模型需要的训练样本量从来不必特别庞大。比如在50个客户部署后发现边缘案例，更多是更新外围架构而非模型本身。以医疗行业日程管理为例，可能有5-10家公司拥有足够数据，但总觉得单靠数据量本身难以形成差异化。早期人们期望需要数十万样本，但现在看来每个领域最终只有少数公司能...

You know, because I I do think that one thing, know, talking about strong opinions loosely held is just on the the proprietary data side and how much that matters, I feel like I've gone back and forth, but I think I've generally been, you know, maybe a little more bearish than you where it does feel like the number of examples you need to feed these models has never seemed to need to be massive, and in some ways the learnings from like, oh, this weird edge case that happened because I'm deployed at 50 clients now, maybe ends up updating the scaffolding you put around the model or something, but the model itself maybe, if we take the scheduling and healthcare example, there may be five different companies or 10 different companies that have enough data, and they have more than certainly the 30 other people that are starting out, but I always wonder, it feels like not sufficient on its own to differentiate, I feel like in the early days, there was this hope maybe that actually you needed tens or hundreds of thousands of examples, and it turns out, feel like in any space we look at, there ends up being a handful of companies

掌握这点。关键是目前模型还没进化到仅用100个数据点就能微调的程度。你说得对，启动飞轮不一定需要独特数据，但让产品触达客户绝对至关重要。完全同意。

that have that. Again, that's that's the thing is that the models are not good enough that you can just you can fine tune a model on, like, a 100 different data points. Totally. So it's not you're right. You're totally right that the unique data may not be super necessary for you to start that flywheel, But I do think that, you know, getting into the hands of customers is super important Totally.

开始吧。

To to start.

最后往往会说，哦，我们需要，在这个我们共同协调的系统中，额外调用一个模型来检查这种特定场景下才会出现的奇怪现象——这种经验只有在大规模部署这些模型后才能获得。

End up saying like, oh, we need, you know, in this, like, system we're orchestrating together, we need an extra model call that checks this weird thing that happens in, like, this specific context that you would only learn from having deployed these models at at massive scale.

完全同意。完全同意。而且我觉得还有一点就是，当我遇到创业者时，我能分辨出哪些人具备那种坚韧的特质。

Totally. Totally. And yeah. And and I think the other thing I would say is that there's a you know, when we when I meet founders, I can tell which founders have that grit. Yeah.

我称之为必须付出辛勤努力。你必须真正下苦功夫才能让事情运转起来，不能只是随便试一次模型就指望成功。你需要真正理解应用场景并据此推进。我认为这不仅仅是模型能力的问题——记住，模型会学习行为模式。

I call it, like, you gotta put in the elbow grease. You gotta you gotta really have the elbow grease to actually make this thing work, and you can't just, like, you know, one shot the model and and hope it works. You have to really understand the use case there and and go from there. I think that what I'm I'm it's more than just like the the model's ability, but I also think that, remember, the models learn behavior. Yeah.

我想强调的是行为模式会随时间改变。看看消费市场就知道——我们从横屏视频转向了竖屏视频。消费者、工作者和企业需求永远在进化，这种潮流变迁正是我认为飞轮效应重要的原因，与其说它增强了模型能力的防御性...

And I do wanna put out there that behavior shift over time. Right? And if you just look at just the consumer market Yeah. You know, we shifted from looking at, what, you know, landscape videos to vertical videos. So I think there's always gonna be an evolution of what consumers and workers and enterprise demand, and there's gonna be shifting those tides, which is why I think that flywheel is important, less so because it makes it more defensible from a the model capabilities get better.

不如说它能让你跟上节奏，真正捕捉到偏好变化的时刻。这才是魔法发生的地方。

It's more so that you can keep up and really detect when when when preferences change. And when and and that's that's that's kind of where really the magic happens.

我记得Swarthcraft的Jon在播客里说过一句很棒的话：最好的评估方式就是把东西直接扔给用户观察反应。这比在实验室里空想要有效得多。

One thing I think Jon from Swarthcraft said on the podcast a while ago that I loved is like, he's the best he's the best evals are just like, throw these things at your users and see what they do. Yeah. Like, you'll learn way more than standing up in your lab and trying to

没错。你需要有个初始假设，但之后就必须推向市场观察实际效果——这就是飞轮。

figure out. Yeah. Yeah. You have to have an initial hypothesis, but after that, then you just gotta go out to market and see what's what's happening, and that's the flywheel.

你显然在ChatGPT上投入了大量精力。GPT-five的公众反应很有趣——现在这个产品被如此多不同细分群体使用，有些群体热情高涨，有些则表现出不满。这种差异化使用场景的呈现非常耐人寻味。你怎么看？

You obviously spent a lot of time working on ChatGPT. I think with GPT-five, it was fascinating to see the reception, because clearly this product is now used by so many different sub segments of consumers, and so I feel like some of those segments were incredibly fired up about it, and others were frustrated with it in some ways. And it was just a really interesting manifestation of just the so many different people using this for different things. What'd you make of that all?

首先我要说明ChatGibidi是整个团队的成果——我在任时有许多杰出的训练后优化专家、产品专家和设计师共同完成的，我个人不敢居功。但GPT-five的反应完全在我预料之中，这和我在Facebook改造首页时的经历如出一辙——人们总会质问'你们动了我的奶酪'。

Well, first thing I wanna say is that, you know, really, ChatGibidi was done by the entire team when I was there, and so I take zero credit for it versus a lot of brilliant post training, brilliant product minds, brilliant designers who really kinda made it all come together. But what you're seeing with g p d five is not, in my opinion, anything of a surprise because this is the same pattern that I dealt with when I at Facebook when we worked on revamping the homepage, everyone's like, where you moved my cheese.

就像，嗯，其实也不算恶意，动态消息（News Feed）刚推出时反响极差，这是众所周知的。

Like Well, also wasn't mean, News Feed famously, like, incredibly poorly received the first time it was it was launched.

是的。那是第一个版本，而我参与开发了现在基本在线的版本。我们内部称之为'否认'，也就是新的流式架构。这个版本同样很难被接受，因为人们会说'你把这里或那里改动了'，懂吗？

Yeah. And so that was the first version, and I worked on the version that is pretty much live now. We internally, we call it denial, which is the the new stream architecture. And that was really hard to receive too because it's like, oh, you move this or this. You know?

这完全是典型的变革管理。我并非要轻视它，但我认为任何有影响力且拥有忠实用户群体的公司都会经历这个过程。所以我非常理解那些仍在团队中应对这些挑战的朋友们。说实话，看到那些反弹时，我一点都不惊讶。

It's it's it's a classic change management. And and and I don't wanna belittle it, but I just think that every company that's consequential and has a really deep following is gonna come is gonna, go through that. So my, my thoughts are with, the, you know, my my friends who are still on the team there of, like, working through that. But, honestly, I saw the the backlash. Was like, this is not a surprise.

这完全在意料之中，因为正确的方向就是走向一个不需要选择模型的未来——你为什么要选择模型呢？我记得在任时，我们成功精简了模型选择器，离职后却发现它又开始膨胀。现在我很高兴看到他们决定'我们要彻底整合'。

It's exactly what's gonna happen because and that's the right the right direction is to go towards a world in which there's no model picker. Because why why why do you need to pick that model? And I remember being there, we had a really successful time where we cleaned up the model picker, and I remember after I left, was like, wow. That thing just starts growing again. And then I'm glad that we're like, hey.

作为用户，我非常欣赏他们整合的目标，但现在控制欲又让功能开始蔓延。这和我见过的无数产品模式如出一辙：在提供便捷魔法与保留必要控制权之间，始终存在精妙平衡。动态消息也面临同样问题。

We're gonna do it all together. And and as a user, I just love that they they had that goal of of collapsing it, but then here, it's it's creeping back up again because people do wanna control. And this is the same, again, same pattern I've seen in so many different products. There's there's a fine balance between how much, magic you provide to the person so that you remove a bunch of the friction and how much control is really needed and by whom. We had the same problem at News Feed.

用户总说'朋友动态太多'、'这类内容太多'，要求'恢复旧版控制权'。明白吗？

People are like, I've seen too many friend stories. I'm seeing too many of this. And people are like, I used bring it back to the way it was. Like, I had control over it. Right?

推特上也是同样情况。有人不要推荐流只要时间线，每个人对控制权的需求都不同。这将成为产品设计中永恒的命题：如何找到最优雅的解决方案。

Same thing you see on Twitter. It's like, oh, like, I wanna I don't want I want I don't want the for you thing. I just want the the the livestream. Everyone has a different level of control that they want, and this is gonna be an age old question in product design of how do you figure out the most tasteful way of doing that.

我特别想厚着脸皮请教你的看法：在加入OpenAI前你在Uber工作，我很好奇ChatGPT作为消费者门户，与其他应用（比如Uber、DoorDash）的交互会如何发展？ChatGPT团队肯定想整合这些服务来增强代理模型价值，但Uber团队恐怕不愿沦为普通API调用。你从双方角度会怎么思考这个问题？

One thing I I was shamelessly curious to get your take on is obviously you've, you know, before working at OpenAI, you were working at Uber, and and I think one of these things that I'm really curious to see how it plays out is just the interaction between, you know, ChatGPT as this consumer homepage, and then all of these other applications that are out there. Because I can imagine if you're on the ChatGPT product team, being able to integrate Uber, DoorDash, all these other products into better agentic models will just make it that much more valuable to end consumers. Then I'm sure if you're on the product team over at Uber, you're like, please do not put me into some commodity like API call. How does this all go, and how would you think about this on both sides?

你的问题框架非常准确。这本质上是利益驱动的问题——总有公司更愿意参与整合，从而迫使其他公司跟进。归根结底，每家公司都在优化自身业务，最终会进行成本效益分析。世界正朝着即时简化的需求发展，这些公司终将别无选择。

I mean, I I really think that you you you have it framed spot on. It's it's it's just about the incentives, and, you know, there's gonna be some company out there who has more incentive to partner with some of the, you know, integration and do the integrations than others, which will force some of the others to come in. At the end of the day, I think everyone has their own business that they're optimizing for, and it just becomes a, you know, cost benefit, you know, analysis, at some level. The world is going to a world where, we are we are going into a world in which simplicity and instantaneousness is what consumers demand. At some point, there won't be really a choice if you're one of these companies.

你必须提供最优质的核心服务——就像用户选择Uber而非Lyft是看中预估到达时间。关键在于如何让智能代理获取这些优势。说实话，我认为界面终将消失，未来我们会通过耳戴设备直接对话。

You'll just have to provide the best service at what you can provide, which is why would I choose Uber versus Lyft. Right? Are my ETAs lower, etcetera, and how do I make sure that the agent is able to access that? So, honestly, you know, if you look on the video, I think that interfaces are largely gonna go away at some point. You're gonna have these devices in your ear that are gonna you're be able to talk to.

他们将采取主动而非被动应对。观察哪些玩家会撤出哪些领域会非常有趣。但归根结底，我认为这一切都会有所收缩。届时界面将演变为那个语音模型，或是那个设备本身——这也正是为何众多企业争相布局硬件领域。苹果和谷歌在这方面显然具有优势，其他公司也能看到不少硬件创新举措。

They're gonna be proactive, not just not just reactive. And it's gonna be really interesting to see, like, which one of the players takes off which piece. But at the end of the day, I I do think it all kind of collapses a bit. And then the interface becomes that voice model, or then the interface becomes that whatever that device is, which is why I think you see a lot of these companies going after the device. Apple, Google have an advantage there, obviously, and you see a lot of hardware initiatives in some of the other companies.

这对优步这类公司纯粹是负面影响吗？过去的核心优势在于拥有用户访问的主页，可以交叉推广快递、外卖等服务。是否存在某种可能，这种演变最终不会对全球用户造成净损失？

Is that just net bad for for these companies like Uber? Like, obviously, one of the, you know, powers in the past was having this homepage that people went to, and you're like, oh, maybe you thought about using a courier, or you thought about, you know, food or other things that you can cross sell. Is there any world in which this doesn't end up as like a net negative thing for viewers of the world?

我们拭目以待。我认为就像iPhone范式曾给优铺创造成功机会那样——当每个人口袋都装着GPS时。

I I we'll we'll see. I think that like the, you know, just the iPhone paradigm gave birth to gave Uber a chance to succeed. Right. When you have GPS in every pocket.

但严格来说，当时技术还不足以支撑整合成单一消费者门户来接入所有应用

But arguably, the tech wasn't good enough to consolidate around one single consumer homepage that could just then access all these other apps

确实。我是说，其实已有早期迹象显现。

for you. Like Yeah. Yeah. I mean, look. I I think that there there are early signs of this.

东南亚和亚洲涌现了许多超级应用，这源于Web2.0甚至Web1.0时代——某些国家用户打字比点击更困难。这些语言环境自然催生了超级应用模式，虽看似杂乱但极具实用性。既然已有先例，那么消费者领域会出现超级应用吗？

So in Southeast Asia, there's a lot of super apps that developed in in Asia as well, and this was born, I believe, out of you know, the the roots were in just even the web two point o or web one point o days where for some of these countries, typing was harder than clicking. And so you just had this emerging behavior, in those languages to to just have a super app like what we would consider cluttered, but, you know, it was highly utilitarian and useful for for some other parts of the world. So it's been done before. Right. We so you so if you're saying, like, is there gonna be super app for the consumer?

语音时代很可能会出现。但优步必须适应这种变化，找到避免被同质化的差异化竞争点。

Probably in in the voice, world. But then, again, Uber's gonna have to find a way to adapt to that, and they have to figure out how to differentiate on some other level they don't get commoditized at at at that level.

我在想最终是否会形成几个巨型超级应用门户，而那些虽擅长预估到达时间、却无法掌控入口的小应用将大幅贬值。

I just wonder if in aggregate, it's like, you know, you end up with a few really big super app homepages and then, you know, smaller applications that maybe are really good at ride ETAs, but without controlling that homepage are are just less valuable.

是的。用户界面重要性将持续降低，甚至比现在更甚。但这些公司仍需建立让智能代理优先选择优步的价值主张。

Yeah. I think the UI will become less and less important, even less so than than today for some of those companies. But, again, there has to be some value prop that that makes it so that your agent would choose Uber on your behalf.

有件事想请教：本·汤普森多次撰文讨论OpenAI同时服务消费者与企业面临的挑战，包括计算资源分配等权衡问题。不知你是否关注过这些观点？

Yeah. One thing I'm really curious for your take on is I feel like Ben Thompson's written a lot about the challenges of OpenAI being both a consumer and enterprise company, and the interesting trade offs that happen in doing that, whether it's where to prioritize compute or relative effort across these things. I'm not sure if you've that stuff, but I'm curious to

我爱的东西我就爱得不得了。

what I love I love that stuff.

你对他写的内容有什么反应？

What's your what's your reaction to what he's written?

首先，我很庆幸自己不用参与其中，因为解决起来会非常棘手。其次，我对Fiji解决这个问题的能力抱有极大信心。不过说正经的，这确实会引发内部斗争，你说得对——这不仅仅是企业级的问题，工作可以按多种维度划分：有消费者端、企业端，还有API和平台层面。

Well, one, I'm glad that I'm not there to to to be a part of it just because it'll be really messy to solve. Two, I had the utmost face faith in Fiji to figure out how to solve that problem. But, you know, joking aside, I think that it is gonna be it it was a struggle internally because you're totally right. It's not just about enterprise, and you you can cut the the the work in many different ways. There was consumer, and there was enterprise maybe, and there was, there was API and platform.

对吧？有趣的是，当初我去OpenAI时，连职称都很难定——既不能叫消费者产品（因为包含企业版ChatGPT），又找不到准确词汇。这恰恰说明那里待完成的工作有多复杂。话说企业级业务包含API吗？

Right? And it was funny because when I when I went to OpenAI, I remember it was really hard to come up with a title because it wasn't consumer because it includes ChatuchPutty for enterprise. It was just there's no good word for it. And that maybe just shows you how complex, like, that place is in terms of all the work that needs to be done. But, yeah, does enterprise include API?

说不准，有些企业可能需要API。这种架构确实很难规划，但我相信Fiji能解决。不过这种冲突确实很大，因为组织结构会随时间变化——我离职时他们把企业版和消费者版ChatGPT拆成了两个部门。

I don't know. Some enterprises might want an API. So it's really hard to figure out how to organize that, and I trust that, you know, Fiji will help, figure that out. But it is a it is a really big conflict because at some point in time, you know, they have to figure and and the organization structure might shift from one time to to another. When I left, they they broke up enterprise and, and consumer CHIPTY into two different groups.

现在听说又合并了，这很合理，毕竟是同一个产品。组织结构会像波浪般起伏，反映公司不同阶段的战略重点。但归根结底，应用部门只有一个CEO，他们需要明确下一步战略重心。就我经历的公司而言，重组不是坏事，而是战略落地的表现形式。

Now I hear it's back together, which makes a lot of sense because it's the same product. So, I think there's gonna be undulations of organizational structures that map the priorities of what the company's gonna do. But at the end of the day, there's there's one CEO of applications, and there's gonna be, there's gonna be a they're gonna need to figure out what is the most important next part of the strategy. But in the companies that I've been at, reorgs are not a bad thing. They are just a reorgs are a way to express your strategy.

等他们明确未来9到12个月的战略重点后，组织结构自然会相应调整到位。

So I think that it will become clear once they figure out exactly the strategy of what's important in the next, you know, nine months to a year that they will make the the the organizational stuff match accordingly.

OpenAI涉猎的广度总让我惊叹——从机器人领域到各种前沿方向。现在基础模型公司似乎开始划定赛道，比如Anthropic专注代码领域放弃消费者业务，但OpenAI依然全面出击，连代码领域也不放过。这几年他们几乎无往不利，你在任期间有遇到过不如预期的事情吗？

Yeah. I'm always fascinated just to the breadth of of things that, you know, that OpenAI tries to tackle, right, and the fact that, you know, they're still kind a hand in the robotics space, and still a hand in all sorts of different areas, and it'll be fascinating to see, and sometimes it feels like we might be getting toward an era of the foundation models picking their lanes, I think Anthropic is like, kind of seems to be punting on consumer, and doubling down on code, but OpenAI seems as ambitious as ever covering They're fighting hell to win code as well alongside It everything feels like OpenAI just hasn't missed. In the last few years, seems like everything, the company just has done so well in so many ways. I'm curious if there's any example you can think of while you were there of something that maybe didn't work, as well as you as you might have thought.

与其说失败，不如说是执行方式的调整。比如我参与过的SearchGPT项目——当时需要独立搜索入口，这对凝聚组织共识很有必要，让大家意识到默认网络搜索对用户体验的重要性。

Well, I think, you know, I I don't think it's things that didn't work. It's more just like, oh, well, we need to do it differently. I think one of the things that I was there was I was there for SearchGPT, which was a separate entry point to search. I think it was really necessary for, the organization to rally behind a singular goal. Again, just, again, expressing the strategy of just getting people aligned that, like, searching the web by default or for that recall is really important, and that's what some people expect.

后来这个功能被整合进主产品，这不算失败，而是找到了更优方案。就像我和寻找产品市场契合的创业者交流时常说的：早期试错很正常。

And then over time, it's just kind of been folded in. So I don't think that was there's no failure there. It's just like a, oh, well, we found a better way to do things. And I think it's the same thing with startups in early stage when I, you know, talk to founders who are preproduct market fit. It's like, great.

你知道吗？要弄清楚，比如你从中学到了什么，并且非常愿意做出那种转变。那种适应性和成长型思维，是所有世代公司共有的特质。

You know? Figure out, like, what did you learn from that and just be really willing to do that pivot. And that adaptability, that growth mindset is just common between every single generational company that's out there.

完全同意。但感觉在这个时代，这种适应性被推向了更极端的地步，因为事物变化的速度实在太快了。而且，你可能已经构建了一些在当时非常有帮助的东西来收集经验，但六个月后模型的发展让它完全过时了。不过你并不会后悔当初的构建，因为那正是获得那些宝贵经验和最终实现广泛传播的必要步骤。

Totally. But it feels like it's on, you know, taken to the even more extreme in this era with just the pace at which things, which do change and Yeah. And the idea that, like, you may have built something that was really helpful to gather learnings at that time, but then is completely obviated by the model six months later, but it's like you don't regret having built that, because that's what was required to have all those great learnings and distribution that you ended up having.

百分百同意，这正是OpenAI、Anthropic以及所有初创公司的创始人令人赞叹的地方——他们总是能说‘啊，我学到了新东西’，然后立刻转向下一个目标。这就是在这个领域工作、见证这些杰出创始人的美妙之处，也是伟大产品诞生的方式。

A 100%, and that's what's awesome about, you know, folks at OpenAI, Anthropic, and, all the companies and all the startups as well is that, you know, the founders are the ones who, you know, are just like, oh, yeah. That that I learned something. Let's on to the next one. Yeah. And that's that's what's awesome about, you know, kind of being in this in this space and seeing a bunch of awesome founders, and you that's how amazing products are built.

你在硅谷一些最具传奇色彩的企业文化中工作过。我很好奇，你如何将OpenAI的文化与你工作过的其他公司进行对比？

I mean, you've worked at some of the most legendary cultures in Silicon Valley. I'm curious how you contextualize the OpenAI culture to some of the other places you've worked.

在我有幸工作过的所有文化中，雄心壮志都是存在的。但OpenAI的梦想比其他任何公司都要宏大——在那里参加全员会议后，你会感觉自己的思维被拓展了，关于通用人工智能到来时可能实现的一切：比如能自主进行研究并持续取得突破的智能体。那种乐观程度...作为一个天生乐观的投资人，我有时需要抑制这种乐观情绪，但OpenAI甚至突破了我的乐观上限，这实在太鼓舞人心了，确实是独一无二的。

In all the cultures I've had the pleasure of working in, ambition was there. I think OpenAI, I would say, dreamed bigger than any of the other companies that I've been at, where, you know, the all hands there, you would just leave with your feeling like your brain just got stretched a little bit in terms of what's potentially possible, when AGI comes and, like, really shooting for that of, like, agents that can do research that will self perpetuate and kind of keep on making advancements. That level of optimism, I'm an optimistic person, and and sometimes as an investor, you gotta tone down that optimism. I'm a very optimistic person, but that stretched my level of optimism, and it was so inspiring. And that's that that was truly just different.

并不是说我经历过的其他文化不乐观，只是层级完全不同。

And I'm not saying the other cultures I've been a part of were not optimistic, it's just a completely different level.

是的。这次对话非常精彩。我们总是喜欢以快速问答环节结束采访，最后塞进几个范围过大的问题。

Yeah. This has been a fascinating conversation. We always like to end our interviews with a with a quick fire round where we stuff in some overly broad questions Okay. At the end.

很好。

Great.

那么首先，当今AI领域有什么是被过度炒作和低估的？

So maybe to start, what's one thing that's overhyped and underhyped in the AI world today?

我之前说过，当今被过度炒作的是智能体。是的，目前它们还做不到真正智能体该做的事，我们正在朝这个方向努力。但我希望当我们更接近目标、当这个词的使用在智识上更诚实时，再来使用这个概念。

I've said this before. Overhyped is is agents today. Yeah. Today, it's it's overhyped because they they don't they don't they don't do things that agents do yet, and we're building towards that. But we're but but I'd love for us to use that word when we get closer to, when that feels to be true, intellectually honest and true.

被低估的领域，我必须说是模型评估。我之前说过，但真的，这是那种让人突然醒悟的事情——等等，你说什么？这就是你需要的？没错就是这样。

Underhyped, I I gotta go with evals. I I've said this before, but, like, truly, it's one of those things that's like, wait. That's what you what are you talking about? Like, that's what you need? It's like, yeah.

这就是你投入心血去塑造这些模型的过程，听起来可能很简单，甚至有点像繁琐的杂活，但这就是关键所在。我认为即便在今天，它的价值也被完全低估了。

That's the elbow grease that you put in to make and shape these models, and it just sounds so simple and maybe sounds so, like, kinda busy worky, but that's the unlock, and I think it's completely undervalued even today.

确实。你觉得现在围绕评估涌现的这些工具，是否还存在根本性问题——这是否因领域差异太大而无法形成通用标准？还是说其实存在共性特征？

Yeah. Do you think any like, obviously, there's all this tooling that's emerged around evals, and it still seems kind of unclear. Is that, like, a problem that is so bespoke to each domain that there's not a common set of characteristics or that there is, like

我认为共性特征是存在的。只是这个领域有时没有得到应有的重视。虽然有些初创公司专注强化学习环境等领域，确实获得了一定关注，但整体而言讨论度还是不足——毕竟这个话题不够'性感'。

I think there's a com I think that there is. I think that it's just it sometimes just doesn't get the love that it that it requires, and and you see some of these, like there's some startups out there, who do, like, RL environments and, you know, kind of the they're really in that space, and so it's getting some love, but I think that it's still under it's not talked about as much because it doesn't feel as as sexy.

这似乎是许多应用中最关键的环节。核心问题在于它能否产品化，还是会演变成定制化咨询服务模式。当然，当然。不过最终答案还有待观察。

I mean, it seems like the most important thing for a lot of these applications. I feel like the big question is whether it can be productized or whether it ends up being these, like, more bespoke consulting like arrangements. Sure. Sure. But I guess to to be determined.

如果你现在不是投资人，而是要去开发一个AI应用，你会选择做什么方向？

Yeah. If you weren't investing right now and you were to go start an AI app and build one, what, you know, what would you go build?

天啊。我可能会深入那些被认为无聊、无人问津，但存在大量人力浪费或优化空间的行业。具体来说，AI语音技术令人惊叹——我们甚至还没开始挖掘它在优化业务流程等方面的潜力。

Oh my goodness. I think I would probably go deep into some business that a lot of people think is boring, that no one pays attention to, that is deeply optimizable in terms of wasted human effort or human effort that could be better leveraged. I'll be more specific. I think that this AI voice thing is incredible. Like, the voice I don't think we've even scratched the surface of what AI voice can unlock in terms of helping business processes, etcetera.

我知道这听起来不符合我在Uber、Instagram和Facebook的工作背景，但我一直痴迷于行业深度洞察。或许这正是当初吸引我加入Uber的原因——解决原子级实体运输的低效问题，这种极客难题太迷人了。所以我会选择开发能解放人力的智能体或语音代理，把人们从不想做的工作中解脱出来。

I know that sounds like probably out of character given I've worked at, like, Uber and Instagram and Facebook, but I've always been super fascinated by how, the people with the right insights and just how deep some of these industries go. Right? And maybe that's why I've I was really attracted to Uber when I joined is that the the the amount of inefficiency in the world of just transporting, you know, atoms from one part to another, it was such a great, you know, nerdy problem that I love. So I think that would be something in that in that way where whether it's some true agent or some voice agent Yeah. That's able to start to, free people from work that, you know, is not the work that they wanna be doing anyway.

再说一次，虽然听起来不刺激，但呼叫中心就是这样一个让我惊叹的领域——我们完全可以提升这个行业。

And and, you know, again, it doesn't sound super exciting, but, like, the call centers is, like, one of those areas that I'm like, wow. Like, we can we can uplevel that.

你之前提到OpenAI的极端乐观文化，这些年肯定也思考过后AGI时代的走向。节目开始前我们聊到你的孩子——随着对世界发展趋势的思考，这是否改变了你的育儿方式？或你想传授给孩子的技能？

You've obviously you know, you mentioned before this OpenAI culture of extreme optimism, and you've I imagine over the last few years thought a lot about, you know, post AGI world and and where things are headed. You know, we were talking before the show about about your kids. I'm curious, like, as you've reflected on where this where the world is headed, has it changed the way you've thought about, like, parenting at all, or, you know, the skills you wanna impart on your kids, or anything like that?

是的，我有过。而且你看，我不知道自己做得对不对，但我认为培养孩子们的好奇心真的非常重要，同时也要让他们明白可以实现自己的愿景，无论那是什么。我在孩子们身上看到这一点，当他们完成一个自己非常自豪的艺术项目时。但你知道吗？

Yeah. I have. And and look, I don't know if I'm doing it right or not, but I I think that it's really important to foster curiosity, in the kids because and and also just to to show them that they can realize their vision, whatever it may be may be. I see this in my kids when they go and, like, create an art project that they're really proud of. But guess what?

等他们长大后，他们将拥有超越当前想象力的工具。我迫不及待想看到他们用未来的工具——甚至是今天尚未接触的工具——为世界带来的创造力。所以对我来说，好奇心的理念以及坚信任何梦想都能实现的勇气至关重要。对现在的孩子们而言，他们绝对可以实现那些愿景。我认为这是其中很重要的一部分。

When they grow up, they're gonna have tools that are beyond a a current imaginations of what's possible. I can't wait to see the creativity that they, you know, bring upon the world with the tools of tomorrow or even the tools of today that they don't have access to, already. So that to me, the idea of curiosity and the boldness to, know that whatever you dream up, you can for certain do. For the kids of today, they can for certain achieve those those visions. That's that's, I think, a a big part of it.

同时我相信基础很重要。你必须深刻理解数学概念，才能学会结构化思考。编程就是这样一件事——尽管现在有Cursor和代码生成模型，但我认为编程本质上是学习一种思维方式：将问题分解为组成部分，而非单纯写代码本身。从这个角度看，Cursor简直是天赐良机，对吧？

And I believe the fundamentals are important. I mean, you have to deeply understand mathematical concepts to appreciate, you know, what to to to think, you know, very in a in a structured way. Coding is is one thing that, like, even though, you know, we have cursor and and the coding models, I believe that coding is about learning a way of thinking, of taking a problem and subdividing it into component parts rather than the actual code writing itself. And and that way, Cursor is such a blessing. Right?

能够不仅停留在高层构思，还能明确自己想要构建的事物的抽象层级，并快速将其实现——让公司、产品或工具以正确的方式构建，具备所有应有的功能特性。

To be able to say, well, it's not just a high level idea, but, like, to be able to know, well, what is the level of abstraction that I wanna architect this thing, and how do I quickly get that to fruition so that this this company or this this product or this tool I wanna build, is is framed in the right way, and it has all the right affordances.

说得太好了。这次对话非常精彩，肯定有很多值得深入探讨的话题。最后的时间交给你了。

Yeah. I love that. Well, this has been a fascinating conversation. I'm sure there's all sorts of threads folks wanna pull on. I'll just leave the last word to you.

大家可以去哪里了解更多关于你的信息呢？任何渠道都可以，现在由你主导。

Like, where can folks go to learn more about you? Any anywhere. The floor is yours.

好的。我很期待结识一些创业者——那些思维清晰、坚韧不拔、有抱负又肯实干的人，他们能看到市场中真正独特的机会。我通常会查看主动来信，但最佳联系方式是找到认识我的人进行引荐，我回复很快。我特别期待看到那些早期创业者的创造力，等不及要和他们合作了。

Yeah. So I I I'd love to meet some founders. I I love people who have that that clarity of thought, that that grit, that ambition, that elbow grease to, like, they see something really unique in the market. I get out of inbound, so the best way to reach me is to find a, person who knows me and get a warm intro, and I respond pretty quickly. But I'm really looking forward to seeing all the creativity of of these early stage rounders that are out there, and I can't wait to work with them.

太棒了。非常感谢，这次交流非常愉快。

Amazing. Well, thanks so much. Really appreciate it. It's been really fun.