Google Labs 与 Gemini 负责人 Josh Woodward：每家公司都要像实验室团队一样运转

Lenny Rachitsky

我的论点是：每家公司都需要开始像实验室团队一样思考和运转，因为我们所依赖的底层技术变化太快，能做的事也就变化太快。大家都得很擅长拿最新模型试一堆东西，在竞争对手、在创业公司把你的午饭抢走之前，先弄清楚现在什么变得可能。

My thesis is that every company needs to start thinking and operating like a labs team, because the underlying tech that we’re all building on is changing so fast, which means what we can do is changing so fast. Everyone has to get really good at trying a bunch of stuff with the latest models and figuring out what is now possible before your competition does, before a startup comes and eats your lunch.

Josh Woodward

你的前提完全正确，不过难点在细节。某种意义上每家创业公司本身就是实验室，真正的问题是：怎样把同样的方法、点子和能量带进大团队，甚至中等规模的团队；又怎样营造一个环境，去培育点子，也培育能在其中茁壮成长的人。

I think your premise is spot on. I think the devil’s in the details with all those things you just said. In some ways every startup is a lab, and the real question is how you bring some of that same methodology, ideas, energy, excitement into bigger teams or even medium-sized teams. And then how you build that environment and nurture and cultivate both the ideas but also the people that can really thrive there.

Lenny Rachitsky

Josh 运营着世界上存续最久、规模最大的实验室团队，他会分享很多非常具体的经验：最好的点子从哪来、怎么判断该放弃一个想法、产品契合长什么样、今天哪些技能在升值，以及为什么各职能角色并没有真正融成一个「万能 builder」。

Josh runs the longest lasting and biggest labs team in the world. And he shares a ton of very specific lessons about where the best ideas come from, how to know when to quit an idea, what it looks like when you have product market fit, what skills are becoming more valuable in today’s era, why roles aren’t actually blurring into one builder role, and so much more.

Lenny Rachitsky

回顾 Labs 乃至 Google 里最好的那些点子，它们来自哪里、有什么规律？

If you were to reflect back on the best ideas that emerged out of labs and even in Google in general, what are some patterns of where the best ideas come from?

Josh Woodward

说实话我希望书末有个标准答案，但它们来自很多地方。一条经验是：你得一直留意那些抓住你眼球的东西。至少在我的经验里，它们很少来自设计冲刺，或日历上专门划出来「去想下一个点子」的时间。更多是有人在走廊里闲逛，或在周末、假期里，Labs 的人忍不住自己动手做，回来分享一段视频说「看这个太神奇了」。

I wish there was an answer in the back of the book, but they come from a variety of places. One lesson is that you always have to be watching for something that catches your eye. They very rarely, at least in my experience, have come from design sprints or times where you set aside on a calendar, okay, this is where we’re going to go come up with our next idea. They usually come when people are swimming in the hallway, maybe on a weekend, often over a break. A lot of our labs teams will be hacking and building on stuff and they just can’t help themselves, and they’ll share a video and be like, look at this amazing thing.

Josh Woodward

第二，它们不出现在常规的地方，你没法「微波炉加热」出好点子。第三，有时关键不在点子本身，而在谁是有意思的人、他们有什么有意思的小项目和热情，并围绕这些创造化学反应。过去四五年里，不管是 NotebookLM、Flow、AI Studio 还是 Google Beam，都是由很小的一群特别着迷、特别好奇的人先做起来的。

Second lesson is they don’t come in conventional places. You can’t microwave good ideas. A third thing I’ve seen is it’s less about the idea and more about who are the interesting people and what are their interesting little side projects and passion things, and creating the chemistry around that. Over the last four or five years, all these ideas, whether they’re NotebookLM or Flow or AI Studio or Google Beam, get started with very small groups of people who are just obsessed and usually very curious about something and they just won’t stop.

Josh Woodward

我们团队里有个说法叫「几乎可能」，我们真的维护一份清单来追踪这些事。当其中一项跨过去，就像物质从固态变成气态那样发生相变：「哦，这下可能了。」你如果身处前沿、玩模型、看 X 上发生的事、手里又有这份清单，看到某样东西就会判断它刚刚变得可能，然后我们想的是怎么围绕它拉一支团队冲上去。

We have a phrase on the team, we talk about what’s almost possible, and we just have a list where we’re trying to track those things. When one crosses over, it’s almost like a phase shift, when something goes from solid to gas. You’re like, oh, that just became possible. If you’re at the frontier playing with the models, watching what’s happening, and you have this list, then when you see something, whether it’s on X or wherever, you’re like, okay, that just now has become possible. Then we try to think about how we get a team around that and just go.

Josh Woodward

光有技术不够，还得和一个有意思的用户问题匹配；如果组合得足够好，人们愿意为它付费。我们很早就开始找这种组合。我们正在密切关注的一个方向是未来的娱乐形态——沉浸式娱乐。在技术层面，如果模型能达到某种延迟、某些语言和模态，你就能开始推演什么会有意思。

Tech alone isn’t enough; you’ve got to match that to an interesting user problem, and ideally put those together in a way that’s so good people will pay for it. One area we’re watching very close right now is what the future of entertainment is going to look like, this whole idea around immersive entertainment. If a model can get to this kind of latency and this amount of languages and modalities, you can start to think through what would be interesting from the tech side.

Lenny Rachitsky

所以你不是在等灵感降临，而是做了大量前置功课：什么几乎可行、未来可能走向哪里、哪些领域会被颠覆。

So this is not just waiting for inspiration to strike. You’re doing a lot of pre-work to think about what is almost working, where we think the future might be going, areas that are going to be ripe for disruption.

Josh Woodward

对。我们那份文档里有 82 条预测，都以「我们相信未来是……」或「我们预测……」开头。做一条预测都很危险，更别说 82 条。但你得对未来有观点，同时知道十次里有九次，或者 9.9 次会错。我们要培养的价值之一就是追逐未来：你几乎得把自己扔出去，尝试活在未来里一点点。

That’s right. In this doc I talked about, we have 82 predictions that all begin with we believe the future is or we predict. This is dangerous ever to make one prediction, let alone 82. But I think you have to almost have a point of view about the future knowing that nine out of 10, or maybe 9.9 out of 10, you’re going to be wrong. One of the values we try to cultivate in the team is how do we chase the future. You almost have to throw yourself out there and try to live in it a little bit.

Lenny Rachitsky

你见过的被扔上墙、被尝试和原型化的点子，可能比任何人都多。你怎么判断某样东西值得加倍投入？产品契合到底长什么样，会有什么让人意外的地方？

You’ve seen more ideas thrown at the wall and tried and prototyped than maybe anyone else. What are signs to you that this is something we need to double down on, and what’s something that might surprise people about what product market fit looks like?

Josh Woodward

最棘手的一点是，它的艺术成分远大于科学。我们很会跑科学实验、做数据驱动的决策、定 OKR，但早期有时只是团队里几个人有信念，说不清楚。稍微有用一点的是，我们常谈怎样走出大楼给人看原型。这时我一直在看人的眼睛，这就是指标——不是 DAU 除以 MAU，也不是 7 日留存，没有这些。就是问「Lenny，你看到这个了吗？你觉得呢？」如果你的眼睛亮了、身体前倾，这就是早期要找的信号。

It is way more art than science. We know how to run science experiments, make data-driven decisions, set OKRs, but sometimes in the early days it’s literally a couple of people on the team who just have conviction and can’t explain it. What I think is a little more useful: we talk a lot about how to get outside the building and just show people prototypes. When you’re showing people early prototypes, you’re looking at people’s eyes, and that is the metric. It is not a DAU over a MAU. It’s not a D7 retention. It’s literally, hey Lenny, did you see this? What do you think about this? If your eyes kind of light up, if you lean in, these are the signals you’re looking for in the early days.

Josh Woodward

对于那些容易兴奋的人，我们会想：我们爱上的是什么问题，而不是什么产品。因为我发现往往要经过三四五次转向，才能做出东西，或者干脆认定没成。至少在我的经验里，人们先爱上某个具体的产品方案，通常会过得很糟。

The people who get excited about things, we try to think about what’s the problem we’re falling in love with, not the product. What I’ve found is it takes three, four, five pivots before you either got something or you’re just like, that’s it, didn’t work. When people fall in love with an actual product solution first, at least in my experience, it’s going to lead to a bad time.

Josh Woodward

过去一年左右，Nano Banana 在 Gemini 上爆红、NotebookLM 的播客走红、Flow 或 Stitch 上发生的一些小爆点，那时你真的只能扑上去、死死抓住。这既让人精疲力竭，也特别有趣。但你几乎得多开很多枪。

In the last year or so, when Nano Banana went viral on Gemini, when the NotebookLM podcasts have gone, when stuff’s happened on Flow or Stitch, you’re literally just swarming it and holding on for dear life. It is both so exhausting and so fun. But you almost have to take a lot of shots.

Josh Woodward

有时团队会等着领导者拍板，但我发现恰恰相反：团队通常比领导者更早知道。信号是：热情开始消退，该试的都试了还是不行。人们知道一个项目、一段关系什么时候不行了。我作为领导者要做的，是营造一种环境，让人们能问那些让人不舒服的问题，敢说「我觉得这行不通」。

Sometimes a team is looking for the leader to make the decision. What I found is the opposite: the team usually knows before the leader knows. The signs are, does the passion start to run out? You’ve tried everything you can think of and it’s just not working. People know when a project’s not working or a relationship’s not working. What I try to do as a leader is create an environment where people can ask those uncomfortable questions and say, hey, I don’t think this is working.

Josh Woodward

举个昨天的例子：我们有个很期待的 Gemini 功能要上线，拿到早期数据，人们去看了对方的眼睛，结果就是不对。团队里的 PM 说「我觉得我们不该发布，它不够好」，我特别为她骄傲，回复所有人说「太棒了，谢谢你这么做」。

Yesterday, just as an example, we had a feature we were really excited about in Gemini that we were going to launch, and we got some of the early data back. People went out, looked in the eyes, and it just wasn’t there. I was so proud of the PM on the team because she was like, I don’t think we should launch this, it’s not good. So I replied all and said, this is great, thank you for doing this.

Lenny Rachitsky

你同时管 Labs（从零到一的实验）和 Gemini 应用（可能有十亿用户）。为什么会是这种结构？

You oversee labs, which is zero-to-one experiments, and then the Gemini app, which is one of the most popular apps in the world with probably a billion users. How did that come to be? Why is that the structure under your team?

Josh Woodward

它们确实是规模的两端。Gemini 有超过十亿用户；而 Labs 早期的零到一项目，我们开玩笑说，一万个鼠标点击就让我们兴奋不已，多数产品的仪表盘根本不算这么低的数。但它们有共同点：即使 Gemini 规模很大，也很有「从零到一的心跳」，吸引的是同一类 builder；我们都是对 AI 着迷的人；第三，团队的优先级顺序是「用户第一，Google 第二，我们的产品第三」，Labs 第三，Gemini 也第三。

They are very different ends of the scale. Gemini, over a billion users. Labs, we joke that sometimes our early zero-to-one projects get so excited about 10,000 mouse clicks, and most other products’ dashboards don’t count that low. But there are a few common pieces. The type of people attracted to these things, even Gemini despite its size, very much has a zero-to-one heartbeat. We’re all AI nuts and maniacs, we love this stuff. And the mantra on the team is users first, Google second, and our product third, so labs third or Gemini third.

Lenny Rachitsky

Labs 现在做的一切都基于 AI 吗？

Is everything you’re doing in labs now AI based?

Josh Woodward

现在全是 AI 驱动的。四五年前刚开始时我们探索过一些别的方向，后来决定至少在可预见的未来全力押注 AI。但我们不只做软件，比如 Google Beam，是个很神奇的「魔镜」，能创造 3D 版本，让你和对面的人像全息影像一样交谈，几乎可以算 AI 硬件。

It’s all AI driven right now for sure. When we started four or five years ago, we had a few other areas we were exploring, and we made the call, at least for the foreseeable future, to go all in on AI. We don’t just do software stuff though. Google Beam is this incredible magic mirror that can create a 3D version, where you can talk to someone almost like they’re a hologram on the other side. You could consider that almost like AI hardware in a way.

Josh Woodward

我觉得消费领域是现在最令人兴奋、竞争也最激烈的地方之一，感觉和「消费已死」恰恰相反。有几个有意思的方向：一是娱乐会去哪——我们会永远刷不同的信息流，还是这项技术能让我们在现实里更亲近、和别人一起做事？二是整个消息聊天领域曾经似乎被「解决」了，现在谷歌内外都有很多实验，不再像以前那么固定。三是回到问题本身：世界上稀缺的是什么？时间、金钱，以及人们能一起拥有的线下体验。

Consumer is one of the most exciting, competitive spaces right now; it feels completely opposite to me. A few areas are interesting. One, where’s entertainment going to go? Are we going to be looking at different types of feeds forever, or are there ways this technology can bring us closer together in the real world or do things with other people? Two, the entire messaging chat space felt like it was solved, and there are lots of experiments happening on that right now both inside Google and outside. Three, going back to the problems: what are things that are super scarce? Time on earth, how much money people have and how they can save money, and experiences people can have IRL together.

Josh Woodward

无论它是聊天机器人还是个人智能体的形态，我觉得很多东西都还有待被发明。

Whether it takes the form factor of a chatbot or a personal agent or whatever that’s going to look like, I think is kind of yet to be invented in a lot of ways.

Lenny Rachitsky

来个尖锐的问题：现在每个新发布的 AI 助理应用都在接入 Gmail、日历、文档，用这些 Google 数据做出神奇的事，而 Google 自己好像没做。是隐私顾虑、官僚体系，还是马上就来？

Let me ask you a spicy question. Every AI assistant app that’s launching these days sucks in Gmail and Google Calendar and Docs, all the Google data, and does all these magical things with it, and Google hasn’t. What’s kept you guys from doing something like that? Is it privacy concerns? Is it bureaucracy stuff? Is it coming?

Josh Woodward

有些部分今天已经有了，比如在 Gemini 里用 Gemini Spark，你可以把这些都连起来，相当强大。我们在迭代的是：即使看其他竞品，也有不同的「模式」要切换，而过去几周的教训是——忘掉模式和开关，就一个提示框，告诉它你想要什么，它直接去做。这是整个行业的 UX 简化趋势。另外人们常常对我说：「我所有东西都在你们这，赶紧让它能用。」他们其实不太愿意把这些连到竞争对手的产品上。请盯着这个领域，我们听到了。

There’s parts of it out there today. In Gemini, in Gemini Spark, you can actually connect all that stuff and it’s quite powerful. Where we’re iterating is that other competitors have different modes you go into, and the lesson even from the last couple of weeks is forget the modes and toggles, just one prompt box, tell it what you want and it just does stuff for you. That’s a broader UX simplification happening across the industry. And one of the main things people will ping me about is, “Oh man, Google, I’ve got all my stuff with you, just make the thing work.” There’s a lot of reluctance to connect it to competitor products. So watch this space. We hear you loud and clear.

Josh Woodward

我们对这件事的思考是「个人智能」：你想要一个个人的、主动的、强大的、为你工作的 Gemini。这一年多来一直是我们的愿景，我们现在也在试很多东西，很快会推出。

The way we think about this is personal intelligence: you want a Gemini that is personal, proactive, powerful, that works for you. That’s been the vision for the last year or so that we’re headed towards, and there’s a lot of stuff we’re playing with right now that will be out soon.

Lenny Rachitsky

你觉得 AI 里什么被低估，如果有的话，什么被高估？

What do you think’s underhyped in AI right now, and what do you think is overhyped, if anything?

Josh Woodward

被低估的，我今年夏天想了很多，在内部写了份 Labs 原则的文档：我们为什么做我们做的东西？我坚信，不管你在 Google 还是创业公司，你通过产品说话，产品其实编码了你的原则。我希望业界对这些原则、价值观、我们想要什么样的未来有更多讨论。被高估的，我已经说了大概一年了，仍然是模型基准。

Underhyped: I was thinking a lot about this over the summer. I wrote a doc internally on labs principles, how and why we’re building what we build. I believe very strongly, whether you’re at Google or a startup founder, you speak through your products, and your products represent and encode those principles. I wish there was a lot more discussion in the industry around that, what kind of future do we want. The overhyped one, I’ve been saying this maybe for a year now, I still think it’s overhyped: model benchmarks.

Josh Woodward

大多数人永远不会在乎 Elo 分数，可能连怎么念都不知道。现在 AI 的讨论大多是速度和参数、基准——它们当然是能力的代理指标，但业界有时会过度偏向，忘了这到底对谁有用、什么产品或功能能带来真正的价值。

Most people in the world are never ever going to care about an ELO rating. They may not even know how to pronounce ELO. The talk in AI right now is about speeds and feeds and benchmarks, and these are proxies for capabilities, but as an industry it can over-rotate to that and lose sight of why this is useful to someone and what’s the product or feature that’s going to deliver real value.

Josh Woodward

关于品味、判断、氛围编程已经有很多讨论了。我在找的是「被错误定价的信号」。第一个是「遗忘速度」：大家常说学习速度，我现在更想看一个人学会某样东西之后，根据测试产品时看到的东西，能多快放下它。第二个是「爆发式耐力」，这个词来自一本关于费德勒的书：你既能全力冲刺，又能照顾好自己和身边的人，持续很久。第三个是协作：即使编程越来越趋向于指挥一堆智能体，我仍觉得和人合作的方式更重要了，要能在与智能体（或一群智能体）和与人之间都成为出色的协作者，能快速建立并放大信任。

There’s a lot out there on taste and judgment and vibe coding. I’m always trying to find the mispriced signals. One is unlearning rate: people talk about learning rate, but I’m looking for how fast someone can learn something and then walk away from it, based on what they’re seeing when they test their product. Another is explosive endurance, from a book about Roger Federer: people can go really hard, the explosive part, but over a while, taking care of themselves and the people around them. And another: as much as coding is trending towards directing lots of agents and not having to work with as many people, collaboration and how you work with people matters even more. Can someone be an incredible collaborator with an agent or a swarm of agents but also with people, and build trust, scale trust?

Josh Woodward

其一，执行成本下降后我们会做更多产品，人们会和不同的人组队做东西。组建新的 Labs 项目小队，过去大约五到七人，现在大约两到三人，最多四人，就像出色的音乐人和不同的人组队演奏不同的曲子。

As the execution cost of making products goes down, we’ll be able to make more products, and people will team up with different people to make stuff. When we’re fielding a new team for a labs project, we used to say it’s about five to seven people. Now it’s about two to three people, maybe four. You can almost imagine super talented musicians who team up with different people and then they play different songs together.

Lenny Rachitsky

你们把设计师、工程师、PM 都拉进 Labs 一起工作，谁最有影响力、最容易领导产品点子，这方面有变化吗？

You have been pulling designers, engineers, and PMs into labs, and they all work together. Has there been a shift in who you find is most impactful and successful in leading product ideas?

Josh Woodward

我们研究了很多，也测试了很多。总体上 PM 这个职能因为天然更横向，似乎适应其他职能更快，我们在公司里跑小型 builder 试点时就看到了。但最好的人可以来自任何职能：战略、运营、BD、市场的人在 Labs 里也能当 builder，而且做得很好。我觉得现在被高估的是「人人都是 builder、这是新职能、一切都一样」。我们看到的是人们仍有各自的专长，就像大学的主修和辅修。如果人人变成 builder，会失去对这些专业的尊重和欣赏。

We’ve been studying this a lot and testing a lot of this too. In general, PM happens to be the function that, by nature because it’s a bit more horizontal, has been able to adapt to some of the other functions faster. We see this when we run little builder pilots across the company. But our best people can come from any function. We’re seeing strategy and ops people and BD people and marketing people be builders in labs and they are really good. What’s overhyped right now is everybody’s a builder and it’s just a new job function and everything’s the same. What we’re seeing is that people still have their specialty, like their major or minor in college. There’s a risk if everybody becomes a builder that you lose a lot of respect and appreciation for the expertise that comes from those specialties.

Josh Woodward

我们给新 PM、APM 或早期职业的人的建议是：氛围编程、做原型很好，但要以成为一名卓越的 PM 为目标，不要丢掉你的主修。

That’s the advice we give especially to a lot of new PMs, like APMs or people earlier in their career: it’s great you’re vibe coding and making all these prototypes, but do that in service of becoming an exceptional, extraordinary PM, and don’t lose that major.

Josh Woodward

我们把 Labs 想成一个个「赛季」：任何时候大约有二三十个实验同时跑，但对每个小团队，我们试着建立里程碑或节奏。我作为领导者学到的最好一课，是把它说出来、宣告出来：人们可以选择加入或不加入；同时明确哪些团队此刻不在这种模式里。比如一个大模型要来，这五个产品要用它来展示——那是爆发期。而在 Google I/O 之后的六月，很多 Labs 团队就是去随便做、去重新发现新东西。人们的直觉是「六月了，没在发布，效率低」，其实那正是下个赛季种子的来源。

We think of seasons in labs. At any given point there are 20 or 30 experiments running simultaneously, but for each individual little team we try to create milestones or a rhythm to the work. The best lesson I’ve learned as a leader is to name that, declare that; people opt in or opt out, but also be very clear when a team is not in that mode. We’ve got a big model coming, these five products are super important to showcase this model, it is explosive time. After Google I/O, a lot of labs teams will sprint, and then in June it’s, hey, just go hack, go rediscover what’s new, go build, have fun. The temptation is, oh man, it’s June, it’s unproductive, we’re not shipping, and no, actually those are where all the new seeds are coming from for the next season.

Josh Woodward

过去人们谈工作与生活平衡，我喜欢「爆发式耐力」这类概念，是因为我们希望作为团队做很多很多年，你不能九个月就把油箱掏空把自己累垮，既要着眼长期，又要认识到眼前这一刻很快。

In the past people would talk about work-life balance. Why I like things like explosive endurance is we want to be doing this as a team for many, many years. You can’t just empty the tank in nine months and kill yourself. You have to have a view towards the long term but also recognize the moment you’re in is fast.

Josh Woodward

前几天有人问 Gemini 的两年路线图是什么，我说我们有六个月的路线图，也有多年的愿景。世界变化太快，我两天前看到的东西，现在已经成了 48 小时内完成的项目，这你没法提前规划。

I got a question the other day on Gemini, what’s our two-year roadmap, and I was like, we have a six-month roadmap on this thing, we have a multi-year vision. The thing I saw two days ago is now a project that happened in 48 hours. You can’t plan for that.

Josh Woodward

我们通常想的是大约六个月，因为那差不多是「几乎可能」的窗口——你假设每家实验室一年出几次新的预训练模型，中间还有些节点。然后我们想象大约一百天里，你能把产品实质性推进到哪。我们最好的 Labs 团队，从点子到一个有意义的里程碑往往是 50 到 100 天，一部分在于怎样设计一种文化和环境去扫除阻碍，让人们尽情去做。规划就是滚动的六个月。

We usually are thinking somewhere on the order of about six months, because that’s that almost possible window where you assume every lab is cranking out some new pre-training model a couple times a year, with individual points in between. And in about a hundred days, we think about where you can move your product meaningfully forward. Our best labs teams go from idea to some meaningful milestone in 50 to 100 days. Part of that is designing a culture and environment where you unblock a lot of stuff so you can just let people cook. We think about planning as almost like rolling six months.

Lenny Rachitsky

Dan Shipper 在我的峰会上讲到，要有一支专门试新东西的前沿团队，让公司其他人不被干扰。你怎么看这种模式？

Dan Shipper gave a talk about how, where the tech is constantly changing underneath you, his advice is to have a frontier team dedicated to testing out the new stuff so the rest of the company doesn’t have to be distracted. How do you feel about that as a model for companies?

Josh Woodward

每家公司都肯定需要这样一群人：他们的工作就是待在前沿，如果他们没听说过过去 24 小时 X 上爆红的最新东西，那就该问问在干什么。他们就是那些在各种东西上订了八个每月 200 美元订阅的人。第二是让这些人能做东西的环境。第三，看看历史就知道：大公司办实验室，通常 3 到 4 年就失去效力，要么是公司永远没法把实验室产出的东西商业化，于是别人看到未来后另起炉灶。所以你需要一群人，能生活在前沿，同时又能和更大的公司连通，别被看成只会搞噱头和玩具的奇怪研发组。

Every company definitely needs a set of people whose job is to be at the frontier. If they haven’t heard about the latest hot thing viral on X in the last 24 hours, what are you doing at your job? They’re the people who have like eight $200-a-month subscriptions on everything. The second thing is the environment in which those people can build. The third part, if you look at the history, a large company who starts a lab, the lab is usually ineffective and fizzles out after about three to four years, or the company can never commercialize the things coming out of the lab. So someone comes by, sees the future, and does it as a separate company. You need people who live at the frontier but can bridge across the bigger company so it’s not just some weird R&D group always coming up with gimmicks and toys.

Josh Woodward

随着 Labs 长大，我们内部发明了几个词：零到一 Labs、一到十 Labs、十到百 Labs，对应实验室生命周期的不同阶段。NotebookLM、Flow、AI Studio 这些规模变大的项目，有挤占零到一项目的风险。零到一阶段看的是前面说的「眼睛」，十到百阶段则看漏斗和 CPA，怎么把人转化到产品的神奇时刻，是一场规模化的游戏。

As labs has grown, the terms we’ve started using are zero-to-one labs, one-to-10 labs, and 10-to-100. A bigger project like NotebookLM, Flow or AI Studio that has grown out of labs risks crowding out the zero-to-one stuff. In zero to one, you’re looking at eyeballs like we talked about. In 10 to 100 you’re looking at waterfalls and CPAs and how we’re converting people to magic moments in the product. It’s a scaling game.

Josh Woodward

一，要创造一个让奇怪的东西能生长的空间，你不能把它塞在现有业务单元下面，它得有自己的独立性，在有些公司意味着要向 CEO 汇报，或者至少保持足够的距离。二，一套共同的价值：在我们这是「用户第一，Google 第二，Labs 第三」，实验室的工作是让用户成功、让 Google 成功，而不是让自己成功。还要想清楚你的实验室是要把东西「毕业」到现有产品，还是要发明新的产品线和品类；我们一开始是毕业型实验室，后来随着 AI 模糊了很多品类而扩大。

One, you’ve got to create a space where weird things can grow, and you can’t just nestle this under an existing business unit. It has to have its own independence, which in some companies means it needs to report to the CEO or have enough distance. The second thing is a set of people and a mindset: users first, Google second, labs third. The job of the lab is to make users successful and Google successful, not to make itself successful. Is your goal for this lab to graduate things to existing products, or to invent new product lines and categories? In our case we started as a graduation lab and it’s expanded over time as AI blurs a lot of categories.

Josh Woodward

三，我可以想象 CEO 们会感兴趣：把实验室不仅用于产品开发或技术探索，也用来挑战公司的工作方式。我们的说法是，Labs 有时会在 Google 制造「好的麻烦」。例如我们是最早看到职能角色在模糊的团队之一，所以我们说需要一条 builder 职级，先在 Labs 试点，再推广到整个 Google。

The third thing, and I could imagine for CEOs listening, is use your lab not just for product development or technology exploration but even for challenging how the company works. The phrase we use on the team is, lab sometimes creates good trouble at Google. We were one of the first teams that were like, we’re seeing the job roles blur, we have to have a builder job ladder and we should use labs as the place to pilot it before we roll it out across Google.

Josh Woodward

我们痴迷于找那些停不下来要做东西的人，我们有一份「Labs 概要」文档，新人入职第一周我会发给他们。他们一直在做东西，求知欲强，在读、在刷 X、在听播客，痴迷于解决别人的问题而不是自己的自我与荣耀，在他们身边会增加能量。面试时如果有人说「让我给你看我刚做的东西」，立刻有可信度；或者我问他们面对未知、没有结构的「黑暗洞穴」怎么找出路，被这种事情激发的人也会大大加分。

We are obsessed about finding people who can’t stop themselves from building. We have a doc on the team, Labs in a nutshell, and the first week someone joins the team I email it to them. They build all the time. They are intellectually curious. They’re reading, they’re on X, they’re listening to podcasts. They are obsessed about solving people’s problems, not about their own ego or glory. When you’re around them, they add energy. In an interview, if someone says, hey, let me show you this thing I just built, instant credibility. Or if I ask them about unknown, unstructured dark caves and which ways forward, people that are energized by that get a huge credibility boost.

Josh Woodward

组织结构上最大的错误？一是人们急着归并、追求整洁，我多年下来的教训是有时得在混乱里待一阵，看会涌现什么，同时仍愿意做调整。二是招人太多太快，团队一两个季度翻倍翻三倍是虚荣指标。我多年前做过一个项目，到了 30 个工程师我很开心，但没有产品契合，只有一个很鼓舞人的愿景，结局你大概能猜到。

The biggest mistake on org structure: people rush to group and consolidate, make it clean, and I’ve learned the hard way that sometimes you kind of have to live in the messiness a little bit and see what emerges. The other common mistake is you hire too many people too fast. It can be so addicting to look at, our team doubled or tripled in the last quarter, and these are vanity metrics that are really bad most of the time. I remember I was on a project years ago and we got to 30 engineers and I was so happy and we had zero product market fit but a very inspiring vision.

Lenny Rachitsky

Logan Kilpatrick 让我问你：五年后 Google 会有一万个产品，还是五个？

Logan Kilpatrick suggested to ask you: in five years, do you think Google is going to have 10,000 products or five?

Josh Woodward

这是我和 Logan 长期争论的问题，我自己的答案还总在变。一种框架是一万个产品就是一万个有各自品牌的界面，复杂得很，我不确定还可能；因为你可以有个「产品工厂」不断吐出产品。我越来越多地想的是，会不会出现更多由模型生成的、更定制化的体验，背后是 Labs 团队做的智能体，你在一个共同的界面里体验它。Labs 里有几个实验在同时验证两边：假如一切收敛到一两个超级应用，世界什么样？假如极度碎片化呢？目前我偏向超级应用在某些场景和地区确实很成功，但对很多人用起来很复杂；而且这还假设点这些体验的大多是人。

This is a long-running debate Logan and I have on the team, and I keep switching my answer. One framing is 10,000 products from Google, 10,000 surfaces with their own brands, all complex, and I’m not sure that’ll be possible, because you can have a product factory that’s just going to turn out these products. I’m more and more thinking about experiences that are more bespoke that the models might make, and there might be labs teams behind it making the agent that does this thing that you experience in a common surface. We have a couple of experiments in labs going on this: if this thing really consolidates down into a super app or a set of super apps, what’s that world going to look like, and if it super fragments, what’s that world going to look like? Right now I’m probably more in the camp that super apps have worked in some use cases in some parts of the world extremely well, but for a lot of people they can be very complex. And this is assuming humans are going to be the ones tapping most of these experiences.

Josh Woodward

我想留给听众的一点：我觉得现在还有很大空间，去想怎么获得关于你正在做的产品和体验的反馈；用 AI 来帮助收集、汇总反馈、获得洞察，会有很多有意思的东西。一个好的 PM 对某类人、对什么会让他们惊喜或解决他们的问题有洞察，那么怎样用 AI 破解这个反馈回路？

One thing I’d leave: I still think there’s a lot of space right now on how to get feedback about the products and experiences you’re building. There’s a lot of interesting stuff happening with AI applied to that. What makes a great PM is having insight about a certain type of people and the types of things that are going to delight them or solve their problem. How do you crack that feedback loop with AI and these new types of products we’re building?

Lenny Rachitsky

我听说你会给队友颁发一些很搞笑的奖项，我最喜欢「TPU 收割者」。分享一下这些奖项和它们的含义吧。

I heard you give these hilarious awards to teammates. My favorite is the TPU harvester. Share these awards you give and what they mean.

Josh Woodward

TPU 收割者：随着产品增长我们需要越来越多的 TPU，想奖励那些做效率优化、回收 TPU 这种脏活累活的人，好让我们开发下一批功能、支撑增长。我们给他们一把用 Etsy 做的迷你耙子，加一笔现金奖励，办成一场庆祝。另一个是金色创可贴，给修「纸割伤」的人——那些加起来会让人烦的小毛病。还有「LLM 低语者」，给那些能以令人意外的方式把模型推到前沿的人，奖品是一副巨大的耳朵和一个很棒的机械键盘，因为他们能听到别人听不到的模型的声音。

The TPU harvester: we just need more and more TPUs as our products are growing, and a few of us wanted a way to reward people who do the grungy work of efficiency and reclaiming TPUs so we can build the next features and fuel our growth. We give out a miniature rake we made on Etsy and a cash award, and we turned it into this big celebration. Another one: people that fix paper cuts, those little things that add up in a product that annoy people. We made a golden band-aid. And the LLM whisperer, people who can push the model to frontier ways in surprising ways: they get a giant set of ears, because they can hear things from the model that others can’t, and a really nice mechanical keyboard.

Josh Woodward

还有一个有点像「秘密社团」：我们给某人的日历发个邀请，看起来有点吓人——四五个团队领导、没有议程；他们一到场，我们就轮流说为什么这么感谢他们在团队里。我总是以「这不是一个坏会议，这是个好会议」开场。

This one is a little bit of a secret society. We’ll put an invite on someone’s calendar, it looks kind of scary because there’s four or five leaders on the team and no agenda. Then they show up and we go around the circle and tell them why we appreciate them so much for being on the team. I always start the meeting with, this is not a bad meeting, this is a good meeting.