OpenAI ChatGPT 负责人 Tibo：Dots、智能体时代与开放生态

Lenny Rachitsky

Tibo，谢谢你来。很酷的房间。你们今天发了一大堆东西——睡眠怎么样？还撑得住吗？

Tibo, thank you for doing this. Nice room. You launched a ton of stuff today. How are you sleeping? Are you doing all right?

Tibo Sottiaux

挺好。团队很强——我睡觉的时候他们大半夜还醒着。办公室有个叫 Library 的房间，我们改成了超级作战室，气氛特别好，大家待到很晚。

I’m doing great. I’m fortunate to have a very good team. While I sleep, they were awake most of the night. We have this amazing room at the office called the library, which we completely repurposed as a mega war room. The vibes in there are just immaculate—people were there until very, very late.

Lenny Rachitsky

你做工程师很久了。现在还写代码、合 PR 吗？还是主要在文档和会议里？

You’ve been a long-time engineer. Are you coding at all anymore? Are you shipping PRs, or mostly in docs and meetings?

Tibo Sottiaux

「写代码」是什么意思？偶尔还是会合一点。大量分析、业务、下一功能、上次发布做得怎么样——很多代码是 Codex 帮我写的。周末偶尔手写一点 LeetCode，当治疗。

What is coding? Yeah, merging some code every once in a while. A lot of code for analysis—understanding trends, the business, the next feature, how previous launches are doing—is written by Codex. Occasionally on weekends I’ll have a little therapeutic moment where I do a little LeetCode by hand. That’s the only actual hands-on coding I do.

Lenny Rachitsky

日常工作你会并行跑多少个智能体？

How many agents do you find yourself running in parallel in your day-to-day work?

Tibo Sottiaux

以前并行得多；后来有了 UltraFast 突破，又能进心流了——更快的智能体帮很大。六个月前那档 UltraFast 很快也会对外。推不同前沿时，我会把智能体团队越做越大；模型再突破，一个更大的智能体就能包办、把一切留在记忆里学，我就又把团队缩回去。扩张—收缩—再扩张。

I used to run a lot more in parallel, then we got a breakthrough with UltraFast, so I’m able to be in the flow again—a faster agent helps a lot. I’m excited that we’re getting six-month-old UltraFast out there soon. As I’m pushing different frontiers, I find myself building larger and larger teams of agents. Then when we have the next breakthrough with models, suddenly a bigger agent can just do all of it and keep everything in memory and learn, so I shrink the team again. Expansion and shrinking, expansion.

Lenny Rachitsky

多说一点——这跟现在大家热衷的 loops、graphs 是什么关系？

Can you say more about that? It makes me think about this whole loops thing and graphs thing.

Tibo Sottiaux

我觉得折腾怎么设 loop、怎么微调，可能让人兴奋过一阵，但不是终局。终局是我们用 Dots 在推的东西：一个极聪明、7×24 工作、懂你目标与偏好、能从反馈学习的智能体。今天这版还不算完美，会从 Pro 用户身上学很多；长远你要的是会学的系统，而不是「我必须把 loop 拧成这样才出结果」。

Having to set up and fiddle with your loops is something people maybe got excited about, but I don’t think that’s the way it’s going to work. The way it’s going to work is what we’re positioning and shipping with Dots: an incredibly smart agent that works 24/7, understands your goals and preferences, and learns from feedback. What we launched isn’t perfect—I’ll learn a lot from making this available to Pro users—but over time you just want a system that learns from what you want to achieve. You don’t want to think about looping it exactly this way to get results.

Lenny Rachitsky

现在有 Codex、ChatGPT Work、消费级 ChatGPT、Dots。听起来你觉得最终会走向以 Dots 为主入口，再去调度别的？

Right now there’s Codex, ChatGPT Work, ChatGPT consumer, Dots. What I’m hearing is you think we’re heading toward Dots as the primary way you talk to AI, and that kicks off everything else.

Tibo Sottiaux

不管叫不叫 Dots，本质是摆脱被技术绑住，拥有永久在线的主动智能：它知道该做什么，任何客户端、任何屏幕都能用。走进会议室它就出现、记笔记；邮件里接上；需要就发短信。不必黏在笔记本或手机上——需要时在，不需要时让开。我等不及这一天；整天扛着笔记本电脑这块砖，是人迁就技术，而不是技术为人服务。

Whether it’s Dots or not, it’s about breaking free from the technology and having this permanent active intelligence that knows everything it needs to do and is available through any client, any screen. You walk into a meeting room, it shows up, takes notes; you pick it back up on email; you text it if you need it. You’re not glued to a laptop or always on your phone—it’s available when you need it and gets out of the way when you don’t. I can’t wait for that, because carrying your laptop as this brick everywhere means you’re tied to the technology instead of the technology working for you.

Lenny Rachitsky

过去两年工作方式变了特别多。你觉得接下来两年会定型，还是还会剧烈变？

So much has changed in how we work over the past two years. Do you think we’re settling into what work looks like for the next couple of years, or is there still more radical change?

Tibo Sottiaux

还会相当剧烈地变。今天语音、多模态输入输出终于开始拼起来，但仍有点笨拙——会一直笨拙到突然不笨拙。那时你会像现在跟我说话一样跟它说话，记忆完美；想头脑风暴就随手涂两笔，有协作画布。我们用 ChatGPT Spaces 发出了信号；你可以想象人类与智能体共享白板协作。很多客户端今天都还没真正跑通，所以还会有一轮大转型。

It’s going to continue to change quite radically. Even today the technology is finally coming together with voice and multimodal inputs and outputs, but it still feels a bit clunky, and it will feel clunky until it isn’t. Then you can just talk to this thing the same way we’re talking right now; it remembers perfectly. If you want to ideate, you squiggle something—you have a collaborative surface. That’s what we signaled with ChatGPT Spaces. You can imagine a shared whiteboard collaborating between humans and agents. None of this is really working today, so we’re going to go through this big new transformation.

Lenny Rachitsky

所以 Dots 是那条未来的大块头——不用再想 Codex 对 ChatGPT？

What I’m hearing is Dots is a big part of that future—you don’t have to think about Codex versus ChatGPT.

Tibo Sottiaux

对。Chat 和 Work 开关我们会合并：大家喜欢 Work 的能力，但有时又想用更快、更舒服的 Chat。最终也会把 Dots 能力直接送进 ChatGPT，抬高大约 12 亿用户的地板，做到真正无缝。Dots 让我特别兴奋的一点是：没有模型选择器，几乎零配置，你只管说话；唯一要配的是想用哪些渠道。很像《Her》那种世界。

Yes. For Codex and ChatGPT, we’re merging chat and the Work toggle. People love the capability in Work, but sometimes you want chat because it’s a little faster and more pleasant. Eventually we’ll ship all of the capabilities of Dots straight into ChatGPT as well and lift the floor for like our 1.2 billion users, and we want it really seamless. One thing I’m very excited about with Dots: they don’t have a model picker. There’s just no configuration. You just talk to it. The only thing you configure is which channels you want to talk on. It feels very much like the Her kind of vision.

Lenny Rachitsky

你会想那部电影吗？这种一直在的助手？

Do you think about that movie much—this idea of this always-there assistant?

Tibo Sottiaux

看过一遍。我更常想三四十年前的科幻有多前瞻——比如《Neuromancer》，还有原版《星际迷航》。星际迷航里技术非常有用：跟电脑说话它就做事，跟飞船说话它就动。这一幕正在成真——我们在进入那个时代。

I watched it once. What I think about more is sci-fi from 30–40 years ago and how visionary it was—books like Neuromancer, and the original Star Trek. There’s something about Star Trek: technology that was very useful; you talk to a computer and it does things; you talk to your spaceship and it moves. This is getting real. We’re entering that era.

Lenny Rachitsky

我们在现场录，门外排队等着见你。你怎么想这份责任——做出影响那么多人生活的东西？

We’re recording in front of a live audience; there was a line out the door. How do you think about this responsibility—building what you’re building and the impact on people’s lives?

Tibo Sottiaux

很重要成为社区的一部分、跟人一起建。我们在一起发现这门技术能做什么。每次跟人聊，都会被提醒：你在做一件我原先没料到的事，它以某种方式影响你、也影响别人。既谦卑又鼓舞，而且是做出真正对人有用的东西所必需的。没有社区我不知道怎么做。

It’s very important to be part of a community and to build with people. We’re discovering this technology and what we can do together. Every time I talk to people I’m reminded: you’re doing this thing which I hadn’t even anticipated; it’s impacting your life and this other person’s life. It’s both very humbling and very inspiring, and necessary to build something truly useful for humans. I don’t know how we would do it without the community.

Lenny Rachitsky

你怎么还有时间刷推、回推？

How do you find time to even tweet?

Tibo Sottiaux

大概每天半小时，滚一滚，有灵感就说点有意思的，经常是当下冒出来的。团队版多 Dot 还没放，但今天放的是主 Dot：可配置、熟悉 Dot 能做什么、接你的应用。很快就能加第二、第三、第四个，组你的虚拟团队。

I spend maybe half an hour per day—just scrolling, then doing something when I feel inspired and have something interesting to say. Often it comes in the moment. We haven’t released team Dots yet, but today we’re releasing the primary Dot which you can configure, get used to what a Dot can do, connect to your apps. Very quickly soon you’ll be able to add a second, third, fourth—as many as you need to create your virtual team.

Lenny Rachitsky

会给它们具体角色吗？

And you can give them specific roles?

Tibo Sottiaux

不一定非要；但有些活很重，比如帮我盯 Twitter，就值得整颗 Dot 专职干。你可以有好几颗这样的。

I don’t find it absolutely necessary, but sometimes I have a thing that requires quite a bit of work—such as monitoring Twitter for me—and that’s the work of an entire Dot. You can have a couple of those.

Lenny Rachitsky

今天这么多发布，Dots 最性感。有没有你觉得被低估、以后会很重要的？

Of all the launches today, Dots maybe feels sexiest. Is there something you think isn’t getting the attention it deserves that will become really important?

Tibo Sottiaux

黑马是生态——全面开放，以及我们对开放的深度承诺。ChatGPT 登录已有 16 家伙伴，我很骄傲。大概去年一时兴起：跟 Pi、OpenCode 的创建者聊，说当然该能用这套登录、用这套用量，虚拟上握个手、相信你们别搞鬼。后来变火，现在做成正式支持、很多伙伴。另一边是把 ChatGPT 的搭建基础设施打开：插件扩展、插件发现，让大家可能触达大约 12 亿用户。Keynote 没细讲的是分成：用量高、受欢迎的插件会拿到一部分收入分成。

The sleeper hit is ecosystem—opening it all up, and the deep commitment we have toward opening it all up. We have Sign-in with ChatGPT, 16 partners. Very proud of that. Last year it was sort of in the moment—I was talking to creators of Pi and OpenCode: of course you should be able to use this sign-on and the usage; we shook hands virtually and trusted you not to do anything shady. That became popular; now it’s an actual supported thing with many partners. The other side is opening up the infrastructure for how we build for ChatGPT—plugin extensions, plugin discovery—allowing everyone to ship something through maybe 1.2 billion users. We also have shared economics we didn’t talk about in the keynote: we’ll pay plugins that are popular and seeing a lot of usage—they get part of the revenue share.

Lenny Rachitsky

举例：有人用 Notion 或 Figma 插件，Notion、Figma 能从应用内用量赚钱？

As an example—someone using the Notion or Figma plugin within ChatGPT Work—Notion, Figma make money from people using it within the app?

Tibo Sottiaux

对。订阅用户有一定用量；当他们把用量花在插件上，或在别的产品里用 Sign in with ChatGPT，我们会有共享经济，让对方得到报酬。

Yes. Subscribers get a certain amount of usage, and when users use that usage with the plugin—or within the other product when they use Sign-in with ChatGPT—we’ll have shared economics where they get remunerated.

Lenny Rachitsky

想让插件被发现、爆起来，有什么建议？

Any advice for people that want their plugins to bubble up and get discovered?

Tibo Sottiaux

把插件做好。我们会调系统，但看的是留存、成功度、质量，再据此在对话里推荐。好插件可以推到很大一块用户；不好就会停推。

Build a good plugin. We’ll tune the system, but we look at retention numbers, how successful the plugin is, the quality—and that’s what we then start to recommend to users in conversations. Your plugin can get recommended to a significant slice of users. If it isn’t quite good, it’ll stop being recommended.

Lenny Rachitsky

所以不是 AEO 堆词，而是持续有用？

So it’s not so much AEO—get the right words—it’s people using it consistently?

Tibo Sottiaux

对。它是不是真的在加效用、让 ChatGPT 能多做一件事。

That’s right. Is it actually adding utility? Is it enabling ChatGPT to do something else?

Lenny Rachitsky

Dots 你们做了多久？为什么出得这么晚？还记得 Codex Cloud 吗？

When did you start working on Dots? Why did it take so long? Do you remember Codex Cloud?

Tibo Sottiaux

记得。大概一年前。回头看 Codex Cloud 用的那小段动画，我觉得那就是 Grok Bot 的灵感——非常像。认真说：长程、持久任务的研究我们做了两年多；连贯的记忆系统大概同样久，很多已直接进了 ChatGPT。很多人有轶事：ChatGPT 记得你很久以前烧烤时酸橙汁烫到手之类的细节——记忆已经很好用。这些叠在核心 harness 和 7×24 生产力上，就是 Dots。安全也很关键，所以用 Astra——我们最安全、最对齐的模型——在安全与安全防护上花了很多力气。

Yeah. That was like a year ago. If you look back at Codex Cloud and the little animation we used, I think that was the inspiration for Grok Bot—really similar. Seriously, on the research: long, persistent, long-horizon tasks—we’ve worked on that for more than two years. Memory systems that are coherent—roughly the same amount of time. We launched a lot of that straight into ChatGPT. So many people have an anecdote like falling ill and ChatGPT remembering the barbecue with lime and tequila and the burn on their hand—memory works very, very well. All of this coming together on top of the core harness and the ability to run 24/7 with good productivity is what we launched with Dots. Getting safety and security right mattered—that’s why we launched with Astra, our safest, most aligned model. A lot more effort went into making Dots really safe and secure.

Lenny Rachitsky

人脑长期还会在哪里有价值？你觉得自己在工作里哪里仍必要？

Where do you think human brains will continue to be valuable over time? Where do you find yourself being necessary versus where AI takes more?

Tibo Sottiaux

这很取决于我们怎么建技术。OpenAI 的设计是把人放中心，做成你意志与品味的延伸、真正赋能。只要继续这样，它就能把你想做的事、创造力和品味放大，像趁手的艺术工具。也许不再有「码农」，但会有比以往更多的 builder。人类想学习、想看别人在建什么——我坐在这儿是在跟你聊，不是在跟 Dot 聊。我觉得这会很久都还成立。

This is very much a function of how we build the technology. The way we design everything at OpenAI is to put humans at the center and build it as extensions of humans—of your will and your taste—a super empowering thing. As long as we continue to do that, it enables you to take whatever you wanted to do, your creativity and taste, and expand that in a way that feels awesome—really this artistic tool. We may not have coders anymore, but we have more builders than ever. There’s something that remains human about that. Humans want to learn and see what other humans are building. I’m talking to you, not to Dot. I think that will remain true for a very, very long time.

Lenny Rachitsky

工程师生活变了之后，上下文切换更多，也有孤独感——整天跟智能体聊而不是跟人。你怎么想这部分影响？

One thing that’s emerged with engineers is a lot more context switching, and a trend of loneliness—talking to agents all day instead of other humans. How much do you think about that impact?

Tibo Sottiaux

一直在想。减少配置疲劳是一块；减少「跟一个智能体单聊、再委派给一堆」这种独自冒险的感觉。很多东西会凑在一起，让体验更愉悦——团队花很多时间想这个。理想是就在物理空间里，像我们现在这样聊：它能观察、听见想法；你在纸上写两笔，它后台开建；你又有新主意，再开另一条；投到屏幕上继续跟人一起聊。没有「对着屏幕想怎么 prompt」的疲劳，变得超自然。今天没完全到，但朝那走。

All the time. Reducing configuration fatigue is one. Reducing the fact that talking to an agent is a solo adventure—just you and one agent, then many, delegating so many things. A lot of things are going to come together to make it much more delightful; I and the team spend a lot of time thinking about that. The ideal way to get things done would be in your physical space, having a conversation like we have now. It’s able to observe, hear the ideas. You jot something on paper; it starts building in the background; you have another idea; it starts another thing; project it on a screen and talk—evolving in conversation with other humans. You don’t have the fatigue of being on a screen thinking about prompting. It becomes super natural. We didn’t fully get there today, but we’ll get there.

Lenny Rachitsky

还能做更多，就有压力必须做更多——为什么不一次跑 30 个智能体？这有解吗，还是人性？

There’s this pressure to do more because we can—why aren’t you running 30 agents at once? Is that solvable, or human nature?

Tibo Sottiaux

Sam 也谈过这个承诺：降噪，让你把注意力花在想花的地方。很多事抢注意力，把音量拧低一点，让你聚焦。我每次度假彻底断联一周，就会开始用更不同、更创意的方式想问题。我很想看能不能把那种状态变成日常——也许少开会、少做一些，休息更好反而更高效。行业得一起想清楚；AI 的承诺不该是每秒多一条 prompt。

We heard Sam about the promise as well: reduce the noise and allow you to spend attention where you want. All these things vying for attention—tone that down and get you to focus. Every time I go on holiday and fully disconnect for a week, I start thinking in different and more creative ways. I’m eager to see if we can make that day-to-day—maybe fewer meetings, maybe you don’t need to do as much, and you’d be more productive if better rested. As an industry we’ll have to figure that out. The promise of AI isn’t just one more prompt per second.

Lenny Rachitsky

最近有没有特别惊艳的 AI 用例？

Is there anything you’ve done really interestingly with AI recently that blew your mind?

Tibo Sottiaux

现场 demo 其实刚挂过，没那么酷。但有意思的是：我的 Dot 意识到 ChatGPT 生产出了问题，在直播 demo 前五分钟 ping 了我——挺让人慌的。「生产挂了。要我试着修吗？」我说你还没到那步，谢谢你想试。然后我联系工程团队，他们后来修好了。它理解：有个叫 DevDay 的重要事、有生产系统、有直播 demo、大概会用这套生产、五分钟后开始——所以该提醒我。这点相当惊人。

The live demo just failed in front of everyone, so not so cool. But actually my Dot realized ChatGPT production went down and pinged me five minutes before the live demo—quite stressful. “Production is down. Do you want me to try and fix it?” I was like, I don’t think you’re there yet, little Dot, but thank you for trying. Then I got in touch with the engineering teams and they fixed it. Having this thing that understands there’s an important thing called DevDay, a production system, a live demo probably using that system, and it’s happening in five minutes—so I should ping him—I think that’s quite remarkable.

Lenny Rachitsky

你怎么看把这些权限都给它？它能碰生产代码库吗？

How do you feel about giving access to all these things? Does it have access to the production codebase?

Tibo Sottiaux

对部分生产系统有权限，但有护栏。我们谈过专家 Dot：额外护栏、额外监控，跑在自有硬件上——有些其实跑在 Mac mini 上。Dots 的根本不同是：harness 不跑在你那台机器上。很多人没意识到——它可以连你想连的任意多台设备。它有自己的电脑，也能连你的笔记本；以后也许连十台，像章鱼一样全控。

It has access to some production systems with guardrails. Specialist Dots are operated with additional guardrails, additional monitoring, on their own hardware—we actually run some on Mac minis. The fundamental difference in how we’ve built Dots is the harness does not run on the machine. What people may not have realized: it can connect to as many devices as you want. It has its own computer; you can connect it to your own laptop as well. Over time you might connect it to like 10 different devices and it can control them all—a little bit like an octopus.

Lenny Rachitsky

招人时，哪些技能在涨、哪些在跌？

In your interviewing, what skills are trending up, and what are trending down?

Tibo Sottiaux

下跌的是打字快——没那么有用了。上涨的是好品味、想用户、连上你在为谁建。很多前创始人现在特别成功；OpenAI 招人也能感到——我们已经有 120 多个前 YC 创始人，像个超级大创业公司。想做出真正要紧的东西、知道什么叫好，比以往更重要。角色也在模糊：也许你以前只做设计或只做工程，现在是你发光的时候。

The skill trending down is typing fast—not that useful anymore. Skills trending up: great taste, thinking about the user, connecting to the audience you’re building for. A lot of ex-founders are incredibly successful right now; we’re finding that when hiring at OpenAI—we have more than 120 ex-YC founders now. It’s like this mega startup. This passion and envy to build something that matters, and know what good looks like, is more important than ever. Roles are blurring—you might have been only into design or only an engineer; now it’s your time to shine.

Lenny Rachitsky

你会想念纯工程的那部分吗？

Do you miss the engineering part of the job?

Tibo Sottiaux

想念过一阵；现在速度真的快起来，那种感觉又回来了。我很想尽量把这摊民主化。要普及到十亿级用户还要一阵；但有了超快速度，六个月 Sole 成本大概和 Astra 差不多，进展连我们内部都吃惊。尤其能语音操控时，你会回到很有创造力的心流——跟过去那种心流不一样，但我不怀念旧的。

I missed it for a while, and now with really fast speeds I’m finding that back. I’m really excited to democratize that as much as possible. It’ll take a while to be ubiquitous for more than a billion users, but with ultra-fast speeds we’re able to achieve it—a six-month Sole costs roughly the same as Astra. The progress is astonishing even for us internally. Especially when you can voice-control it, you get back into a really creative state of mind and endless flow. It’s different from the flow of the past, but I don’t miss it.

Lenny Rachitsky

给应届生、职涯早期的人什么建议？很多人想当下一个 Tibo。

Advice for new grads, people early in their career? A lot of people want to be the next Thibault.

Tibo Sottiaux

下一个 Tibo 已经在 OpenAI 了，叫 Ahmed Ibrahim。他入职时是应届生，现在负责整个 OpenAI 算力舰队和应用侧，责任极大，也建了很多 Codex harness。让他突出的是：特别善良、特别会协作，总是优先解真正重要的问题而不是把自己放第一；而且是海绵，学得极快，用最新技术把学习速度推到我从没见过的程度。几个月到一年里进步惊人，现在很多最难啃的发布我都非常信任他。

The next Thibault is already at OpenAI—his name is Ahmed Ibrahim. He was a new grad when hired; he’s now responsible for all of OpenAI’s compute fleet and applied—a massive responsibility. He’s built a lot of the Codex harness. What made Ahmed stand out is he’s incredibly kind, incredibly collaborative, and always tries to solve the problem that is really important without putting himself first. And he is just a sponge—learning so fast about everything, using all the latest technology to learn at speeds I haven’t seen before. In a couple of months and a year his progress was fascinating. He’s now someone I trust a ton with the most gnarly launches.

Lenny Rachitsky

在 OpenAI 工作，有什么会让外人吃惊？

What would surprise people about what it’s like to work inside OpenAI?

Tibo Sottiaux

也许不惊人，但确实是大量前创始人在做自下而上的点子。比如 Decisions API：很快拼起来。我们发现 Luna 模型很好，做点约束采样，就能以不同于 Responses API 的形态上线，某些场景更快、更爽。掀开引擎盖：一个 Slack 频道，最初四个人周末黑客，然后像公司食堂一样谁都能碰，大家兴奋起来，加视觉输入，觉得比外面强，越来越多人加入。感觉完全无结构，却一路推到上线；我们尽量把质量门槛拉高，有些东西今天本可以发、我们又按住多烤一会儿——下周、再下周还有发布。自下而上的能量，再有人站出来把能量导入正轨。

I don’t know if that’s surprising, but as I said, a lot of ex-founders doing a lot of bottoms-up, exciting ideas. Decisions API came together very quickly. We realized we have a really good model with Luna; we can do some constraint sampling and ship it as a different shape to the Responses API—faster, quite delightful for some cases. Peel under the hood: a Slack channel; initially four people hacking on a weekend; then it was made as accessible as company food; people got excited, started to build; “we can support visual inputs, this is even better than whatever is out there”; more people pitching in. It feels completely unstructured, then take it all the way and ship. We try to hold a high bar for quality and stop things before they go out, or they need more baking. Many more things we were maybe going to ship today at DevDay—we held them back a little and spaced them out. More launches next week and the week after. Incredible bottoms-up energy, and people assuming roles to channel that productively.

Lenny Rachitsky

OpenAI 很独特的一点是你好像有很大自治——随时按重置、在 Twitter 上直接说。怎么谈这种信任？

Something unique about OpenAI is it feels like you have a lot of autonomy—pressing the reset button whenever you want, saying things on Twitter. Can you speak to that cultural philosophy?

Tibo Sottiaux

自治多，就要扛住。我们信任人做正确选择；出事就快修，或从错误里学。到目前为止很有效，是很赋能的环境。必要时或觉得对的时候我可以按重置，这是特权，也让我能以别的方式贴近社区——不必过一层层审批。

You have a lot of autonomy and then you have to own up to it. We trust people to make great choices. When things go wrong, fix it quickly or learn from the mistake. So far so good—it’s a very empowering environment. It’s true I can press the reset button whenever necessary or whenever it feels right. That’s a privilege, and it allows me to be close to the community in ways that wouldn’t otherwise be possible. I don’t have to run it through an echelon of approvals.

Lenny Rachitsky

这么大自治下，你自己搞砸过什么？

Is there something you screwed up with that amount of autonomy and trust?

Tibo Sottiaux

有时我本可以让团队更有动力去做不那么复杂的东西，而不是继续堆复杂度。Codex 早期我造成过几次故障——入职大概第三天就把生产搞挂了，居然还没被开。我学到了教训。

At times I could have made teams more inspired to build things that are less complex, and just really continue to strive toward simplicity versus complexity. Early days in Codex we had a couple of outages I caused. Very early for me at OpenAI I took production down like day three. I was still employed after that. I learned my lesson.

Lenny Rachitsky

人们还没定价进去的、方向性变化是什么？

What do you think people aren’t pricing in in terms of where things are going?

Tibo Sottiaux

好多。多数互联网上的动作将由智能体完成。模型会以相当惊人的速率变得更便宜、更快。我们终于能把各模态无缝拼在一起。看外面在建的东西，我常觉得你们差一点就懂了——如果硬逼自己想象这一切一年后大概好十倍，你会用完全不同的方式建。

There are so many things that are not yet priced in. The majority of actions on the internet will be taken by agents. Models are going to become cheaper and faster at rates that are quite incredible. We will finally be able to integrate all modalities together in a way that is very seamless. Often when I look at what people are building out there, you’re not quite getting it. If you were just pushing yourself and really imagining all of this being roughly 10 times better than it is today in a year, you would build in a different way.

Lenny Rachitsky

智能体成为主要流量之后，二阶效应呢？

Do you see second-order effects from a world where most actions—most traffic—come from agents?

Tibo Sottiaux

很多。首先，产品要对智能体成功，就得按那个量级建。我们跟 Notion 走得很近：他们做 MCP 之后，突然对大量能干活的智能体开放，流量暴涨，系统和对账都承压。做产品的人会卡在「要不要做人机界面」——可以拖一阵，但不可避免：多数东西会被智能体用。反过来，为人类做真正爽的、吃满多模态的新体验，我觉得投资严重不足。

There are many. First, if you want your product to be successful for agents, you have to build for a certain level of scale. We’ve worked closely with Notion; when they built their MCP, suddenly it’s available to all these agents that can actually do the work—they saw a ton of traffic come in. That puts strain on the system; you have to figure out the economics. There’s tension if you’re building products: do you build an interface or not? You can hold that back for a while, but it is inevitable—the majority of things are going to be used by agents. Building toward that future is very important. On the flip side, building absolutely delightful new experiences for humans that really benefit from all the modalities is something I feel is underinvested in.

Lenny Rachitsky

过去一年你改观过什么？

Is there something you’ve changed your mind about in the past year?

Tibo Sottiaux

我本以为要再过一两年才能到今天这种模型能力（比如 Astra），所以修正了先验。我也没料到自己会这么依赖语音——大量靠听写或直接打电话给智能体，那种感觉好太多。招人上：我低估了更年轻一代的能量与人才，以及他们会最先拥抱变化、想清楚怎么驾驭这一切——我也改了招人策略。新人职场其实有优势：没有旧工作方式的包袱，可以 all-in AI，吸收和学习快得惊人。

I was expecting us to reach the level of capabilities of the models we have today—say Astra—maybe in a year or two, so I had to revise my priors. I also did not expect to rely so much on voice; I do so much through dictation or just calling my agent, and I hadn’t anticipated how much better that feels. On hiring: I hadn’t realized the incredible talent and energy younger generations have, and how they would be the ones embracing all this change first and figuring out how to harness it—so I had to change my opinion on hiring strategy there. New people in the workforce actually have an advantage because they haven’t worked in a certain way and can go all in on AI—and the ability to absorb and learn super fast.

Lenny Rachitsky

模型变强很快。你们放慢了一些训练、在「pace the frontier」。怎么看风险与对齐？

Models are getting really smart fast. You’ve slowed down some training; everyone’s pacing the frontier. How do you think about AI risk?

Tibo Sottiaux

对我来说 pace the frontier 就是提前投、而且投得比我们以为「该投」的还多：对齐与安全、安全防护、护栏。我们越来越多算力花在二级监控上——主智能体干活的算力之外，还有监控它的算力：别采取过高风险动作，发现可能被 prompt 注入就打断干预。API 栈上很大一部分投入其实在安全栈。你要极度 hardened。我们还没放出 Astra 之上的下一档能力；放出的是接近 Astra 智力、但高效得多的档，我们觉得舒服。会继续大笔投入。新闻里也有：我们有 6.1 Astra 但没发——这点我很骄傲。

For me, pacing the frontier is really investing ahead—and investing way more than we even think we ought to. Investing in alignment and safety of the models, security, guardrails. We’re spending more and more compute on secondary monitoring: compute going to the primary system—the agent doing work—and compute monitoring the primary agent so it’s not taking too high-risk actions, interrupting if it looks like it’s getting prompt-injected. The majority of our investment on the API stack is going into the safety stack. You want to make sure you’re extremely hardened. We haven’t yet released the next step up in capability beyond Astra. We’ve released a level similar, near Astra intelligence, but much more efficient, and we feel very comfortable about that. We’ll continue to invest a ton. It was also in the news that we had 6.1 Astra and then we didn’t release it—and that is something I’m very proud of.

Lenny Rachitsky

所以你乐观我们能解决对齐、不会搞砸？

So you’re optimistic we’ll solve alignment and AI doing really bad things?

Tibo Sottiaux

是。也感到深深的责任。想想 OpenAI 的激励：我们为大约 12 亿人做产品，不能搞砸。给这么多人、包括并不技术的人手里塞什么，必须做对。我们不在那上面赌博。

Yes. And it’s also a deep responsibility. If you think about the incentives for OpenAI: we build products for 1.2 billion people. We can’t screw that up. What we put in the hands of so many people, including people who are not very technical themselves—you have to get it right. We don’t gamble there.

Lenny Rachitsky

最后一问：当前产品状态里有什么最烦你、巴不得马上修掉的？

Final question: what’s something that annoys you about the current state of the app that you’re like, goddamn it, we got to fix that?

Tibo Sottiaux

模型几乎到了，但还没到让应用几乎完全消失的程度。我等不及把体验压到极简本质。我自己都会被模型选择器、推理力度、要不要多智能体或 Ultra、它到底干什么搞得很累——简直需要模型选择器博士。我想尽快把这些全干掉。

I feel like the models are almost there, but not quite for the app to almost completely disappear. I can’t wait to get to this super essence of simplicity. I myself even get fatigued with the model picker and the reasoning efforts and whether to use multi-agent or Ultra or what it even does—you kind of need a PhD in model pickers. I just want to get rid of all of that as quickly as we can.

Lenny Rachitsky

Tibo，谢谢。

Tibo, thanks for doing this.

Tibo Sottiaux

谢谢邀请。

Thanks for having me.