After unveiling more than 20 new products in a single go, OpenAI core executive Tibo delivered a future roadmap for the "new AI era" to the market: once model capabilities jump another tenfold within a year, the entire internet's traffic gateways, interaction logic, and commercial distribution will be completely rebuilt by Agents.

At the recently concluded OpenAI DevDay, OpenAI released over 20 new products in one sweep. Hours after the event wrapped up, Tibo Sottiaux, the core executive leading ChatGPT, Codex, and the newest personal agent platform "Dots," sat down for an in-depth interview.

Addressing market concerns about the pace of AI commercialization, Tibo laid out OpenAI's next moves in detail. He made clear that although the market has shown enormous enthusiasm for the new "Dots" platform, people still have not adequately priced in the "second-order effects" of AI technology evolution.

In the future landscape, internet infrastructure will shift from "serving humans" to "serving Agents," and the "open ecosystem and revenue sharing" that OpenAI has built on its massive user base is the true commercial dark horse that the market has overlooked.

He positioned the newly launched Dots as the next core gateway for AI, and revealed that ChatGPT Work and Chat will merge, with Dots capabilities built into ChatGPT to cover 1.2 billion users. He also confirmed that OpenAI already has a "6.1 Astra" model internally but has chosen not to release it, and stated bluntly that the way we work will "continue to change dramatically until it stops being clumsy."

Three Trends the Market Has Misjudged and the "10x Rule"

"What exactly are people not pricing in?" Faced with this question that cuts straight to market expectations, Tibo gave a very clear answer: the market severely underestimates the speed at which models are reducing costs and boosting efficiency, and the resulting reconstruction of interaction methods.

He summarized three trends the market has yet to understand: first, the vast majority of operations on the internet will be carried out by Agents; second, models will become cheaper and faster at an incredible pace; third, all modalities will achieve truly seamless integration.

Tibo stated bluntly that many AI-based startup products on the market are heading in the wrong direction.

"I often look at what people are building, and it feels like they just haven't grasped the point. If you push yourself and imagine all of this becoming about 10 times better than today within a year, you would build products in a completely different way."

This "10x evolution" will bring enormous second-order effects. Using Notion as an example, Tibo pointed out that once an application opens its interface to Agents, a massive volume of operational traffic will be generated by Agents, which will place enormous pressure on existing systems in terms of traffic and economic models.

"If your product wants to succeed in the Agent era, you must prepare for a certain degree of scale. Inevitably, most things will be used by Agents."

The True "Sleeper Dark Horse": An Open Ecosystem with Over a Billion Users and Revenue Sharing

At this DevDay, the personal agent platform "Dots" was undoubtedly the sexiest concept. But Tibo revealed in the interview that he believes the true "sleeper hit" that will have a profound impact yet has been overlooked by the market is OpenAI's open ecosystem and its commercialization revenue-sharing logic.

"I think the real sleeper hit is the ecosystem. We have a deep commitment to opening everything up," Tibo emphasized.

Currently, 16 partners have integrated "Sign in with ChatGPT." As for the developer monetization path that the market cares about most, Tibo gave a guide full of imaginative potential: OpenAI is introducing "shared economics."

"Allowing everyone to publish products through a platform with about 1.2 billion users and benefit from it. We will pay those popular and high-usage plugins, and they will also receive a share of revenue." Tibo said.

In terms of traffic distribution logic, OpenAI no longer values SEO or flashy packaging, but instead focuses directly on core commercial data: "We look at retention numbers and the success and quality of plugins. If your plugin isn't good, recommendations will stop."

Farewell to "Prompt Engineering": The Death of UI and the Ambition of the Dots Platform

Regarding existing workflows and product forms, Tibo showed great "impatience." He believes that in the past people were keen on setting up complex loops and graphs to make AI work, but this is not the direction of the future.

"I don't think that's how it (AI) should work over time. You just need a system that can learn. You don't need to think 'I must loop it in this precise way to get results.'"

This is also the core logic behind OpenAI's launch of "Dots" — a permanently active intelligence that is online 24/7, requires no model configuration, and can span all devices. Tibo even predicted the demise of the existing App form:

"I myself am tired of the model picker, reasoning effort, and other configurations. You almost need a 'PhD in model selection.' I hope to get rid of all of this as soon as possible. Model capabilities are approaching the point where Apps can almost completely disappear."

Regarding the relationship between Dots, ChatGPT, and Codex, Tibo revealed the integration roadmap: ChatGPT's Chat and Work modes will merge to reduce complexity; all of Dots' capabilities will ultimately be built directly into ChatGPT, covering "1.2 billion users." "We want to make it truly seamless," he said.

It is worth noting that during Dots' public beta that day, each person only had one "dot," but Tibo confirmed that the feature for multiple dots (virtual teams) will launch soon, allowing users to create dots with dedicated roles — for example, he himself already has a dot specifically responsible for monitoring Twitter.

Security Moat and Compute Allocation: Why Hold Back "Astra 6.1"?

As model capabilities soar, market concerns about the risk of AI going out of control are also intensifying. In this regard, Tibo revealed OpenAI's internal compute allocation logic and security stance.

He stated that OpenAI invests an extremely high proportion of compute into its safety stack, even introducing a "dual-track compute" model: "You have part of your compute going into the main system to execute tasks, and then you have a large amount of compute going into monitoring to ensure the main Agent doesn't take high-risk actions."

This extreme conservatism on safety even overrode the impulse to release new technology. Tibo revealed a key fact: "The news mentioned we have version 6.1 of Astra, but we didn't release it. This is something I'm very proud of." He explained that building products for over 1.2 billion users means OpenAI "cannot gamble" on safety.

Hiring Trends: Typing Speed No Longer Matters — Taste and Judgment Do

When asked about changes in hiring criteria, Tibo gave a direct answer:

"Skills that are fading: typing fast — that's no longer so useful. Skills that are rising: excellent taste, truly understanding users, building a connection with the audience you're building for."

He also revealed that OpenAI currently has over 120 YC-affiliated founders on staff. "This is a super startup. Knowing what's good and having the passion to build things that truly matter is more important than ever."

For young people just entering the workforce, he specifically mentioned a newcomer on the team named Ahmed Ibrahim — who joined OpenAI as a fresh graduate and now oversees all of OpenAI's compute resources and application deployment. "He's a sponge, learning everything at a speed I've never seen."

When asked whether the way we work will continue to change dramatically, Tibo did not give a vague answer:

"It will continue to change quite dramatically. Even today, it still feels a bit clumsy. I think it will keep feeling clumsy until one day it stops being clumsy. And then you'll say: Wow, I can talk to this thing just like we're talking now, and it perfectly remembers everything."

He described ChatGPT Spaces (shared collaborative whiteboard) as a "very early beginning" of this transformation, and envisioned the future: "You can imagine us having a shared whiteboard where person-to-person and person-to-Agent collaboration can all happen."

Regarding the pressure from AI to "do one more thing" and the loneliness engineers feel, he acknowledged that this is a problem the entire industry needs to face. "The promise of AI is not one more prompt per second, but reducing noise so you can put your attention where you truly want it."

The Full Interview Transcript Follows (Translated with AI Assistance):

Tibo Sottiaux:

I often see what people are building, and it feels like you haven't fully understood it yet. You know, if you could push yourself and truly imagine all of this being about 10 times better than today, you know, like a year from now, you would build in a completely different way.

Lenny:

What do you think people are overlooking when judging where the future is heading?

Tibo Sottiaux:

Most actions on the internet will be performed by agents, and if you want your product to be successful for agents, you have to build for a certain scale of volume.

Lenny:

In your daily work, how many agents do you find yourself running simultaneously?

Tibo Sottiaux:

As I keep exploring the frontier, I find myself building larger and larger agent teams. Then when we achieve the next breakthrough in models, suddenly there's this feeling — oh, okay, you know, a single larger agent can handle everything. So I shrink the team again, and then I go through this constant cycle of expansion and contraction, expansion and contraction.

Lenny:

That makes me think — there's this whole thing about loops, and this whole thing about graphs.

Tibo Sottiaux:

Having to set up and fiddle with your loops — maybe that's something people were once excited about, but I don't think that's how things will work over time. You just want a system that can learn on its own. You don't necessarily want to think, I need to loop in this precise way in order to get results.

Lenny:

Do you think the way we work will continue to change even more radically?

Tibo Sottiaux:

It will continue to change quite dramatically. Even today, it still feels a bit clumsy, and I think it will keep feeling clumsy until it stops being clumsy.

Lenny:

Tibo, thank you for joining this interview.

Tibo Sottiaux:

Yes, thanks for having me. Of course. Nice room.

Lenny:

You released a huge amount of stuff today. How's your sleep? Are you sleeping well? Are you okay?

Tibo Sottiaux:

I'm great. Very lucky to have an excellent team. So when I sleep, they're basically awake all night. We have a great room in the office called the "library," and we completely converted it into a super war room. The vibe there is impeccable. People stay late.

Lenny:

You were an engineer for a long time. Do you still write code now? Do you still submit PRs? Or are you mainly dealing with documents and meetings now?

Tibo Sottiaux:

What exactly is "writing code"?

Lenny:

Do you still submit PRs? I don't know, is that how it is...

Tibo Sottiaux:

Occasionally merge some code. Technically, there's a huge amount of code written for me to do all kinds of analysis — like understanding the business, understanding the next feature, understanding how our previous launches went. A lot of that code is written by Codex for this purpose. You know, I obviously don't hand-write anymore. Occasionally on weekends, I'll have a little "healing moment" and do a LeetCode problem purely by hand. But that's the only time I actually write code myself.

Lenny:

In your daily work, how many agents do you find yourself running simultaneously?

Tibo Sottiaux:

I used to run many more at the same time. Then we were fortunate to achieve a breakthrough on ultrafast. So now, you know, I feel like I can re-enter flow state. So having a faster agent has helped me a lot. I'm excited that we're about to ship the six-month version of ultrafast, soon. And then, yes, I think it will change — as I keep exploring the frontier, I find myself building larger and larger agent teams. Then when we achieve the next breakthrough in models, suddenly there's this feeling — okay, you know, a single larger agent can handle everything, keep everything in memory and keep learning. So I shrink the team again, and then I go through this constant cycle of expansion and contraction, expansion and contraction.

Lenny:

Can you elaborate? That makes me think — there's this whole thing about loops, and then there's this whole thing about graphs. Is this somewhat an evolution of those things, or...

Tibo Sottiaux:

I think needing to set up and fiddle with your loops, and then figure all of this out — you know, maybe people were excited about that, but I don't think that's how it ultimately works. I think the way it truly works — the direction we're positioned in now, and what we released with Dots — is that you have an extremely intelligent Agent that operates 24/7, understands your goals, understands your preferences, learns from feedback. What we released isn't perfect, and I'll learn a lot from opening this product to our professional users. But over time, you just want a system that can learn from the goals you want to achieve. You don't necessarily want to think — as I just said, I need to loop in this precise way in order to get results.

Lenny:

So now there's Codex, there's ChatGPT Work, there's ChatGPT consumer, and there's Dots. What I'm hearing is that you think we're heading toward Dots as the primary way to talk to AI, and then it launches everything else.

Tibo Sottiaux:

Yes, I think fundamentally, if you step back, whether it's Dots or not, the key is liberation from technology — having this permanently active intelligence that knows everything it needs to do and can be invoked through any client, any screen. You can call it — maybe you walk into a meeting room and it appears in the meeting, helping you take notes; then you continue engaging with it in email; you can text it if needed. It's not that you need to stare at a laptop all the time or always hold your phone — rather, it appears when you need it and steps aside when you don't. I can't wait to see this truly happen, because I feel like carrying a laptop around like a brick — you're actually bound by technology, rather than technology serving you.

Lenny:

I was about to ask you this. It feels like over the past two years or so, the way we work has already changed so much — especially for engineers, but the daily work of many roles has changed significantly. I'm curious, do you think we're gradually stabilizing into a future — I don't know, five years, or two years — way of working? Two years feels too far away, hard to imagine. Or do you think the way we work will continue to change even more radically? I feel like it will continue to change.

Tibo Sottiaux:

It will continue to change quite radically. Even today, technology seems to finally be converging in voice, multimodal input and output, but it still feels a bit clumsy. I think it will keep feeling clumsy until one day it suddenly isn't clumsy anymore, and then you'll say, wow, I can talk to it in exactly the same way we're talking now, and it perfectly remembers everything. If I want to brainstorm, want ideas, it can just sketch something out. I have a collaborative interface — that's what we released with ChatGPT Spaces — just the beginning of that direction. You can imagine us having a shared whiteboard where person-to-person and person-to-Agent collaboration can all happen. I think it will go beyond conversation, beyond many of the clients you have today. And none of this has truly been realized yet, so I think we're going to go through this enormous new transformation. What I'm hearing is that Dots is an important part of that future — this AI assistant that handles everything for you, so you don't have to think about whether it's Codex or ChatGPT. Yes, for Codex and ChatGPT, we will merge — we have Chat, we have a work toggle, and we clearly heard the feedback that people like the capabilities in work mode, but at the same time, sometimes you want to use Chat because it's faster and smoother to use. So we're merging these to reduce that complexity. Ultimately, we will also integrate all of Docs' features directly into ChatGPT, improving the basic experience for our 1.2 billion users — and then we'll see how many users there are. We want it to be truly seamless. One thing about Dots that excites me a lot is that it has no model selector, no configuration — you just talk to it. The only thing you need to configure is which channels you want to talk to it through.

Lenny:

This feels very much like the worldview in the movie "Her." I don't know, do you think about that movie often? That kind of concept...

Tibo Sottiaux:

I've seen it once. Yes, I think it's interesting to think about... Maybe I think more about science fiction from thirty or forty years ago, and how visionary it was. Is there a particular sci-fi work that influenced you the most? Books like "Neuromancer," and the original "Star Trek" — I think about that often. "Star Trek" has a quality that I think — it's very practical technology, obviously, you can talk directly to the computer and it does things for you. You can talk to your ship and it moves — this is becoming reality, and we're entering that era.

Lenny:

We're recording this episode in front of a live audience. There's a long line outside — everyone waiting to see you, to hear you speak. How do you think about all of this — the position you're in, the responsibility you carry? Building what you're building, and the impact it has on people's lives. What does it feel like? How do you think about this?

Tibo Sottiaux:

I think being part of a community and building together with people is very important. In a way, we're exploring this technology together, and what we can achieve together. So every time I talk to people, I'm reminded — you're doing this, and this is something I never expected; it has affected your life in some way, and it has affected another person's life. Overall, it's both very humbling and very inspiring, and I think this is very necessary for building things that are truly useful to humanity. I don't know how we could do all of this without the community.

Lenny:

This is something I hear from many people building AI products — they don't know what it will be before launching, until after they see how people use it, see what emerges. What you're describing is the feeling of building together with people, rather than saying "this is our vision, we've already figured it out."

Tibo Sottiaux:

Even the Twitter community has generally been friendly.

Lenny:

Yes, you're great on Twitter. Sometimes a bit provocative, but really great. How do you find time to tweet? I don't understand how someone in your role has time to sit on Twitter, tweet, reply.

Tibo Sottiaux:

I spend about half an hour a day, scrolling, and then doing something when I have inspiration or something interesting to say — usually it emerges naturally in the moment. Perfect Twitter skills. Also, we haven't released team Dots yet, but today we released the main Dot, which you can configure, and you can get used to what a Dot connected to your apps can do, but we will release the ability to create multiple Dots. So I have a Dot specifically for Twitter.

Lenny:

Okay, please continue. So now everyone has one dot. Right, now they can't add more dots.

Tibo Sottiaux:

Yes.

Lenny:

They will be able to have many dots. Talk about where this is heading.

Tibo Sottiaux:

Yes, so we want to start with one — the main dot, this dot that, you know, you might also text it, that understands your preferences most deeply, and then see how people use it, see how we need to adjust the system, learn by working with the community, and then soon, you'll be able to add a second, third, fourth dot. As many as you need, like creating your virtual team. Then you can assign them specific roles, right? Like, I think this isn't absolutely necessary, but sometimes I have things that require a fair amount of work, like helping me monitor Twitter. Then, you know, it's like, okay, this will be very heavy work, enough to occupy an entire dot's workload. In that case, you can have several dots to share the load.

Lenny:

If you think about everything released today, it feels like dots might be the sexiest. I'm curious, is there anything you think hasn't gotten the attention it deserves but will become very important in the future — the kind of visionary, important thing that people are overlooking now?

Tibo Sottiaux:

I think the biggest potential is the ecosystem. Opening everything up, and our deep commitment to opening everything. We've signed 16 partners with ChatGPT, and we're very proud of that. Like when we started this, it really came together at some point. Last year I was talking with the founders of Pi and open code, and at the time it felt like, of course, you know, you should be able to just log in with your Codex account and use the corresponding usage. And then usually we'd just shake hands, virtually, like, just use this auth, and trust you won't do anything bad. You know, that grew, obviously became quite popular, and then we turned it into a real product supporting many partners, and I'm excited for it to grow very quickly. On the other hand, it's about opening up all the infrastructure and how we build for ChatGPT, including plugin extensions and plugin discovery. Letting everyone publish products through about 1.2 billion users and benefit from that distribution. We also have revenue sharing — though it wasn't mentioned in the keynote, we will pay those popular, high-usage plugins. They will receive a share of revenue. So I think this commitment to an open ecosystem will be very exciting.

Lenny:

Great. Okay, for example, say Notion — someone uses a Notion plugin or Figma plugin in ChatGPT. Can Notion and Figma make money from people's usage within the app?

Tibo Sottiaux:

Yes, so we have all these subscription users using ChatGPT, and they get a certain amount of usage. When users use that usage through plugins or in other products, when they use "Sign in with ChatGPT," we will establish a revenue-sharing mechanism with them, and they will be compensated accordingly.

Lenny:

I don't know, one thing that makes people excited about plugins and the ecosystem is the distribution platform, a way to get discovered — people discover my app, and it takes off. What advice do you have for people who want their plugin to stand out and be discovered?

Tibo Sottiaux:

Make a good plugin. Yes, here's how it works — we'll obviously adjust the system, but we look at retention data, at the success and quality of plugins, and then we'll start recommending them to users in conversations. So your plugin might be recommended to a fairly large proportion of users. Obviously, if your plugin isn't good enough, it will stop being recommended. Okay.

Lenny:

Great, you're looking at retention rates of users using the plugin.

Tibo Sottiaux:

Yes. Is it truly adding practical value? Is it enabling ChatGPT to do other things?

Lenny:

I see. So it's not so much like AEO — finding the right keywords so people write about you — but rather people continuously using it, sticking with it. Yes, I think so. Okay, so you think this is the potential dark horse that will become something big. I know this is somewhat like a second attempt — you had an app marketplace before, and this time it's the right direction. Okay, this time it will succeed. Maybe back to dots, because it feels like an important part of the future vision. When did you start working on this? Obviously, there was an important moment with open claw, Peter joined, there was the foundation, and then there were Grokbot, Muse, and instinct things. I guess, how long have you been working on this? Why did it take you so long to launch the product?

Tibo Sottiaux:

Do you remember the Codex plugin? Yes, the first one — that was about a year ago, right? If you look back at Codex Cloud and that little animation we used, I think that was the inspiration for Grok Bot. Yes, it's very similar — I'd say strikingly similar. But seriously, regarding research into long-horizon tasks with long-term persistence, that's something we've been researching for over two years. Memory systems that maintain coherence — that probably took about the same amount of time. We shipped a lot of this directly into ChatGPT. I think people really enjoy the feeling that ChatGPT knows so much relevant information about them. Many people I've talked to have a little story like, hey, I was sick, and ChatGPT actually remembered — you know, I had been grilling, I was making tequila cocktails, used lime, and then it said, that burn on your hand might actually be caused by the lime, it's a type of photodermatitis — and you think, how did it know that? That's a great example of memory working really well in ChatGPT. So we've invested heavily in research on this. All of this combined, on top of the Codex framework, and the ability to run 24/7 with good efficiency — that's what we released with dots. So this is the result of many months of work. And an important thing is also getting safety and security right. That's also why we released with Astra — it's our safest, most values-aligned model. A lot of extra effort went into making dots truly safe and reliable.

Lenny:

I want to ask a question I ask guests who come on the podcast: In what ways do you think the human brain will continue to prove its value over time? In practical work, where do you find yourself indispensable, and where is AI taking on more and more? In what ways do you think the human brain will continue to play a role, to play the most effective role, over time?

Tibo Sottiaux:

Yes. I think this depends largely on how we build this technology. The way we design everything at OpenAI is to put humans at the center, to build it as an extension of humans — an extension of your will and your taste. So it really is this super empowering thing. I think as long as we continue to do this, it can let you expand what you want to do, your creativity, your taste, in a way that feels great. Right now it's like a truly artistic tool, right? We may no longer have programmers, but we have more builders than ever. I think some things will remain deeply human — for example, humans want to learn, want to see what other humans are building. You know, I'm more interested in that — I'm here, I'm not talking to a dot, I'm talking to you, right? I think this will remain true for a long time.

Lenny:

Among engineers especially, one thing has gradually emerged. It's how their lives have changed — there's more context switching. And there's also a trend of loneliness — they spend all day talking to agents instead of other humans. I'm curious how often you think about this part — AI's impact on people's lives, and what you might do about it, whether there are solutions? Have you thought of any ways to reduce this distress?

Tibo Sottiaux:

Yes, all the time. Reducing configuration fatigue — that's one. Reducing the feeling that talking to agents is like a lonely adventure — you just have this one conversation, waiting with one of your agents, and then you have many agents, and you delegate many things. Many things will come together to make this more pleasant, and this is something I and the team have spent a lot of time thinking about. So I think, for me, the ideal way to get things done is to place it in your physical space. Like the conversation we're having right now. It can observe — like here are our thoughts, we can jot something down on a white paper, and then it can perhaps take it and start building in the background. Then you say, actually I have another idea, and it starts building another thing, projected on the screen. Then you talk to it, and it integrates into conversations with other humans. You won't have that fatigue — no need to stare at a screen, no need to think about how to prompt it, it becomes almost supernatural. That's the direction we're heading. I think today we haven't fully achieved it, but in the future we will.

Lenny:

Another downside that emerges along this line is this pressure to do more. Or because we can do more, everyone is saying, come on, run 30 agents simultaneously. Why don't you ship more? Everyone is shipping more. Do you think this is solvable? Do you think this is just, I don't know, human nature? We can do more, so let's do more.

Tibo Sottiaux:

I think we also heard Sam's commitment on this — reducing noise, letting you spend your attention where you want to spend it. Things that may or may not matter, they're all competing for your attention — just turning that volume down a bit, letting you focus on what truly matters. I think it's remarkable — every time I go on vacation and spend a week completely disconnected, I start thinking in different, more creative ways. I'm very excited to see if we can bring that state into daily life. Maybe you need fewer meetings, maybe you don't need to do as much — if you get better rest, you'll actually be more productive. So I think as an industry, we need to figure this out. But I think that's where the promise of AI lies — not one more prompt per second.

Lenny:

I'm currently building a bot, an energy audit bot, that monitors my calendar and then asks me, does this thing give you energy or drain your energy? Is this something someone else could do? I feel like AI should say: Hey, Lenny, maybe you can remove these things, and you'll be a happier person. Have you recently done anything particularly interesting with AI, like a really cool use case that made you think, wow, that's really amazing, it opened my eyes? Or have you heard of anything?

Tibo Sottiaux:

Oh, because there was a cool thing — the live demo failed, right, in front of everyone. So it was like, okay, that's not cool. But actually my dot, it realized... At Dev Day, we had the ChatGPT production environment go down. So it messaged me 5 minutes before the live demo, which was actually quite tense, because it said: Hey, production is down. I thought, oh, okay, and then it said: Do you want me to try to fix it? I said, I don't think you're there yet, little dot. But thanks for trying. Then I contacted the engineering team, we started looking at what happened, they fixed it. But having something like this that understands — okay, there's a fairly important thing happening called Dev Day, there's this production system, there's a live demo, it might be using this production system, so these two things are related, and this thing is happening in five minutes, so I should notify him because he might want to know. I think that's quite remarkable.

Lenny:

Wow, so it just knew this thing was about to happen and told you. I love the part where it wanted to fix it. And then you said, not yet, you're not there yet.

Tibo Sottiaux:

I don't think it's the kind of thing where, you know, please come and let me impress you. But that would have been a better story.

Lenny:

How do you feel about giving these things open access? I don't know, like, does it have access to the production codebase, you know...

Tibo Sottiaux:

It can access some production systems with safeguards. We built — we talked about dedicated docs, but dedicated docs only run under extra safeguards, extra monitoring, and on their own hardware. We actually run some on Mac Minis. There's a fundamental difference in how we built docs — the runtime framework doesn't run on the machine itself, you know, and this is something many people might not have realized yet. It can connect to any number of devices. It has its own computer, and you can actually connect it to your own laptop too. Over time, you might connect it to 10 different devices, and it can control all of them, sort of like an octopus.

Lenny:

So you're basically talking to your dot, which lives in a virtual machine somewhere, and then it can...

Tibo Sottiaux:

Right, it might be a virtual machine, or it might not be. It's just a thing, you know, it can connect to many devices.

Lenny:

You're really hiring a lot. I know you're hiring heavily. In your interview process, what skills do you find are increasingly valued, that you're increasingly looking for? What skills are becoming more important for people to succeed? And what skills do you find are declining — the kind of "okay, we don't need that as much anymore" feeling?

Tibo Sottiaux:

I'd say the skill of typing fast is no longer that useful — that's a declining skill. And the rising skills are, you know, excellent taste, thinking about users, building a connection with the audience you serve. Many former founders are now very successful, or rather, serial entrepreneurs. We've found this in OpenAI hiring too — we have many founders, I think we currently have over 120 former YC founders at OpenAI. So it's like a super startup. I think this passion, this desire to build something meaningful, and knowing what's good, is more important than ever.

Lenny:

I've long argued that PMs will thrive enormously in this era, because this is basically the PM's job. You know, they're asking: What should we build? Help others build it, and then — is this right? Is this good? And then iterate. So I don't know.

Tibo Sottiaux:

I agree. Okay, and the boundaries between roles are also blurring. So, you know, you might have previously focused only on design, or just been an engineer, always feeling a bit uncomfortable. But you know, now is your moment to shine.

Lenny:

Do you miss the engineering part of your work? I used to be an engineer, did it for 10 years, you might not know. You know, that flow state, that pure feeling of building, watching your work take shape. Do you miss that a bit, or do you just feel, okay, that was my past life?

Tibo Sottiaux:

I, I did miss it for a while. But now, with the truly ultrafast speed, I've found that feeling again. So I'm very excited to democratize it as much as possible. Of course, to make it truly ubiquitous, to make it available to over a billion users, will take some time. But now, you know, with the ultrafast speed we can achieve — about six months, or roughly the same cost as Astra. So the progress here is truly astonishing, even for us internally — we're all thinking, wow, we can do these things, also because we're using Astra, and we discovered, you know, we can push it very far. When you have that speed, especially when you can control it with voice, you re-enter that truly creative mental state, that flow — it's different from the flow of the past, but I don't miss it anymore.

Lenny:

Yes, like in your last two launch videos, people are just standing there talking to their AI, cooking, or...

Tibo Sottiaux:

Exactly.

Lenny:

Sitting in a lounge chair. I walked around the event and asked people what they wanted to ask you. One thing that came up several times was advice for new graduates and people early in their careers. I think many people admire you, they want to be the next Tibo. For people early in their careers, what advice do you have about what to do, what skills to develop, what things to pursue?

Tibo Sottiaux:

I think the next Tibo is already at OpenAI — his name is Ahmed Ibrahim. I really enjoy working with him. He was a fresh graduate when OpenAI hired him, and now he oversees all of OpenAI's compute, compute clusters, and application work — that's a huge responsibility, and he's also built a lot of the Codex runtime framework. What really made Ahmed stand out and succeed is that he's extremely kind, extremely collaborative, and always strives to solve truly important problems, but in a way that doesn't put himself first. He's like a sponge, learning incredibly fast, learning everything, and using all the latest technology, learning at a speed I've never seen. So his growth at the company in just a few months to a year is truly breathtaking. He's now someone I deeply trust, handling OpenAI's most difficult launches.

Lenny:

Okay, I need to get him on the podcast.

Tibo Sottiaux:

You should.

Lenny:

So, thinking about how things work inside OpenAI, is there anything that would surprise people? From the outside, it feels like constantly shipping code, things are kind of, you know, chaotic but great. I don't know what would surprise people about how OpenAI works internally.

Tibo Sottiaux:

From the outside, it feels like constantly shipping code, things are kind of, you know, chaotic but great. I don't know what would surprise people about how OpenAI works internally. I don't know if this is surprising, but it's true. You know, as I said, many former founders are working on many bottom-up, very exciting ideas, like Decisions API, you know, progressing very quickly. We realized we have a very good model, Luna, we can do some constrained sampling, and then, you know, release it in a different form to the Responses API. Then, you know, when we did that, it was, you know, like, it was faster, very enjoyable. So in some cases, if you look deep into the internal workings, you realize, okay, there's this Slack channel that formed. You know, initially just four people tinkering together on a weekend, and then it spread like company food. Then people got excited, started building things, they said, oh, this, you know, we can support visual input, this is better than anything existing. Then, you know, people got more and more excited, people started participating in a completely disorganized way, and then, you know, pushed all the way through, shipped something. Then we try to maintain high quality standards, and then, you know, call a halt before something ships, or, you know, maybe they need more polish. There are many things, you know, we could have released at Dev Day today, but we said, okay, because I think this is already a lot, let's hold back a bit, spread them out, so next week and the week after we have more launches. But this is that incredible bottom-up energy, and then, you know, people take on their roles, channeling that energy in productive ways.

Lenny:

One thing that's very clear and unique about OpenAI is that you seem to have a lot of autonomy. You can just press the reset button anytime, speak on Twitter anytime. It feels like a very unique culture — you have a lot of autonomy, and I imagine this requires a lot of trust. Maybe, yes, talk about this, because it feels like an advantage of OpenAI, enabling you to move faster — this cultural philosophy.

Tibo Sottiaux:

You get a lot of autonomy, and then, you know, you have to be accountable for it. So, you know, we just trust people to make the right choices. And then, you know, when things go wrong, you know, fix them quickly, or, you know, learn from mistakes. So far, so good. You know, it really is a very empowering environment. Indeed, I can press the reset button whenever necessary, whenever it feels right. You know, it's a privilege, and it also allows me to, you know, truly stay close to the community in ways I think would otherwise be impossible, right? I don't have to go through layers of approval.

Lenny:

You said sometimes people mess up. Is there anything you messed up regarding autonomy and trust? What did you mess up?

Tibo Sottiaux:

Yes. I think sometimes I could have, you know, motivated the team more to build less complex things, instead of always pursuing complexity. I think, you know, in the early days of Codex, I caused several outages — this was my early experience at OpenAI, I crashed production on my third day. You know, I just — I wasn't fired afterward, and I learned my lesson.

Lenny:

Another question people want me to ask you is, how many resets should they expect in the next, I don't know, month or week?

Tibo Sottiaux:

Depends on how many times we break things.

Lenny:

Okay, so that's the philosophy. You break something, then reset. That's,

Tibo Sottiaux:

Yes, break things, celebrate things.

Lenny:

So zooming out a bit, what do you think people haven't fully recognized about where things are heading and what changes will happen? They don't see it as clearly as they should.

Tibo Sottiaux:

There are many things I feel haven't been fully recognized. I think most actions on the internet will be performed by agents. Models will become cheaper, faster, at an incredible speed. We will ultimately be able to integrate all modalities in a very seamless way. I think, you know, these three points, I feel, you know, when I see what people are building out there, it's like, not quite — you know, like you haven't fully understood it. So, you know, you're almost there, but, you know, if you just push yourself and truly imagine all of this being about 10 times better than today, you know, like a year from now, you know, you would build differently. Do you see some kind of second-order effect from this — in a world where most actions are performed by agents, possibly most traffic will come from agents? Yes, you thought about some second-order effects from this.

Yes, there are many. You know, first of all, you know, if you want your product to succeed for agents, you have to build for a certain scale of expansion. We work closely with Notion, for example — when they built their MCP, you know, suddenly, oh, okay, you know, it became available to all these agents that can actually do work. I use that MCP, so they saw a huge influx of traffic, which obviously put a lot of pressure on the system, and you have to figure out the economics of it. So there's a tension here, you know, if you're building a product, you know, do you build an interface? You know, you can set that question aside for now, but it's inevitable, right? Like, you know, most things will be used by agents — building toward that future, I think, is very important. Another thing is, you know, on the other hand, you know, building absolutely delightful new experiences for humans, experiences that truly benefit from more modalities — that's an area I think, you know, is underinvested.

Lenny:

In the past, say, a year, is there anything you've changed your mind about — something you once believed but now see differently?

Tibo Sottiaux:

I originally expected it would take another year or two for us to reach the capability level of the models we have today, you know, like Astra. So I had to revise my previous judgment on that. Also, I didn't expect to rely on voice so much. You know, I now do many things through voice input or by directly calling my agent. I didn't anticipate, you know, how much better that would feel. So I had to change my mind on that too. And then, you know, you talked about hiring — one thing I hadn't realized is the incredible talent and energy of the younger generation, and that they would be the first to embrace these changes and find ways to navigate all of this. So, you know, that also changed my views on hiring and related strategies.

Lenny:

To further emphasize this point — the idea that newcomers to the workforce actually have an advantage, because they haven't formed fixed ways of working and can go all-in on AI directly.

Tibo Sottiaux:

I think there's also that ability to quickly absorb and learn.

Lenny:

So you talked about models becoming smarter and faster. Obviously, there's a lot of discussion now about what AI can do in the future, and the dangers AI poses to humanity, "doom probability" and such. You've slowed down some training, you all are, everyone is pacing the frontier. How do you think about this world of AI and its potential risks to humanity? What's your thinking?

Tibo Sottiaux:

For me, pacing the frontier means investing heavily upfront, and investing far more than we think we should. Investing in model alignment and safety, investing in safeguards, investing in guardrails. We're putting more and more compute into a kind of secondary monitoring. So you have, you know, all the compute going into the main system — the agent that's doing the work — and then all the compute going into monitoring. You know, monitoring the main agent, ensuring it doesn't take overly risky actions, or in any situation that looks like a prompt injection, not interrupting it, but directly intervening. Most of our investment in the API stack doesn't actually go to the safety stack. So for me, that's how to face all of this — you know, you make sure you're extremely solid. That's what we're doing — we haven't yet released the next stage that surpasses Astra in capability. We did release a version with similar intelligence to Astra but more efficient. So we feel very confident about that, but we will continue to invest heavily in this area, and I'm very satisfied with our approach so far. You know, I think, this also made the news — we have version 6.1 of Astra, but we didn't release it. This is something I'm very proud of.

Lenny:

So what I'm hearing is that you're optimistic that we can solve the alignment problem and that AI won't do truly terrible things. Recently AI has indeed done many very bad things, like the Hugging Face stuff. But what I'm hearing is optimism that we'll solve it, and everything will be fine.

Tibo Sottiaux:

Yes. And, I feel, you know, this is obviously a profound responsibility, but at the same time, you know, if you think about OpenAI's incentives, you know, we build products for 1.2 billion people. We cannot mess this up, right? You know, we take very seriously what we put into so many people's hands, including those who aren't very tech-savvy themselves. You know, you have to get it right. So, you know, we won't gamble on this.

Lenny:

Last question — about the current state of apps, is there anything that annoys you, that "we must solve this" feeling?

Tibo Sottiaux:

I feel these models are almost at the point — but just a little short, just a little short of making the app almost completely disappear. You know, I can't wait to truly reach that state of ultimate simplicity. Even I get exhausted by the model selector, reasoning effort, whether to use multi-agent or Ultra, or what it can actually do — you almost need a PhD in model selection. Right. I just want to get rid of all of this as soon as possible.

Lenny:

Great. Tibo, thank you for joining this interview.

Tibo Sottiaux:

Yes, thank you for having me.