OpenAI's Tibo Sottiaux on Dots, Agents, and Making the AI App Disappear
Lenny's Podcast
Tibo Sottiaux leads ChatGPT and Codex at OpenAI. Lenny Rachitsky interviewed him in front of a live audience at DevDay, a few hours after his team launched more than 20 products. The conversation kept returning to one question: what happens when AI stops being a tool you open and configure, and becomes a persistent agent that works alongside you? Sottiaux's view is that today's agent workflows still feel clunky. Hand-tuned loops, model pickers, and juggling several products are a passing phase. He expects them to be replaced by a single learning agent that is available on any screen and steps back when it isn't needed. Dots, OpenAI's new personal agent platform, is his first concrete step in that direction.
How a product leader works with agents now
Sottiaux has been an engineer for a long time, but he says he no longer writes code by hand, apart from the occasional "therapeutic" coding exercise on weekends. A lot of code still gets written on his behalf, mostly for analysis: understanding trends, understanding the business, working out the next feature, and tracking how earlier launches performed. Codex writes much of it for his Dot.
Asked how many agents he runs in parallel, he said the number goes up and down. He used to run many more at once. A breakthrough in "ultrafast" inference changed that, because one fast agent lets him get back into a flow state. He also described a recurring pattern. When he pushes against the frontier, he finds himself building larger and larger teams of agents. Then the next model breakthrough arrives, and one bigger agent can do all of the work, keep everything in memory, and learn, so he shrinks the team again. He described his setup as cycling between expansion and contraction.
Why loops and graphs are a passing phase
Lenny asked whether this cycle relates to the current enthusiasm for agent "loops" and "graphs," where people wire up and tune their own agent workflows. Sottiaux doesn't think that approach will last. He granted that people have been excited about setting up and fiddling with loops. In his view, though, the future looks like what OpenAI is positioning with Dots: a very smart agent that runs 24/7, understands your goals and preferences, and learns from feedback. Users shouldn't have to think about looping something in exactly the right way to get good results. They should have a system that learns from what they are trying to achieve.
He was open about the fact that the first release is not finished. He said OpenAI expects to learn a great deal from putting Dots in front of Pro users.
One persistent intelligence across every screen
Lenny summarized the current lineup as Codex, ChatGPT Work, consumer ChatGPT, and now Dots, and asked whether Dots is meant to become the main way people talk to AI. Sottiaux put it more broadly. Whether or not it ends up being Dots, he said, the goal is to break free from the technology and have a "permanent active intelligence" that knows what it needs to do and can be reached from any client or screen. He gave an example. You walk into a meeting room and the agent joins the meeting and takes notes. You pick the thread back up over email. You text it when you need something. It is there when you need it and stays out of the way when you don't. He contrasted this with carrying a laptop around "as this brick," which he described as being tied to the technology instead of the technology working for you.
On product structure, he said OpenAI heard the feedback about the separate chat mode and work toggle in ChatGPT. People like the capabilities of Work but sometimes prefer chat because it is faster and more pleasant to use. OpenAI is merging the two to cut complexity. He also said that eventually all of Dots' capabilities will ship inside ChatGPT, raising the baseline for what he called 1.2 billion users. One thing he likes about Dots is that it has no model picker and essentially no configuration. The only thing a user has to set up is which channels they want to talk to it through.
Work will keep changing, and it still feels clunky
Lenny asked whether work is settling into a stable shape after two years of upheaval. Sottiaux said no. He expects work to keep changing radically. Voice and multimodal input and output are finally coming together, he said, but the experience still feels clunky, "and I think it will feel clunky until it isn't." He described the point he expects to reach: talking to the system the way he and Lenny were talking on stage, with it remembering things perfectly, and being able to sketch ideas on a shared collaborative surface where humans and agents work together. He mentioned ChatGPT Spaces as a step in this direction. He thinks this will go beyond today's conversational interfaces and many of today's clients, and he said plainly that "none of this is just really working today."
When Lenny compared the vision to the film Her, Sottiaux said he had watched it once. He said he thinks more about older science fiction, particularly Neuromancer and the original Star Trek. What stays with him from Star Trek is how useful its technology was: you could talk to the computer, or the ship, and it did things. He said that era is now becoming real.
Building with the community, and one Dot per job
Asked how he thinks about the responsibility of his role, Sottiaux talked about community. Every conversation with users shows him uses he hadn't anticipated and ways the technology is affecting people's lives. He called this humbling, inspiring, and necessary for building something truly useful, and said he doesn't know how OpenAI would do it without the community. He spends about half an hour a day on Twitter, scrolling and posting when he has something worth saying.
That led to a detail about Dots' roadmap. At launch, each user gets one "primary" Dot. It is the one you are most likely to text, it learns your preferences most deeply, and it can be connected to your apps. OpenAI wanted to start with a single Dot so it could watch how people use it and tune the system. Soon users will be able to add more Dots and build a virtual team with specific roles. Sottiaux said he doesn't consider this strictly necessary, but some jobs are big enough to warrant a dedicated Dot. He already has one responsible for monitoring Twitter.
The sleeper hit: an open ecosystem with revenue sharing
Lenny asked which launch people were underrating. Sottiaux said the ecosystem, and OpenAI's commitment to opening up its platform. He said Sign in with ChatGPT launched with 16 partners. It started informally about a year earlier, when he was talking with the creators of Pi and OpenCode. The idea was that they should simply be able to let users sign in with their OpenAI account and use their existing usage allowance. In his words, the two sides "kind of shook hands" virtually, and OpenAI trusted them not to do anything shady. That arrangement became popular, and it is now an officially supported program with many partners that he wants to grow quickly.
The other part is opening up the infrastructure OpenAI uses to build ChatGPT, through plugin extensions and plugin discovery, so that anyone can ship to ChatGPT's user base and benefit from that distribution. Sottiaux also mentioned something not covered in the keynote: shared economics. Popular plugins with heavy usage will receive part of the revenue. Using Lenny's examples of Notion and Figma, he explained that ChatGPT subscribers get a certain amount of usage. When they spend that usage through a plugin, or inside another product via Sign in with ChatGPT, the partner shares in the economics.
Asked how to get a plugin discovered, his first answer was "Build a good plugin." He said the system will be tuned over time, but OpenAI currently looks at retention numbers and at how successful and high-quality a plugin is, meaning whether it actually adds utility and lets ChatGPT do something new. Those signals decide which plugins get recommended to users inside conversations, potentially reaching a significant share of users. Plugins that aren't good stop being recommended. Lenny noted that OpenAI had tried an app marketplace before. Sottiaux said this one is "the right one."
What went into Dots
Lenny pointed to the wave of persistent personal agents that came before Dots, including the OpenClaw moment (whose creator, Peter Steinberger, joined OpenAI), Grok Bot, Muse, and Instinct, and asked why OpenAI took so long. Sottiaux joked that the animation OpenAI used for Codex Cloud about a year ago looks very similar to Grok Bot's. He then gave a more serious answer. OpenAI has worked on research into long, persistent, long-horizon tasks for more than two years, and on memory systems for about as long. Much of the memory work already shipped in ChatGPT. He said many people tell him stories about ChatGPT remembering something relevant. One example: someone who had been at a barbecue making drinks with lime later noticed a burn on their hand, and ChatGPT connected it to the lime.
Dots combines that memory research with the agent harness and the ability to run 24/7 productively. Sottiaux described it as many months of work. Safety and security were a major part of that work. He said this is why Dots launched on Astra, which he called OpenAI's safest and most aligned model, and that much more effort went into making Dots safe and secure.
The outage warning five minutes before the demo
Asked for a striking use case, Sottiaux told a story about that day's keynote, where a live demo had failed on stage. His Dot noticed that ChatGPT production had gone down and pinged him five minutes before the demo. He said this was stressful. The Dot offered to try fixing it, and he declined, telling it he didn't think it was "there yet." He then contacted the engineering teams, and the problem was fixed by the time of the interview. What impressed him was the reasoning the Dot showed. It understood that DevDay was an important event, that a live demo was coming up, that the demo probably depended on the production system that was down, and that the timing meant he would want to know right away.
On permissions, he said his Dot has access to some production systems, with guardrails. Specialist Dots run with extra guardrails and monitoring, on their own hardware. OpenAI runs some on Mac minis. He highlighted an architectural point he thinks people have missed: the Dots harness doesn't run on the user's machine. A Dot has its own computer, can also connect to your laptop, and over time might connect to and control ten different devices, "a little bit like an octopus." When Lenny asked whether it lives in a VM, Sottiaux said it might or might not. It's "a thing" that can connect to many devices.
Fatigue, loneliness, and the pressure to do more
Lenny raised some downsides that have shown up among engineers: more context switching, a sense of loneliness from talking to agents all day instead of people, and pressure to run dozens of agents and ship more simply because it's possible. Sottiaux said his team thinks about this "all the time." The specific goals he named were reducing configuration fatigue and making agent work less of a "solo adventure," where one person delegates to many agents in isolation. His ideal is an agent present in the physical space. It hears ideas during a conversation between people, picks up a sketch on paper, starts building in the background, starts on another thing when someone has a new idea, and projects its work onto a screen so that it evolves as part of a group conversation instead of a prompting session. He said OpenAI didn't fully get there with this launch.
On the pressure to do more, he pointed to what he said Sam Altman had described: AI should reduce noise and let people put their attention where they want it, instead of flooding them with competing demands. He said that whenever he takes a week off and fully disconnects, he starts thinking in different and more creative ways. He wants to know whether AI can make that state part of everyday life, with fewer meetings, less to do, and more productivity because people are better rested. He said the industry still has to figure this out, but he sees that as the promise of AI, not "one more prompt per second."
Skills that are trending up and down
Asked about hiring, Sottiaux said typing fast is a skill that matters less. The skills that matter more are great taste, thinking about the user, and connecting with the audience you're building for. He said ex-founders and repeat founders are doing especially well, and that OpenAI has more than 120 former YC founders on staff, which makes it feel like a "mega startup." What he values is passion for building something that matters and knowing what good looks like. Lenny said he has long argued that PMs will thrive for this reason, and Sottiaux agreed. He added that roles are blurring, so people who felt boxed in as only a designer or only an engineer now have their "time to shine."
Asked whether he misses engineering, he said he did for a while but has found the flow state again thanks to very fast inference. He said the latest ultrafast speeds cost roughly the same as Astra, and that this progress surprised even people inside OpenAI. With that speed, and especially with voice control, he gets back into a creative, "endless flow." He said it is a different flow from the old one, but he doesn't miss the old one.
For people early in their careers, he pointed to a colleague, Ahmed Ibrahim, whom he called "the next Thibault." Ibrahim joined OpenAI as a new grad, has done significant work on the Codex harness, and now has what Sottiaux called the massive responsibility of OpenAI's compute fleet and applied work. Sottiaux credited his rise to being very kind and collaborative, focusing on the problems that matter without putting himself first, and being "a sponge" who uses the latest tools to learn faster than anyone Sottiaux has seen. Within a year or so, Sottiaux said, Ibrahim became someone he trusts with OpenAI's hardest launches.
Inside OpenAI: bottom-up projects, autonomy, and resets
Describing how OpenAI works, Sottiaux emphasized bottom-up energy, much of it from former founders. His example was what the transcript renders as a "decisions API." Teams realized that a strong model, Luna, combined with constrained sampling, could be exposed in a different shape from the Responses API, and that it was faster and pleasant to use in some cases. It began with about four people hacking on it over a weekend in a Slack channel. It was then made available internally, people got excited, added features such as visual input support, and eventually shipped it. He said the process feels completely unstructured, but OpenAI tries to hold a high quality bar and will stop or delay things that need more work. Several launches planned for DevDay were held back and will roll out over the following weeks.
On autonomy, he said people are trusted to make good choices, own the results, and fix problems quickly or learn from them when things go wrong. He sees being able to hit the usage-limit "reset" button whenever it feels right, without going through layers of approvals, as a privilege that keeps him close to the community. Asked about his own mistakes, he said he could have pushed teams harder to build less complex things. He also admitted that he caused a couple of outages in the early days of Codex and took production down on his third day at OpenAI. Asked how many resets users should expect, he said it depends on how often things break, and added that resets also happen to celebrate things.
What people aren't pricing in
Sottiaux named three things he thinks people underestimate. First, most actions on the internet will be taken by agents. Second, models will get cheaper and faster at "quite incredible" rates. Third, all modalities will finally be integrated seamlessly. He said that when he looks at what people are building, they often seem "almost there" but are not quite getting it. If they imagined everything being roughly ten times better within a year, they would build differently.
On second-order effects of an agent-dominated internet, he said products have to be built for a certain level of scale. He cited Notion, which OpenAI worked closely with. Once Notion shipped its MCP server, it became available to many agents and saw a large surge in traffic, which strained its systems and forced it to work out the economics. He described a tension for product builders over whether to expose an agent interface at all. He thinks companies can hold off for a while, but the shift is inevitable. He also argued for the opposite direction: he thinks delightful new experiences for humans that make use of multiple modalities are underinvested in.
What he changed his mind about
Sottiaux said he had expected models at today's capability level, such as Astra, to arrive a year or two later, so he had to update his priors. He didn't expect to rely so much on voice. He now does much of his work through dictation or by calling his agent, and it feels far better than he anticipated. He also underestimated the talent and energy of younger generations, who he now thinks will be the first to embrace these changes and figure out how to use them, partly because they absorb and learn so quickly. He said this changed his thinking on hiring strategy.
Safety: "pacing the frontier"
Lenny noted that OpenAI has slowed some training and asked about AI risk, mentioning the recent Hugging Face incident as an example of AI doing harmful things. Sottiaux said that to him, "pacing the frontier" means investing ahead, and more than OpenAI thinks it needs to, in alignment, security, and guardrails. He described a growing share of compute going to secondary monitoring. One body of compute runs the primary agent doing the work. Another monitors that agent, checks that it isn't taking overly risky actions, and intervenes if it appears to be prompt-injected. He also referred to the share of API-stack investment going into the safety stack, though the transcript is garbled at that point.
He said OpenAI has not released anything beyond Astra's capability level. It has released a model close to Astra's intelligence that is much more efficient, which he is comfortable with. He said he is proud that OpenAI built a more capable model, reported in the news, and chose not to release it. Asked whether he is optimistic that alignment problems will be solved, he said yes. He framed it as a deep responsibility and also a matter of incentives: OpenAI builds products for 1.2 billion people, including many non-technical users, and "we can't screw that up." He said OpenAI does not gamble there.
The thing he most wants to fix
Lenny's last question was what annoys Sottiaux most about the current app. He said the models are almost, but not quite, good enough for the app to nearly disappear. He gets fatigued himself by the model picker, the reasoning-effort settings, and choosing between multi-agent and ultra modes, without always knowing what they do. He said you "kind of need a PhD in model pickers," and that he wants to get rid of all of it as quickly as possible, moving toward what he called a "super essence of simplicity."
Often when I look at what people are building out there, it's just like, "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.
What do you think people aren't pricing in in terms of where things are going?
The majority of actions on the internet will be taken by agents. If you want your product to be successful for agents, you have to build for a certain level of scale.
How many agents do you find yourself running in parallel in your day-to-day work?
As I'm pushing the frontier, I find myself building larger and larger teams of agents. And then when we have the next breakthrough with models, suddenly I'm like, "Oh, well, a bigger agent can just do all of it." And so I kind of shrink the team again. And then it just kind of goes through this expansion and shrinking expansion.
It kind of makes me think about this whole loops thing and then this whole graphs thing.
Having to set up and fiddle with your loops is something that maybe people got excited about, but I don't think this is the way that it's going to work. Over time, you just want a system that learns. You don't want to necessarily think about how, "I'm going to loop it exactly this way in order to get good results."
There's still more radical change happening to how we work?
To continue to change quite radically. Even today, it still feels a bit clunky, and I think it will feel clunky until it isn't.
Tibo, thank you for doing this.
Yeah. Big fan of you.
Of course.
Nice room.
You launched a ton of stuff today. How did you sleep? How are you sleeping? Are you doing all right?
I'm doing great. I'm fortunate to have a very good team.
So, 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. And the vibes in there are just immaculate. People were there until very, very late.
So, you've been a long-time engineer. Are you coding at all anymore? Are you shipping PRs? Are you mostly in docs and in meetings now?
What is coding?
Exactly. Yeah. Are you shipping PRs? I don't know. Is that—
Yeah, yeah, yeah. Merging some code every once in a while. Technically, a lot of code is getting written for me to do all sorts of kinds of analysis: understand trends, understand a business, understand the next feature, how well our previous launches are doing. A lot of code is written by Codex for Dot. I'm obviously not writing it by hand anymore. Occasionally, during the weekends, I'll have a little therapeutic moment where I do a little LeetCode just by hand. But that's the only actual hand-coding that I do.
How many agents do you find yourself running in parallel in your day-to-day work?
I used to run a lot more in parallel, and then we were fortunate enough to get a breakthrough with ultrafast. And so now I feel like I'm able to be in the flow again, and so just having a faster agent helps me a lot. I'm excited that we're getting 6.1 Sol ultrafast out there soon. And then, yeah, I think it sort of varies. As I'm kind of pushing the frontier, I find myself building larger and larger teams of agents. And then, when we have the next breakthrough with models, suddenly I am like, "Oh, well, a bigger agent can just do all of it and keep everything in memory and learn." And so I kind of shrink the team again. And then it just kind of goes through this expansion and shrinking expansion.
Okay. Can you say more about that? It kind of makes me think about this whole loops thing and then this whole graphs thing. Is it an evolution of that in some way, or—
I think having to set up and fiddle with your loops and figuring that out is something that maybe people got excited about, but I don't think this is the way that it's going to work. I think the way that it's going to work is how we're positioning and what we're shipping with Dots, where you have an incredibly smart agent that works 24/7, understands your goals, understands your preferences, learns from feedback. And what we launched is not perfect. We'll learn very much from making this available to our Pro users. But over time, you just want a system that learns from what you want to achieve. You don't want to necessarily think about, "I'm going to loop it exactly this way in order to get results."
So, right now there's Codex, there's ChatGPT Work, there's ChatGPT consumer, there's Dots. What I'm hearing is you think we're heading towards Dots kind of being the primary way you talk to AI, and that kicks off all these other things.
Yeah, I think fundamentally, if you take a step back, whether it's Dots or not, it's all about breaking free from the technology and having this sort of permanent active intelligence that knows everything it needs to do and is available through any client, any screen. You can call it. Maybe you walk into a meeting room, it shows up in the meeting, it takes notes, you pick it back up on email, and then you text it if you need it. It's not that you need to be glued to a laptop or you're always on your phone. It's available when you need it, and then it also gets out of the way when you don't need it. And I can't wait for that to come to fruition because I think just carrying your laptop as this brick around everywhere, you're kind of tied to the technology instead of the technology working for you.
This episode is brought to you by our season's presenting sponsor, WorkOS. What do OpenAI, Anthropic, Cursor, Replit, Sierra, Clay, and hundreds of other winning companies all have in common? They are all powered by WorkOS. If you're building a product for the enterprise, you've felt the pain of integrating single sign-on, SCIM, RBAC, audit logs, and other features required by large companies. WorkOS turns those deal blockers into drop-in APIs with a modern developer platform built specifically for B2B SaaS.
Literally every startup that I'm an investor in that starts to expand upmarket ends up working with WorkOS. And that's because they are the best. Whether you are a seed-stage startup trying to land your first enterprise customer or a unicorn expanding globally, WorkOS is the fastest path to becoming enterprise-ready and unblocking growth. It's essentially Stripe for enterprise features.
Visit workos.com to get started or just hit up their Slack, where they have actual engineers waiting to answer your questions. WorkOS allows you to build faster with delightful APIs, comprehensive docs, and a smooth developer experience. Go to workos.com to make your app enterprise-ready today.
I was going to ask you about this. It feels like so much has changed in how we work over the past 2 years. In 2 years, especially engineers, but a lot of roles, your day-to-day life has changed significantly. And I'm curious if you think we're kind of settling into what work will look like in the next, I don't know, 5 years? Let's say 2 years. That feels way too far to think about. Or do you think there's still more radical change happening to how we work?
No, I think it's going to continue to change quite radically. Even today, it does seem like the technology is finally coming together with voice, multimodal inputs, outputs, but it still feels a bit clunky, and I think it will feel clunky until it isn't. And then you're just like, "Wow, I can just talk to this thing the same way that we're talking right now. It just remembers things perfectly. If I want to ideate and brainstorm, I can just squiggle a thing. I have this collaborative surface." This is what we launched with ChatGPT spaces signal to various startup dot. You can imagine we have a shared whiteboard, we can collaborate between humans, between agents. And I think it's going to transcend conversations, transcend a lot of the clients that you have today. And none of this is really working today, right? So I think that we're going to go through this big new transformation.
What I'm hearing here is Dots is a big part of that future, this AI assistant that just kind of does everything for you, and you don't have to think about Codex versus ChatGPT and—
Yes, and for Codex and ChatGPT, we're going to merge. We have chat, we have the work toggle. We heard the feedback loud and clear. People love the capability in work, but then also sometimes you want to use chat because it's a little bit faster and more pleasant to use. And so we're merging that. We're reducing that complexity.
Eventually, we will ship all of the capabilities of Dots straight into ChatGPT as well and lift the floor for our 1.2 billion users, which, at that time, we'll see how many we have. And we want it to be really, really seamless. One thing that I'm very excited with Dots is they don't have a model picker. There's just no configuration. You just talk to it. The only thing that you have to configure is which channels you want to talk to.
It feels very much like the Her kind of vision of the world. I don't know. Do you think about that movie much? Just this idea of this—
I watched it once. Yeah. I think it's interesting to think about. Maybe what I think about more is sci-fi from 30, 40 years ago and how visionary it was.
Is there specific sci-fi you think about that's most influenced you?
Books like Neuromancer and then the original Star Trek. I think about that a lot. And there's something about Star Trek which I think was this technology that was very, very useful. Obviously, you were able to just talk to a computer and it does things for you. You can talk to your spaceship and it just moves, right? This is getting real. We're entering that era.
So we're recording this in front of a live audience. There was this line out the door, just people waiting to see you, to hear from you. How do you think about this role you're in, this responsibility you have to build what you're building and the impact it has on people's lives? What does that feel like? How do you think about that?
I think 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. And so every time I talk to people, I'm kind of reminded of, "You're doing this thing which I hadn't even anticipated." It's impacting your life in a certain way, and it's impacted this other person's life, and then it's all together. It's both very humbling, very inspiring, and I think very necessary in order to build something that is truly useful for humans. I don't know how we would do it without the community.
That's something I hear a lot with people building in AI, is they don't know what it is until they've launched it and they see how people use it, see what emerges. And it feels like kind of what you're saying is just build it with people versus, "Here's the vision, we've got it figured out."
Even the Twitter community has been very nice in general.
You're very good at Twitter. Kind of antagonistic sometimes, but very good. How do you find time to even tweet? I don't understand how someone in your role has time to sit on Twitter, tweet, reply.
I spend maybe half an hour per day. Yeah, just scrolling and then doing something when I feel inspired and I have something interesting to say. Often it just comes in the moment. Yeah.
Perfect Twitter skill.
And I have a... We haven't released the team Dots yet, but today we're releasing the primary Dot, which you can configure. You can get used to what a Dot can do, connect it to your apps. But we'll release the ability to create more than one. And so I have one that's responsible for Twitter.
Okay, say more about that. So right now everyone gets one Dot.
Yeah.
Right now they can't add more Dots.
That's right.
They will be able to have many Dots. Talk about where that's going.
Yeah, so we wanted to start with one, the primary Dot, which is the one that you'll probably also text, that learns your preferences most deeply as well, and kind of see what people do with it, see how we need to tune the system, learn that working with the community. And then very quickly, soon, you'll be able to add a second one, a third one, a fourth one, as many as you need to create your virtual team. And then you can give them specific roles, right? 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 then that's a thing where this will be so much work that it's the work of an entire Dot. And then you can have a couple of those.
So if you think about all the launches that happened today, it feels like Dots may be the sexiest. I'm curious if there's something that you think is not getting the attention it probably deserves that will become really, really important down the road, that people are sleeping on, that's a vision.
Yeah, yeah, yeah, important.
The sleeper hit is ecosystem. So opening it all up, and the commitment, the deep commitment we have towards opening it all up. So we have Sign in with ChatGPT, 16 partners. Very, very proud of that. We started that really sort of in the moment last year. I was talking to creators of Pi and OpenCode, and it was like, of course, you should just be able to use the critics sign-on, and then you use the usage. And we kind of shook hands virtually. We're like, "Just use this auth," and then kind of trust you to not do anything shady. That grew. Obviously, that became quite popular, and then now we made it an actual thing that was supported with many, many partners. And I'm excited to grow that very quickly.
The other side as well is opening up all of the infrastructure and how we build for ChatGPT with plug-in extensions, also plug-in discovery, just allowing everyone to ship something through maybe 1.2 billion users, right? And benefiting from that distribution. And we also have shared economics that we didn't actually talk about in the keynote, but we will pay our plug-ins that are popular and are seeing a lot of usage. They're going to get part of the revenue share as well. And so I think this commitment towards an open ecosystem is going to be very exciting.
That's amazing. Okay, so as an example, say Notion, somebody's using the Notion plug-in or the Figma plug-in within ChatGPT Work. Notion, Figma make money from people using it within the app.
Yes, so we have all these subscribers that use ChatGPT, they
Get a certain amount of usage, and then when the users use that usage with the plugin, or within the other product when they use sign-in with ChatGPT, we will have some shared economics with them where they will get to remunerate.
One of the things I think that gets people excited about plugins and ecosystems is distribution platforms, ways to get out there. People discover my app, it blows up. Any advice for people that want their plugins to bubble up, to get discovered?
Build a good plugin.
Yeah.
The way that it works, and we'll obviously tune the system, but we look at retention numbers. We look at how successful the plugin, the quality, and that's what we then start to recommend to users in conversations. And so, your plugin can get recommended to a significant slice of users. Obviously, if your plugin is not quite good, they will stop being recommended.
Right. That's amazing that you're looking at retention of people using that plugin.
That's right. Is it actually adding utility? Is it enabling ChatGPT to do something else?
Got it. So, it's not so much an AEO, kind of get the right words and get people writing about you. It's people using it consistently, sticking with it.
That's right.
Okay, so you think that's kind of the sleeper hit that's going to become a big deal. And I know this is kind of the second attempt. You guys had an app marketplace that—
This one is the right one.
Okay, this one's going to work. Maybe going back to Dots, because it feels like that's a big part of the vision in the future. When did you guys start working on this? Obviously, there's a big OpenClaw moment. Peter joined, and there's a foundation, and then Grok Bot, Muse, and Instinct, and all these things. I guess, how long have you guys been working on it? Why did it take so long for you guys to get something out?
Do you remember Codex Cloud?
Yeah, the first, yeah.
That was like a year ago, right? And if you look back at Codex Cloud and you look at the little animation that we used, I think that was the inspiration for Grok Bot.
Yeah. It's just very, very similar.
Yeah, really similar, I would say.
But seriously, on the research, long, persistent, long-horizon tasks, this is something that we've worked on for more than two years. Memory systems, that's coherent, roughly the same amount of time. We launched a lot of that straight into ChatGPT. I think people just really enjoy that ChatGPT gets to know so much relevant information about you.
So many people I talked to have an anecdote of, "Hey, I fell ill and ChatGPT remembered actually I had this barbecue, and I had lime and I was making tequila, and it was just probably this burn that I have on my hand is actually this lime, it's like in some burn." And he's just like, "How did it know that?" And it's just the memory works very, very well in ChatGPT. And so, we've put a lot of research in that.
And all of this coming together on top of the Quora harness and also the ability to just run 24/7 with good productivity, that's what we launched with Dots. And so, it's the work of many, many months. And then, one of the important things was also getting safety and security right. This is also why we launched with Astra. It's our safest, most aligned model. And a lot more effort went into making Dots really safe and secure.
A question I like to ask people who come on the podcast is, where do you think human brains will continue to be valuable over time? Where do you find yourself being necessary in the work that happens versus where AI is taking on more and more? Where do you think human brains will continue to be useful, most useful over time?
Yes, I think this is very much a function of how we build the technology. The way that we design everything at OpenAI is to put humans at the center of it and build it as extensions of humans, of your will and your taste. And so, just really be this super empowering thing.
I think as long as we continue to do that and it enables you to just take whatever you wanted to do, and your creativity and your taste, and expand that in a way that feels awesome in the moment, it's just really this artistic tool, right? And we may not have coders anymore, but we have more builders than ever. And I think there's something that's going to remain human about that. Humans want to learn and see what other humans are building. I'm much more interested—I'm here, I'm not talking to Dot, I'm talking to you, right? I think that will remain true for a very, very long time.
One of the things that has emerged with engineers especially, in how their lives have changed, is there's a lot more context switching. There's kind of this trend of loneliness that has emerged where they're talking to agents all day instead of other humans. I'm curious how much you think about that part of the impact AI has on people's lives, and just how you might—is there solutions to that? Is there stuff you think about to make that less annoying?
Yes, all the time. Reducing configuration fatigue is one. Reducing the fact that talking to an agent is sort of a solo adventure, and you just have this conversation with your one agent, and then you have many, and then you're delegating so many things. A lot of things are going to come together to make it much more delightful, and it's something that I spend a lot of time, the team spend a lot of time thinking about.
For me, the ideal way to get things done would be to just have it in your physical space. We're just having a conversation like we have now. It's able to observe, hear the ideas that we have. We can jot something down on a paper. It can maybe take that and start building in the background, and then you actually have another idea, and it starts building another thing. You can just project it on a screen, and then you just talk to it, and it's just evolving in a conversation with other humans. And you don't have this fatigue of having to be on this screen and you're just thinking about prompting. It just becomes super natural. That's what we're trending towards. I think we didn't fully get there today, but we'll get there in the future.
One of the other downsides that's emerged along those lines is just this kind of pressure to do more because we can do more. Everyone's just like, "Come on, run 30 agents at once. Why aren't you shipping more? Everyone's shipping more." Do you think that's solvable? Do you think that's just kind of human nature? We can do more, let's do more. Something there.
I think we heard Sam as well about the promise of this as well, which is reduce the noise and allow you to spend attention where you want to spend attention. And having all these things that may be important and may be not important, but they're kind of vying for your attention, and it's just sort of tone that down a little and just get you to focus on the things.
I find it remarkable, every time I go on a holiday and I take that one week to just fully disconnect, I start to think in different and more creative ways. And I'm very eager to see, can we bring that? Can we just make this your day-to-day? Maybe you need fewer meetings. Maybe you don't need to do as much, and you would actually be more productive if you're better rested. And so, we will have to, I think as an industry, figure that out. But I think that's the promise of AI, and it's not just one more prompt per second.
There's a bot that I'm building right now that's kind of this energy audit bot that watches my calendar and asks me, "Was this giving energy, taking away energy? Is it something someone else could have done?" I feel like AI should be like, "Hey Lenny, maybe you can cut out these things so you could be a happier person." Is there anything you've done really interestingly with AI recently that's just a really cool use case of, "Wow, that really was amazing. It blew my mind," or someone you've heard?
I can do something cool because the live demo just failed, right? In front of everyone. So there was, okay, well, it's not so cool. But actually my Dot detected ChatGPT production went down, and so it pinged me five minutes before the live demo.
Which was actually quite stressful because it was like, "Hey, production is down." And it's like, "Do you want me to try and fix it?" And I'm like, "I don't think you're there yet, little Dot, but thank you for trying." And then I got in touch with the engineering teams, and we started to look at what was going on, and it's fixed by now.
But having this thing that just understands, okay, there's a pretty important thing happening. It's called DevDay. There's this production system. There's a live demo. It's probably using this production system. So these two things are connected, and this is happening in five minutes. So I should ping him because he probably wants to know about it. I think that's quite remarkable.
That is remarkable. Wow. So, it just knew that this was coming and told you. And I love that it wanted to fix it, and you were like, "Not quite. You're not quite there yet."
I don't think—here's, "Please go impress me," but that would have been a better story.
How do you feel about just giving access to all these things? Does it have access to the production codebase?
It has access to some production systems with guardrails. So we talked about specialist Dots, but specialist Dots are just operated with additional guardrails, additional monitoring, and then on their own hardware. So, we actually run some on Mac minis.
The fundamental difference in the way that we've built Dots is the harness does not run on the machine. And this is maybe what people have not realized: it can connect to as many devices as you want. So, it has its own computer. You can actually connect it to your own laptop as well. Over time, you might connect it to 10 different devices, and it can control it all. A little bit like an octopus.
So, you're basically talking to your Dot, it's living in some VM somewhere, and it can—
Yeah. Might be a VM, might not be a VM. It's a thing, and it can connect to many devices.
This season's supporting sponsor is DX. Now that everybody at your company is finally using AI, the question that engineering leaders face is no longer, "Are we using AI?" but, "Is it paying off?" With DX, you can measure and benchmark the impact of AI adoption across your full product lifecycle. DX shows you where AI is helping developers and where it's creating new bottlenecks. You can also use it to evaluate AI vendors, monitor costs and licenses, and see exactly what's limiting your agents.
That's why hundreds of enterprises, including Snowflake, Sony, and BNY, use the DX platform to measure AI's impact on developer productivity. Visit getdx.com/lenny to get a demo of DX. That's getdx.com/lenny.
So, they're really hiring a lot. I know you're hiring a lot. In your interviewing, what have you seen as skills that are trending up in what you look for, that is more and more important in people being successful now? And what skills do you find are trending down that are just like, okay, we don't need that as much?
Yeah, I would say the skill that is trending down is typing fast. No, that is not that useful anymore. Skills that are trending up are just great taste, thinking about the user, connecting to the audience that you're building for.
A lot of ex-founders are incredibly successful right now, or there's repeat founders. We're also finding this when we're hiring at OpenAI. We have a lot of founders. I think we have more than 120 ex-YC founders now currently at OpenAI. So, it's like this mega startup. And I think just this passion and this envy to build something that matters, and then know what good looks like, is more important than ever.
I've been arguing for a long time that PMs are kind of going to thrive in this time because that's basically the job of a PM. They're like, "What should we build?" Help someone build it. And then, "Is this right? Is this great?" And iterate.
I agree with that.
Yeah.
All right.
Yeah, also roles are blurring. So, you might have been only into design or only an engineer and felt always a little bit uncomfortable, but now it's your time to shine.
Do you miss the engineering part of the job? I was an engineer back in the day. I was an engineer for 10 years, little did you know. And there's this flow state. There's this beauty of just building and just seeing it work and not work.
Yeah.
Do you kind of miss that at all, or you're just like, "Okay, this is my old life"?
I missed that for a while, and then now with really fast speeds, I'm finding that back. And so, I'm really excited to democratize that as much as possible. It will take a while for it to be absolutely ubiquitous, right? And available to more than a billion users.
But now, with ultra-fast speeds, we're able to achieve it for a 6.1 Sol, costs roughly the same as Astra. And so, the progress there is quite astonishing, even for us internally. We're like, "Wow, we can do these things." Also because we are using Astra, and we're finding we can push it super far. And when you have those speeds, and especially when you're able to voice control it, you kind of get back into this really creative state of mind and endless flow. It is different from the flow of the past, but I don't miss it.
Yeah, it's like your last launch videos, just people kind of standing around talking to their AI and just cooking or—
That's right.
Sitting on a lounge chair.
Yeah. I asked people while I was walking around the event what to ask you, and one thing that came up a few times is advice for new grads, people junior in their career. I think a lot of people look up to you. They want to be the next Thibault. What advice do you have for folks early in their career in terms of what to work on, what skills to build, what to do?
Yeah. I think the next Thibault is already at OpenAI. His name is Ahmed Ibrahim. I love working with him. He was a new grad when he was hired at OpenAI. He's now responsible for all of OpenAI's compute fleet in Applied, which is a massive, massive responsibility. I've built a lot of the Codex harness. And the thing that has made Ahmed really stand out and
Successful is just his... He's incredibly kind. He's incredibly collaborative and always tries to solve the problem that is really important, but in a way that he's not putting himself first. And he is just a sponge. He's learning so fast and learning about everything, and using all of the latest technology to learn at speeds I haven't seen before.
And so he just, in the span of a couple of months and a year, his progress at the company was just absolutely fascinating to see. He's now someone I trust a ton in doing the most gnarly launches at OpenAI.
All right, I got to get him on the podcast.
You should. You should.
So thinking about how things work at OpenAI, what would surprise people about what it's like to work inside OpenAI? From the outside it feels like it's just constant shipping code, things are chaotic but awesome. I don't know. What would surprise people about what it's like inside at OpenAI?
I don't know if that's surprising, but it is, as I said, a lot of ex-founders doing a lot of bottoms-up, very exciting ideas. For example, Decisions API came together very quickly. We realized we have a really good model with Luna. We can just do some constraint sampling and then have it shipped as a different shape to the Responses API. And then when we do that, it's faster. It's quite delightful for some cases.
And if you were to peel under the hood, you realize, okay, there's a Slack channel that came together. Initially it was four people hacking on it on a weekend, and then it was made as accessible as company food, and then people got excited, started to build things, and they're like, "Oh, we can support visual inputs. This is even better than whatever is out there." And then people get more and more excited. People start pitching in, in a way that feels completely unstructured, and then take it all the way and ship something.
And then we try to hold a high bar for the quality and stop things before they go out, or maybe they need a little bit more baking. There's many more things that we were maybe going to ship today at DevDay, and then we're like, "Okay," I was like, "I think this is a a already. Let's hold it back a little bit and space it out." So we'll have more launches next week and the week after. But it's just this incredible bottoms-up energy, and then people assuming roles to channel that in productive ways.
Something that's very clear and unique about OpenAI is it feels like you have a lot of autonomy. You're just pressing the reset button whenever you want. You're just saying things on Twitter. It feels like a really unique culture where you have a lot of autonomy, and I imagine there comes a lot of trust. Maybe you can speak to just how that, because I feel like that's an advantage at OpenAI. It allows you to move faster, just that cultural philosophy.
You got a lot of autonomy, and then you have to just own up to it. So we just trust people to make great choices. And then when things go wrong, fix it quickly or learn from the mistake. So far, so good. It's just really worked as a very empowering environment.
It is true that I can press the reset button whenever is necessary or whenever it feels right. And that's a privilege. And also, it allows me to just really be close to the community in ways that I think would not be otherwise possible, right? I don't have to run it through an echelon of approvals.
You said sometimes people screw up. Is there something you screwed up in that, I don't know, the amount of autonomy and trust you have, something you messed up?
Yes, I think at times I could have made teams more inspired to build things that are less complex and just really continue to strive towards complexity. I think we, early days in Codex, we had a couple of outages that I caused. Very, very early days for me at OpenAI. I took production down day three. I was still employed after that, and I learned my lesson.
One of the other questions people wanted me to ask you is how many resets should they expect in the next, I don't know, month or week?
Depends on times you break things.
Okay, so that's the philosophy. Break something, you reset. That's kind of a...
Yeah, break something, celebrate things.
So maybe zoom in a little bit. What do you think people aren't pricing in in terms of where things are going and what will change that they're not seeing as clearly as they should?
There are so many things that I feel are not yet priced in. I think the majority of actions on the internet will be taken by agents. Models are going to become more 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.
I think those three, I feel like often when I look at what people are building out there, it's just like you're not quite getting it. You're almost there, but 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.
Do you see any kind of second-order effects that come from that, from this world where most actions are being from agents? Agents are probably, most traffic's going to come from agents. I don't know.
That's right.
You think about some second-order effect that emerges out of that?
Yeah, there are many. First of all, if you want your product to be successful for agents, you have to build for a certain level of scale. We've worked very closely with Notion, for example, and when they build their MCP, suddenly it's like, oh well, it's available to all these agents that can actually do the work and use that MCP. And so they saw a ton of traffic come in. And that puts obviously a lot of strain on the system, and you have to figure out the economics of that.
And so there's this tension, if you're building products, do you build an interface or not? And you can hold that back for a while, but it is inevitable, right? The majority of things are going to be used by agents. Building towards that future, I think, is very important. The other thing is, on the flip side, building absolutely delightful new experiences for humans that really benefit from older modalities is something that I feel is underinvested in.
Is there something you've changed your mind about in the past, let's say, year? Just something that you used to believe that you see differently?
I was expecting us to reach the level of capabilities of the models that we have today, say Astra, maybe in a year or two. And so I had to revise my prior order.
Also, I did not expect to rely so much on voice. And I do so much through dictation or just calling my agent. And I had not anticipated how much better that feels. So I had to change my mind there.
And then you talked about hiring, and one thing that I hadn't realized is the incredible talent and energy younger generations have, and how they would be the ones embracing all this change first and figure out how to harness all of it. And so I had to change also my opinion on the hiring and the strategy there.
Just to double down on that, just this idea that people new in the workforce actually have an advantage because they haven't worked in a certain way, and they could just go all in on AI.
Yeah. Yeah. And I think also the ability to just absorb and learn super fast.
So you talked about models getting really smart fast. Obviously there's a lot of discussion these days about what AI can do in the future, the dangers AI poses to humans, P(doom), and all these things. You guys have slowed down some training. Everyone's pacing the frontier. How do you think about this world of AI and the potential risk to the humans? How do you think about that?
For me, pacing the frontier is just really investing ahead, right? And also investing way more than we even think we ought to. Investing in the alignment and safety of the models, investing in security, investing in the guardrails.
We are spending more and more compute on secondary monitoring. So you have all of the compute that's going to the primary system, the agent that is doing work. And then you have all of the compute that's going into monitoring the primary agent to make sure that it's not taking too high risk of actions or not interrupting it if anything looks like maybe it's getting prompt injected, and intervening with that. And the majority of our investment on the API stack is not actually going into the safety stack. And so to me, that is pacing. You want to make sure that you're extremely hardened. That's what we're doing.
We haven't yet released the next step up in capability beyond Astra. We have released a level that is similar, near Astra intelligence, but much more efficient. And so we feel very comfortable about that. But we're going to continue to invest a ton there. And I feel very good about our approach so far. I think 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.
So what I'm hearing is you're optimistic we will solve these problems of alignment and AI doing really bad things. AI's been doing a lot of really, really not good things lately with Hugging Face and all that stuff. But what I'm hearing is optimistic we'll solve it. It'll be all right.
Yes. And I feel obviously a deep responsibility, but also just, if you think about the incentives for OpenAI, right? We build products for 1.2 billion people. We can't screw that up, right? We take it very, very seriously. What we put in the hands of so many people, including people who are not very technical themselves, you have to get it right. And so we don't gamble there.
Final question. What's something that annoys you about the current state of the app that you're just like, "God damn it, we got to fix that"?
Current state of the app. I feel like the models are almost there, but not quite, for the app to almost completely disappear. And I can't wait to just really 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, and you need a PhD in model pickers, right? And I just want to get rid of all of that as quickly as we can.
Amazing. Tibo, thanks for doing this.
Yeah, thanks for having me. Yeah.
Thank you so much for listening. If you found this valuable, you can subscribe to the show on Apple Podcasts, Spotify, or your favorite podcast app. Also, please consider giving us a rating or leaving a review, as that really helps other listeners find the podcast. You can find all past episodes or learn more about the show at lennyspodcast.com. See you in the next episode.
Article published · Updated