Tibo Sottiaux

Cool. No, we're working things through. We have some more creative solutions than others. But this is the shape that we think we'll go out with roughly in two weeks. But it will feel significantly better. Two weeks is eternity in software development these days. So yeah.

Well, tell me about why you think that's the right move.

So it's really about making it work across all devices. And for me, just really what you want is, you want a highly, highly personal entity that you can reach from anywhere at any time. And that's not yet— what you've seen is not yet that thing. But you want it to work on your phone. You want it to work on web. You want it to work on desktop. You want everything to be in sync. You want your agent to have— It should feel like this one entity. Most of our users are on web, mobile. We've got two million businesses on ChatGPT. So we're kind of very eager to bring everything that has made Codex successful to them in a way that's also lightweight and appropriate. And we did an analysis as well of how many actually of the hard tasks that people give Codex around general work require local file system and computer access — [the answer] is actually surprisingly little.

Yeah. I bet.

So what we've done is, we're now capable of running all of that, but outside of your computer environment.

You did the hoist. Will it be able to call down to my computer?

Eventually. Yeah.

And then you start to really see, you have the agent, it's able to act on any device, depending on whether it's connected to it or not, with a variety of permissions. Maybe it can not just access your computer, it can also access a Dev Box or multiple Dev Boxes.

Yeah. Well, I guess I'm curious about— I wanna back up, actually. Tell me about when did you start working? Were you working on Codex from the very beginning? 'Cause when Codex Cloud launched and in the GPT-5 days, I was mostly talking to Alexander. So I'm kind of curious about your story with Codex and how you came to it.

There's many beginnings to Codex.

Yeah. And 'cause it's the same name for a lot of different products too.

Yeah. I got involved with coding agents around December 2024. And this was in the pursuit of writing better infrastructure and accelerating the research at OpenAI. And so what we started to work on was a series of models and way to interact with them to write code really fast. And we trained agents specifically on our monorepo, Python, to write good infrastructure.

Oh, you're training it on your monorepo?

Yeah. That was in December of 2024. Python only really, specializing on OpenAI infrastructure libraries.

Oh, for internal use?

Yeah, for internal use only.

'Cause I thought that your monorepo was your benchmark for your new coding agent. Okay. Yeah.

It is as well, but the origin of how I got involved in Codex was just really accelerating research within OpenAI. Accelerating specifically writing infrastructure for the system.

Mm-hmm.

That was my background. And so did this for a bit and then as things started to work, we got increasing adoption, increased our ambitions step by step. Eventually to the point where we thought we should combine forces with another effort, which was called Scientist Assistant, and there was a merger of teams, and that's when I got to know Alexander. So as I'm working with Alexander, we were working on what's referred as the cloud sprint, which was roughly a seven-week sprint. And then we came out with Codex, the cloud version, which was really exciting. It was a little bit too early. Eventually, it's what we're going back to. What everyone will go back to. But it did not have PMF at the time, and it was too frictionful.

Tell me about that 'cause I talked to Alexander about it, I talked to Greg about it. I think Greg's framing was interesting, which is something like we went from Copilot to you have a chat inside of your IDE to the next step was, in my mind, Claude Code being like, "We don't need the IDE at all. We just have an agent on your computer." You guys skipped that step initially.

We didn't skip it. We had it internally. We did not release it. We went through that phase very quickly. We built a version of that and then discarded it after a month to go for the full cloud version.

Why?

Because we were optimizing also for what we thought we would benefit from the most if we were willing to invest a lot in the setup. And so it was very successful internally. 'Cause everyone having perfectly reproducible environments, provisioned dev boxes at scale, seemed like the way to go. And it was making people a lot more productive. But then externally that was way too high friction. And then with models getting better and better, also realized you wanted to pull in a lot more context from the system and you kind of felt trapped inside of that little VM. So we went back to local. Then again, it was too early.

Yeah, I feel like something really shifted. Oh, by the way, I should tell you the framing of the article.

I'm happy to talk about it and check—

Yeah, totally. The framing of the article is how Codex became king. I feel like there was this... Well, the last year or so has been very Claude Code Anthropic focused in terms of the public narrative. And we were very early to Claude Code, and the team now is almost everyone is using Codex as their daily driver. And so I think there's going to be a shift in the meta over the next month or two. When that happens, wanna have an article that's like, “Here's how that happened from the inside,” an oral history of that. So that's kind of what I'm putting together, and the sort of framing all the interviews. I do feel like I was very around Codex... Well, actually, before we get into that, tell me about the decision to split off the model line. 'Cause that feels like one of the things that it was one of the ingredients to success here. Tell me about that.

So there was no decision to split off the model line.

Mm-hmm.

It was rooted in the very beginning of this research acceleration effort, which I was part of, where we were training agents specifically on coding, to help us write infrastructure and do research faster. And we were doing so off of things that were so far from what we were doing for external models. This is just really a research exploration and an optimization of our velocity. We were seeing rapid gains on coding. And so we continued doing that, and what we developed is a series of research methodologies that were very different and were hard to reconcile for a while. One of the relevant things here is, after the Cloud sprint, we did not immediately find PMF. I took over the team, and then I changed the direction quite a bit. We pivoted to local. We had a bit of a lull with the Rust rewrite of the CLI,

Oh, interesting. I didn't realize you rewrote it.

Yeah. So the Codex CLI initially was written in TypeScript. Then there was a Rust rewrite, where we decided... that it wasn't the CLI was the way to interact with the agent, but just really the underlying infrastructure was our agent, which was our most precious infrastructure. And we needed to take great care in it and think about the correctness, the security, sandboxing, and all of that was low-level system design. And we wanted to invest in it. And for the long term, there was absolutely no lack of conviction. It's like 100% conviction that this is very, very core to our success. And so we decoupled the textual — the TUI implementation — from the agent implementation very early on. But that also led to, we were kind of slow to have really good CLI features, because we were so busy developing the agent.

That's interesting.

And then we kind of went back to it.

Tell me about the open source decision. Does that connect to the open sourcing of the TUI decision?

Yeah. So the open source decision was a decision I made with the team. One is, when I grew up, I loved open source. It's how I learned everything. Two, it felt that if we were right and this was going to change how code is written, and very quickly so, open source would struggle, and we kind of also wanted to be part of the ecosystem instead of just causing trouble and not having solutions. Just being part of it. And then the third thing was, it feels like it was a very — still very early days — and there was a whole bunch of explorations necessary, and nobody really knew exactly what a good agent would look like, what a good implementation would look like. And experimenting with that in the open, and showing that there's not really magic to it, and everyone can fork it and contribute back, and then we can benefit from the ideas — seemed very appealing to me. So that's why we decided to do it open source.

Got it. Okay. So it sounds like... 'Cause my memory of some of this was, there was a point before 5.2. There was maybe there was like five Codex, 5.1 Codex, where it was very slow, powerful, hard to use, but often right. And it sounds like maybe during that time — tell me if the timeline lines up — during that time, part of what you were doing was, yes, making better coding, but also building the underlying fundamentals of the agent itself, like the agent harness, that made it so that the visible progress to users was a bit slower than you would've liked. Is that a good summary?

Yeah. We took the general cultures, like, move fast on things where you can move fast. And so you want to experiment and learn, and then for some of the fundamental things, take your time and invest in the long term success. And so for sure, we were really doing a lot of underlying investments during that time. And then also this was our specific strategy on the model as well, which was to focus on improvements that would benefit — helping solve hard engineering problems. Not necessarily being a model that would do things that are cool, or kind of perceived as cool. I think around the same time, Gemini was just investing a lot in front end. I was discussing with people over there, like, “Oh, what are you focusing on?” And these kind of things. And it was really like optimizing the model for this wow factor. We made the opposite decision, and we were like, “Okay, how can we develop an accelerating flywheel as quickly as possible? What is the best model as a tool look like, the best agent as a tool look like?” And then build a lot of that underlying infrastructure. And at some point, we had this sort of realization that we should continue maintaining the CLI, but the biggest next investment should be a native desktop application. And just really optimizing for what the future of agents were gonna be. Which is a lot of multitasking, probably also you want visual output, you want to interact through voice. It has to be this very rich interface if you're gonna spend all your time in it. And just doing all that in your terminal felt very silly when you played it all out. And so decided to invest a lot in that.

Tell me about that.

That's when Andrew got involved. He was like, I found him. He was just the best person for this.

Yeah. He seems amazing at it. Well, tell me about that because there was that time where TUIs were so hot. And even people internally who are non-technical were starting to use TUIs for everything. So it was a pretty, and still is to some extent, a pretty counterintuitive bet to be like, no, no, actually GUIs are still better strictly. Tell me about that decision. 'Cause you can imagine being like, “Holy shit, everyone's talking about Claude Code, and it's all about the terminal and blah, blah, blah. We have to catch up to that or whatever.”

Mm. I don't believe in looking at the competition to make good decisions. I think it's important to reach conclusions from first principles. And based on the information that you have. And the information that I had at the time was that our models were rapidly getting better. And we were seeing the models do longer and longer horizon tasks successfully. We're seeing models capable of ingesting a lot of data outside of code bases as well in order to inform their work. And then we were early at OpenAI as well on voice and multimodal. We had state-of-the-art image generation, state-of-the-art advanced voice mode that shipped two years ago. And so this idea of playing it out where is this very rich interaction, it just felt completely silly to expect that that would happen in a terminal. But obviously requires a lot of conviction to then go and do it.

Yeah. And tell me about that because you have to convince other people that you're right. Maybe if you're not looking at the competition, someone is. I guess I'm curious about the culture. It feels like there's a culture of a lot of parallel bets here, and seems like you're able to take some risks, but I'm curious how that works.

Mm. It was clear and controversial to them within the team that desktop was our main bet. The desktop application native application was our main bet. And so it was not a small parallel side thing. Obviously with anything controversial, people don't necessarily see it until it happens. But I would say the conviction of the team after talking it through and explaining why we were making that bet, was people were excited about it.

Yeah, yeah.

Especially 'cause it took us less than a month for a lot of the team to have switched it and feel that this was actually going to work. I think pre agents, maybe this would have taken us four to five months, and it would've been very hard to convince people because we were kind of able to make rapid progress.

That does make a big difference. Tell me about that.

Mm. Yeah, I think it takes way less talking now to convince people. And you can just go and do it in like two days. And then show the thing. And that's a lot more convincing than any document you could write before. And so that was kind of also playing out in real time, which was very cool.

And at what point did you decide to do the desktop app as the main bet?

This was in December of last year.

Okay. And that was around when 5.2 came out? Okay. What about, if you were ever gonna split out a model line again, and maybe a product too, in the same way where it's like you had a ChatGPT desktop app, you didn't necessarily have to do a separate Codex desktop app. When is it the right move to do that kind of split?

Right. So we were also optimizing for velocity first and foremost. Just deep sense that all of this was going to change again and again in the course of a couple of months, and you need to optimize for how quickly you can ship and move. And that we needed to start with a very forgiving audience as well.

That's interesting.

And doing that with the ChatGPT desktop app felt that was not fulfilling either of those.

And by forgiving audience, usually people don't think of developers as being particularly forgiving.

I think they're amazingly forgiving. And they are very good with feedback as well, which makes it delightful to work with. If you can sort of navigate the fact that you get like a thousand different feelings thrown at you. Shipper: [laughs]

Yeah.

Well, they understand what it means to build a product, whereas Joe in Montana who uses ChatGPT to get answers doesn't understand what's going on under the hood.

Yeah. And if it doesn't work, that person will probably stop using your product, whereas someone who's an early adopter and a developer will come and bother you and be like, “Hey, you broke me. And I wanna know about that, and I wanna fix it. And I wanna learn from you.”

That feels like another good generalizable strategy is serving builder type people with new AI stuff 'cause they're gonna figure out all the different corners and all the ways that you might not even realize, and then you can translate that to everybody else. Is that consciously something that you guys are doing?

Yeah, I think there's a lot that we learned from that strategy. And now we're reproducing in other corners.

What are the subtleties of—

On different models, we will train other specialized models when we feel like it allows us to learn something more quickly. Or when there are research reasons to do it.

What are the—

We did one for life sciences. Rosalind. Yeah.

Interesting. And what are the subtleties of doing that sort of strategy well, the serving builder type strategy?

The subtlety is, and this is something that we are still figuring out for Codex, is that you can build amazing technology that is then overfit, right, to this type. And so it requires then quickly understanding how you can pivot it and position it to be broadly applicable. Or to generalize it, and this is exactly what we're doing with Codex, right? So we're... Okay, the technology is amazing. The agent can do all these things. It's clearly much more useful than just for coders and builders. We're seeing very quick adoption from people who are doing all sorts of general, day-to-day work. But it has been quite optimized for technical people. And so that's the trade-off.

Yeah. How are you thinking about how to push it back well?

So a lot of experimentation. We have a broad network with whom we push early ideas, you included.

Yeah.

And then also we have a lot of diversity at OpenAI itself in ways of working. And so we're able to ship a lot and iterate quickly within OpenAI internally.

When did Codex sort of escape containment in the technical builder side into some of those other functions of OpenAI, internally?

I'd say it really started to pick up, what was it? Six weeks ago roughly. Something like this.

That was when we hit 85%. Yeah.

Yeah. That's when it was clear, when Sarah Friar did an all-hands and explained how the last fundraise was in great part orchestrated by Codex. It's like, okay. Seems like we've got something quite general here. And then we were seeing a lot of stories across comms, marketing, and sales. So we had this desire of just really generalizing it and solving some of the hard problems around security, safety, when you don't... And having the power of an agent that uses code under the hood, but where code is the tool and not the main output. Shipper: [laughs]

I feel like auto review was a big moment where people could really just start feeling confident.

Yes. So we developed internally. Externally we shipped it as a review for me or auto review. And you know what it does, right? It's this second agent that verifies all the actions of the primary agent, but against the original intent. And it reduces approvals by like an order or two of magnitude.

I love it.

That's a reason as well, why I think the team cares deeply about the experience, but also uses it every day. And so we're all using the product that we ship. There's not a different version of Codex internally than there is externally. That was like an important decision. We could have optimized it all kinds of different ways internally than the external version, but the majority of OpenAI is on the Codex desktop app all day, every day. We have a very, very active feedback channel with over 250 internal messages of feedback every day, which is now in great part triaged and actioned on by Codex itself. And so it's kind of like this, yeah, building itself.

That triage is so crazy 'cause you'll share something and it'll be like, “Oh, someone's already shared this. Here's where it's being worked on.” Or it will just correct you and be like, “That's not right.” Or like, “Yes, this is the right use case.” It will give you education, which is kind of crazy.

That's so cool.

Yeah. The Codex we spin up on feedback has access to the roadmap, has access to all of the documentation that we wrote and has access to the code base, and is able to sort of tell whether something is a regression or not. And then if it's a regression, writes a PR, assigns it to the person, based on the blame layer. And so it's just a highly efficient sort of magical experience. We wanna bring this to external too.

I guess what's it been like for you going from... Codex for a while was mostly under the radar unless you're a technical person. You ran a Super Bowl ad. It's become a thing both internally and externally. What's that been like for you, that ride?

It was awesome to see. For almost 15 years, I've always built tools and products for people. I love to see the broad use and utility that people get out of it. I always want to put something out there that is the best version of itself, and seeing the wider and wider adoption in something that we initially built with a high degree of ambition, but was just internal only, going to then external and then continuing to build it with the community, then going from there. And it's the very beginning still.

Yeah. Do you have any specific moments where you're like, "Oh shit, this is a real thing"?

I think there were a lot of small moments. Like the Super Bowl ad was definitely like, oh wow, we're actually doing this.

Yeah. Hard to put it back in the box when it's in front of that many million people.

Yeah, yeah, yeah.

Were you watching—I assume you were watching the download numbers when it went live?

Well, we had a Codex party where we were watching it live—

Yeah. Okay.

and then we were sharing the numbers.

What was that like?

It was exciting. It was exhilarating. The vibes have been really good. I think once we got past the initial period, which is like pre-GPT-5, I would say—GPT-5 made a dent and then 5.2 just really started to accelerate. I feel like the vibes online, the community, it's been really good since the start. And I've always wanted to sort of like be as nurturing of that as I possibly can. And sometimes it's challenging, with scale and capacity and all these things. But always try to be very transparent with what we're dealing with.

Yeah. Mm-hmm.

So no, I mean, it's been awesome. It's been magical.

Do you actually have a big red button that resets rate limits?

I have a big red virtual button. There's a better version coming out today, actually. So up until now, when I press it it actually resets everyone's rate limit. People have been complaining about that, so we changed it where we now give everyone their own button.

Hmm. Oh. Cool.

So you can press it yourself—

Whenever you want. Yeah. That's great. Very cool. Sottiaux: [laughs]

I think people will prefer that.

Yeah, I love that. Katie, any questions on your end?

I just wanna know, this is kind of my guiding question as a non-technical person who didn't touch Codex because our Naveen used Codex, and so I was like, "Oh, it can't be for me too." Why am I, as a writer and non-technical person, why is Codex so good for me?

Yeah, she uses it every day.

I live in Codex.

[laughs]

Thank you.

And then—

So why?

Yeah, just what is it, what changed or what has always been true about it—I don't know which it is—that now makes it so generalizable?

Yes. So this is just, I think, the natural conclusion of doing a very powerful coding agent, is that over time, the agent needs access to broader and broader context. Ultimately, as much context from the entire world as we can give it. And it needs to be able to perform a broad set of actions to be useful for the builder, right? And so it tends towards not a coding agent, it tends towards a very general, useful assistant that can help you with everything. And when you realize that, which we did, it's like Codex is not for technical people, it's not for builders. Codex is for everyone. We're building a very, very general kind of assistant. And so we started then to take steps to make the Codex app more friendly. And if you're not technical, we don't show you code, we don't show you bash commands to approve that you wouldn't be able to possibly understand. We added extra security controls. We added the auto review. And so like evolved to a place so that it could actually provide that utility for everyone. You still have the name Codex— Parrott: [laughs]

which a lot of people like. Some people are a little bit intimidated by it. Yet to find out if it will be sort of accepted by society. I hope that it will take on a life beyond coding. And Codex will become synonym with a very powerful kind of agent or assistant that can help you with pretty much anything you throw at it.

Well, 'cause it's interesting 'cause when I see the word Codex, I don't think of code. I think of a codex, like a book, like the Bible.

The scroll.

Again. Because—

Its definition.

Yeah. So brand wise, there's this merge happening, and what will be the name?

Mm-hmm. So Codex will move to ChatGPT, and so the overall experience will be ChatGPT. But you will be able to use Codex within ChatGPT.

Okay. Interesting.

But the version of ChatGPT that I would be using would be basically the Codex desktop app with ChatGPT added, but it's sort of positioned as ChatGPT with Codex inside it.

Yes. So we're not taking any of the functionality away. It is exactly the same app, right? But we're actually replacing the ChatGPT desktop app with it. So the Codex app becomes the ChatGPT desktop app, and through that becomes ChatGPT on desktop, right? But it's all the same functionality that you use.

Why not just keep it separate?

We're considering everything. We're learning fast and adapting fast. So one of the things that's for sure is people who love to just use the Codex app as a separate thing will be able to continue to do that. You'll not be forced, not for a while, to move. And then we'll continue to adapt and learn.

Does it go to being like, "I have to use the ChatGPT desktop app."

Would you rather have a toggle or a separate app?

I don't want any toggles. No toggle.

Yeah, just a very pure a pure experience. And yes.

Yeah.

Behavior change is just so hard. That's almost harder than the capability of the technology it seems—

Mm-hmm, mm-hmm.

[Behavior change] is getting however many millions or billions of people you have to accept this new thing. I imagine that has to be a very complicated process to help people understand this new thing that they're gonna be seeing that is different, so much more powerful, but helping them see how it can help them. I was really impressed with the rollout of the communications around the plugins, for example. How do you think about handling human behavior change and educating people to use the app?

So there's two populations we have to bring along, right? So there's the Codex set of users today who are like—

Us.

Yeah. Used to everything that it does. And I think that one we'll manage fairly easily, figure out some details to not upset the, the—

The fans.

The fans of this world. The other thing which I'm spending more of my time on is how do we bring the one billion, almost one billion active ChatGPT users along and have them discover that now our models can do so much more. And it's a really exciting opportunity. And so we're working on how do we introduce it to them. And this is also why we're doing the move of bringing Codex to ChatGPT, to just really decrease the friction, and make it super easy. Again, we have over two million businesses using ChatGPT. To date are not benefiting from Codex.

Mm. Mm-hmm.

Why not just... And I'm honestly asking this question. I could imagine, for example, it just still for them looks like ChatGPT, but when you give it a query that verges on something that Codex can do, it's like, "Hey, I have all these new powers. Do you wanna allow me to access your computer," and whatever and there's no tabs. It's just like... Yeah.

Yeah. We've played with routers in the past—

Yeah. Yeah.

or using intent detection and it's very hard to do well.

You can't really read the user's mind, right? Until we have chips in people's brains, you won't really know what they want.

Even then.

Yeah. They're like, "Sign me up." And so the future is one unified mode that does everything, where ChatGPT can just do things for you. We're not there yet. And— Shipper: [laughs]

It's like a change management.

a lot of people just rely on how fast and delightful ChatGPT is, for all sorts of things in their lives. And we don't wanna disrupt that. We wanna respect that. At the same time, we want to show you what's now capable, what you now can do with a Codex-like agent. Especially, I think, I use it on my phone all the time now. Once it's synced across device, it's really good. So we're trying to do it thoughtfully and gradually. And then eventually enough people will kind of be used to these kinds of capabilities and then we'll be able to kinda unify everything.

You feel like you need another playground with less pressure to do more different things once you do this?

I'm sure that will emerge. Yeah.

Do you have a sense for along what dimension or what you need, why you need a playground?

Yeah. We have a couple of things that are brewing right now, which I'm really, really excited about, which feel qualitatively different from anything that we've put out. And so we're exploring a little bit about how we're gonna bring that to the world. And it might be separate. Yeah.

Awesome.

I'm excited for it.