Host: An interesting story, which you recently shared on X, concerns how you were part of the team that created this internal Google bot that was ChatGPT, but a year before ChatGPT. It caught on like wildfire. It was more than just a research project for Codex. You built it in Rust, and at that time the model was still not in distribution for Rust. It turned out that it is quite clear that Rust as a language in reality will become quite quickly useful for agents if we decide to attach certain efforts before that.

Host: When someone joins Codex, what do you tell them? How is it here? Is everything done? What you hear the most when you have questions is, for example, "Did you ask Codex?" Newcomers are still surprised—what can you ask it? Practically whatever. Creation was the interesting part, but maintenance was painful. Maintenance is something like the tax you pay over time to make everything work, but I think this changes so that there will be a lot of things simply automated.

Host: Do you still have the concept of code checks? The role of code review is changing, and now it is like this: "I think that Codex is one of the most popular tools for coding based on artificial intelligence today." But how did everything get started?

Host: Many of you know today's guest Thibaut thanks to his generous and frequent engagement with Codex users. He was also there when Codex as a product was launched, and from then on has led the wider Codex team. Today we will consider how Codex started and why it was built in Rust and became open source; how code checks are changing inside Codex and OpenAI teams; what it means when maintenance and re-architecture become incredibly cheap; what the merger looked like between Codex and ChatGPT; and many more underrated engineering challenges of this project. If you want to understand how the teams inside OpenAI plan to check and supply software, this episode is for you.

Host: This episode is presented by TurboBuffer, an incredibly scalable, fast, and cheap hybrid search engine built on top of object storage by a team of engineers I really liked after spending time with them. TurboBuffer is a tool that companies like Anthropic, Notion, Cognition, and Harvey use for connecting their AI products to huge volumes of unstructured data. When I talked to engineers who use TurboBuffer, the theme that always arises is reliability and performance at scale.

Host: The reason for this relates to TurboBuffer's architecture. TurboBuffer uses object storage only for state, and NVMe SSDs with cache memory for compute. Data in TurboBuffer is organized into namespaces. You can imagine a namespace as a database table, a search index, or an S3 prefix, depending on the background you come from. When a namespace is not queried, it remains in cheap object storage without associated computational expenses. When a namespace is active, TurboBuffer transfers it to hot caching tiers, so requests are executed very quickly. This fundamental design simplifies scaling to hundreds of millions of namespaces. If you create a multiplayer AI product, each user and their agent can have their own dedicated search index without overhead expenses. And every namespace can contain hundreds of millions of documents without any special configuration. You can scale TurboBuffer with practically no restrictions, while performance, reliability, and operating model remain consistent. If you need to connect AI to large amounts of data, TurboBuffer should be your first choice. Check it out at turbobuffer.com/pragmatic.

Host: Thibaut, welcome to the podcast. So good to have you here.

Thibaut: Thank you for inviting me. So good to see you again.

Host: It's nice to do this. Last time we did it in person, now we do it through video. First, I would like to ask you, how did you get into the field of technology? When was your first time realizing you wanted to work with computers?

Thibaut: That's a good question. It was a long time ago, very long ago. My parents actually decided to move from Brussels, where I was born. They just thought it was great to buy a small house and repair it, but it was located in the middle of a village where almost nothing was happening. I think there were about 200 people living there. Not many people with whom I could talk or make friends. And so I was stuck. It was around when I was eight years old. I was stuck with computers, and those were early days for me and the internet; this was my way to learn. And everything else comes from this. So, I am grateful to my parents because they moved to a remote place, and I didn't have any other choice but to get interested in computers.

Host: As soon as you finished high school, you continued your studies and joined university, right? So you actually studied it properly.

Thibaut: Yes. I studied mathematics, applied mathematics at the university. I entered quite early, and that's why I also finished earlier. I thought for a long time that I really wouldn't succeed in academia and quit teaching. I had small companies and a little consulting business while I studied. I worked in banks. I was very interested in supply chain problems and applied mathematics, and I liked to sell solutions and learn many things through this. In the end, I found myself in the startup world in Belgium. I did that for a while, and then moved to London to work first at Google, then DeepMind, and then eventually moved here to OpenAI in California. I love the Californian weather. We can talk about it; it was very good.

Host: Immediately after university you founded a startup, right? Did you have the startup bug or entrepreneurial bug?

Thibaut: Yeah, so this startup was completely dedicated to pharmaceutical supply chains. We investigated the sub-supply chain for clinical tests and tried to optimize and solve questions like, "Should we produce more medicine? Where should it be sent? How do we avoid waste?" And thanks to this, clinical tests became more effective. This used traditional methods not related to machine learning—traditional optimization methods, solving Monte Carlo simulations, and stochastic multi-step optimization problems, actually. We also applied this in the steel industry and to electrical power grids in Europe. It was focused on small forms of optimization tasks that were very interesting. To this day, that company still exists, and I think they perform some of the most interesting work, though it changes a lot with modern artificial intelligence, of course.

Host: But it's interesting because you say, "Oh yes, it was not machine learning, these were simply traditional things." And then you dive into modeling, Monte Carlo optimization methods, and algorithms. It feels like you just got completely absorbed, isn't that right? Was it something like: "Okay, here is the problem space. How can I use the mathematics I've learned—or haven't learned yet—to simply go deeper and deeper?" Do I understand that correctly?

Thibaut: Yes, that's why I was obsessed with applied mathematics. The idea with theoretical mathematics, science, or physics is that you do it because there is something waiting to be discovered, and there is something beautiful about that. It all reduces to patterns and expanding frontiers, but you don't always know how you will apply it. And then there is the real world—there are all these cool problems that are simply lying around, and I was very interested to see how I could make the world a better place, and how I could apply complex mathematics just to optimize real-world systems.

Guest: ...the world around me? And there was a lot of dissertations that stood behind this startup.

Interviewer: Yes. And then, after the startup, you ended up at Google, first at Google London. It was in 2015, and I remember that in 2015 Google was really, truly a competitive place to work, perhaps the same competitive as OpenAI today from the point of view of industry or prestige. First you worked on Maps, and then moved on to DeepMind. Can you tell a little about what you were working on, and why did you have already moved on from this very interesting field, which is clear you liked it, for example, optimization, logistics, and all that?

Guest: Yes, I didn't start with Google Maps. I started with a project which was intended and aimed to make websites faster, and make websites faster especially on mobile devices. At that time, you know, Google tried to see the transition from desktop computers to mobile devices and saw how more and more traffic comes to mobile phones, so they wanted to get ahead of this, so financed a number of initiatives and projects. I worked on one of them. It was done because it was a small group inside the advertising organization, in order to compensate for the loss of income from advertising through traffic switching to mobile devices.

Guest: And I worked on this for approximately two years. And then it was canceled. And although it was technically challenging, the most interesting thing is that through deciding complex technical tasks, I learned a lot. What did I learn? About having product-market fit, not having the necessary users, not having the right feedback communication loop, and not blindly trusting your product manager when they say the project is going well, when in fact things are not going well.

Guest: And then, you know, one day, a vice president arrived from California. And he just said, "Oh, yeah. We are canceling this project. Too bad you only have hundreds of users. It's clearly not the scale of Google." And then it was incredible, people were surprised. And the lesson I carry with me is to always doubt, always think deeply about the actual influence you have, but also about the importance of the overall project to which you are making your contribution.

Guest: And then I moved to Google Maps. Google Maps was very interesting; I worked on the reviews, and then, in about a year, I couldn't ignore DeepMind. It was a special place headquartered in London. So many wonderful things were happening. These were the early days, with rumors about things like AlphaGo. And they seemed to do extraordinary things, and really tackle the most difficult problems that can be solved. And with my background, that was obviously very attractive.

Guest: Started there, I worked a lot on research infrastructure and research tools. It became the topic I was working on for almost ten years, and this is also very similar to how I approach things: creating tools and products that help make others more effective and bring them a lot of utility. At first I did this for research, and then over time began to think about things much more generally and broadly. Eventually, that led to where I am now.

Interviewer: Yes, and recently you shared an interesting story on X about how you were part of the team that created an internal Google bot that was similar to ChatGPT, but a year earlier. Can you tell about it? This is a new story; I haven't heard of it before.

Guest: This was part of DeepMind. There were also several attempts. There was, for example, Brain, which at that time was separate. They had their own attempts to develop large language models, but this was definitely something that was being investigated. This was not the main direction of DeepMind. DeepMind was very focused and busy, for example, thinking about grand challenges and games, and thinking about RL not in the linguistic sense.

Guest: So there was a group that promoted large language models, thinking about: what if large text corpora are everything? What if we push language to the maximum and simply scale language models—will this be enough to get general intelligence? These were heated debates at that time.

Guest: And then our group decided to persistently work on this. And it seemed very natural: when I was creating tools together with others for research, we naturally thought: "What can we do with this model? How should we present it to researchers? How can they inspect incoming and outgoing data?" And finally it took the form of a chat system.

Guest: So we created it independently. It was a lot of fun. At first the models were a bit absurd, not very consistent or useful, but it was very fun to play with them. It caught on like a wildfire; everyone across DeepMind was using it for small talk. It was primarily a research project built by researchers for researchers.

Guest: And over time, there arose a desire to launch it as an external product, but DeepMind was just not configured for that. Google had a rigid, established way to launch products, with a whole mechanism and production stack. That stack was heavily optimized over many years to do things reliably well, but it also made it very, very difficult to create an environment for genuine product innovation.

Interviewer: And then I wanted to ask what made you look around or maybe even consider OpenAI, but I believe that you have partially answered this question. I'm just putting myself in your place: in 2023 or 2024, you are inside Google, you publish amazing papers, conduct really good research, and make super interesting things, right? Expanding the boundaries of what was done earlier, and within the company where you have already progressed so far. For people who feel comfortable or good where they currently are, what motivated you to continue to investigate what else could be out there?

Guest: Yes, I was very comfortable. I was doing my thing... it was a great place. But in fact, I had a strong desire to meet wonderful people and join a mission that I truly believe in, where I felt that people deeply understand this mission and care about directly and positively impacting the world in a massive way. Not in a detached way where people say: "Oh, we just do our research work here, and finding how to make it useful is someone else's job." I wanted to join a group where all parameters would be considered together, where research and product would be co-designed. OpenAI was that disruption.

Guest: ChatGPT was taking off. I met several people from OpenAI and thought: wait, what? You only have about 20 people working on ChatGPT? That is a very small number, and it was extremely inspiring. How does it work? How do you succeed in supporting a product of this scale and level of autonomy with only 20 engineers?

Guest: And as I dug deeper, it was simply an amazing group of people with an amazing mission—super talented and super focused. That just strongly attracted me. And then I immediately joined the preliminary reasoning efforts, as is typical for OpenAI. I joined and thought: "Yes, something big is happening. We are running reasoning models. This is a new paradigm. Let's move quickly forward."

Guest: And approximately a month later, the company launched the o1-preview model. It was very nice to be a part of this. I wanted to be a part of a place that quickly develops, cares deeply about real-world impact, stays in tune with the world, and actively listens to what is happening.

Host: ...with me when creating the Codex, when created products, it was looks like a community, listened to the community, simply focused on very intense reverse cycle communication. And then created something, about what do you just want to take care and, you know, take care of the usefulness that it gives to the world. And then, of course, you started quite quickly work on the Codex. So you joined in 2024 year. Can you take us back to past, as you thought about artificial intelligence, degree Master of Laws and Code? I know that there were then ASWE efforts. We also talked about it under time of deep immersion that we did in pragmatic engineering, autonomous engineering software software. ASE, yes, that's right pronounced within the company.

Guest: Hmm, we don't have any more. ASE efforts. Well, you know, This is Codex. Hmm, but actually for me it is that I joined, I started to create infrastructure for research. Many of of what I was doing previously, it was large-scale, hmm, data storage, analysis, and then tools for understanding of educational runs. I...I did many different things throughout his years. Hmm, but it was always about about that create for others and make them faster, and just take great care about, you know, do it well. AND then with the help of tools and infrastructure new things are possible. AND that's when I joined OpenAI, in I was the same. idea. And then, with previous by reviewing what I I want, and some of later models, It became very clear, what do we need use yourself models to help we have to move faster. So I just got carried away by this idea about restrictions, how do we we will use these models for themselves research. So I gathered with others by research people. We started teaching models. We started to create small agents. They were real predecessors of the Codex. AND it was like, what we taught internal models to they were very skilled at the base OpenAI's Python code. And then, hmm, very skilled, you know, having a good taste in architecture, good taste, you know, in code style. It was only in Python. And then the idea was that what could we do use this for very fast construction infrastructure and, you know, help researchers to code rather. And then, you know, and then we would moved faster. AND then over time, when you just promoting it and simplify to the point, you do a lot, you know, we found that we can very quickly to achieve significant progress, and then very quickly study. And then Greg and Sam, you know, people with extraordinary support; and also, eh -uh, Greg was very adamant that we will not to focus only on yourself, but also on benefit of the world. So he just encouraged us think about it not only as a tool for OpenAI itself, but also how about something that we actually transform on the product. And exactly then we united research efforts with these efforts A3, and we started creating one thing. And this led to a sprint, which was similar to initial cloud the code that we launched, which didn't actually have PMF, because it was there a little too much friction. And we too launched the CLI code, and then continued to promote, but always there was this idea: how can we to force models really help here?

Host: You mentioned that first started create this model for learning on code Python and actually helped me improve it. But then you accepted it interesting solution that for Codex you created it on Rust. And at that time the model was not in distributions for Rust, isn't that right? She was not as good as was in Python or TypeScript. Why did you accept this? decision? Did you expect? you, that she will catch up, Did you think that productivity more important, or because it was very counterintuitive? Most others built systems actually weren't built on Rust. They were built on the TypeScript distribution, Python or something else.

Guest: Yes. From the very beginning, as before, we thought about the interface product and agent as about different things. Ago it was very important, um, create agent core so that it was reliable, safe, um, designed for efficiency and scaling. AND working on projects during many years, you said, "Hey, this "interesting thing." something on like: "Hey, we need scale this to scale, for example, the largest center data processing". Hmm, decisions taken on early stages, are quite important. Until you you sacrifice too much at high speed. So it's like a compromise. , but we had very amazing developers Rust. Our internal the models were not bad in Rust. Hmm, and then you also receive a lot of, uh, validation during compilation. This, you know, is static. checked and everything like this. And this too great for agents. So it turns out that, you know, it was quite it is clear that Rust as the language will actually be good enough for agents, uh, enough quickly if we we decide to attach to this requires some effort. But, first of all, we focused on correctness, as well as on efficiency.

Host: Interesting. So you say, You know, it's worth it. In your case it was worth it to think ahead about where you want to go, that this thing be, and, for example, about the choice languages. It is clear that for through agents can be rewritten a bunch of things, and it's easier than in the past, but still possible, let's say, save money for yourself in processing, adding the correct one, probably a frame, or, or, well, you know, the basic basis of that, What are you building on? wrong?

Guest: I think we could to succeed if wrote this in TypeScript, or, you know, maybe, even in Python, and then everything would be fine, but then, you know, we would rewrote this in some moment. But to be very clear separation between oneself an agent who can to exist independently from the product, it was very important principle. And if you write everything in one code base one language, then inevitably you you will be a little careless, and, um, you intertwine things more than necessary, and this will interfere further innovations . And it was very important as the edge of rust, in a certain sense, which was very useful for this.

Host: One interesting solution, which you accepted, unique for everyone main laboratories , is a built-in open weekend code, right? The CLI has open weekend code, SDK and server applications also have open weekend code. When and why did you do this? decided? This is not something by itself understandable, especially, You know, we used to go there. jokes about what OpenAI has something closed, but actually everything on the contrary, where is it open, while some competitors provide solutions with a closed weekend code, which again , I think it's very easy understand why you Do you want something from closed on weekends code. Why did you want to open weekend code?

Guest: There was something really It's cool to have a code. with open source code, so that, in essence, you create a code agent. So we thought: well, if you have it, you, obviously, direct This is for myself. And, you know, maybe you can create a community participants who use this for its improvement, and then, you know, you you can do a lot with to learn this. Also, at that time we it was very clear, what if we want to be successful, then open source itself will change, and the role the code itself too will change, so be it part of this community seemed important, not separated from her. I think, you know, difficult to decide problems if you don't you see them for yourself eyes. And one more thing was that it still seems early, but at that time it was very early. Hmm, it seemed like we had some ideas about, how to solve well problems, and we together they developed, with training, research, and all boils down to, to express the model opportunities as flexible as possible and in the best way. But we also don't have had all the answers, and we were very open about: "Hey, this is what it looks like good belt security, this is how we talk about "That's what we think." We did, like, a few very technical, like, deep dives

Interviewer: and blog posts, and we [talked] a lot about it. They talked and thought: "You know, the world is just immense. There is, you know, insanely smart people. We will also to be inspired by others projects with open source, so let's just make it even the playing field and we will encourage a lot of skill and research at this stage." It was, you know, about a year later, a year and a half later, that a long time ago review AI framework. But looking back or taking into account experience, what are your advantages seen, such engineering advantages, team benefits engineers from open source? And honestly, what is difficult in open source, or wrong? Apparently, there is disadvantages. I just trying to get honest opinion on both parties.

Respondent: Yes, there is only one drawback, and it has its price, isn't that right? The advantages are because it's great create in open space. It's also great to have little repository. Every time, when we hire someone, and he joins Codux team, he seems to already seen repository before, he reviewed the PR, he thinks that adaptation is complete.

Respondent: Yes, it's like adaptation is complete. Yes, it's like using Codux to review repository together with you, and you put them question, but this as if it's not a secret the problem you are having you can start right away work productively. We get a lot good contributions. Although we also get tsunami of random things.

Interviewer: Of course, you and everyone others. Yes, open the code changes. I think this is one of them examples.

Respondent: This is true. And then for me, and for many people, it's just brings a lot energy to simply to be part of communities and do direct contribution. Not just to say that we take care of community, but actually do things that, you know, they see, they cost us effort, right? We don't need this do. Disadvantages consist in the fact that this is separate from the rest our code. Sometimes we I have to draw artificial boundaries and work with several repositories. When we are working on something special fascinating, and we create it is open, then, sometimes we find that others copy this, you know, before we can do it to release. And this is a little sad, but also it's like part of the game, do you understand? This is like to build in open space, and it's something like the contract you signed. This is something like, you can copy it. In us too very much permit license, but it pinches a little, when you are working on something and you think: "Oh then, thirdly, all others think: us overload by incidental contributions, and we have to to deal with this additional tax." But it pushes us, you know, too to try to decide problem, right? What, in my opinion, good.

Interviewer: And besides the open code, one thing that surprised me in Codex, and I didn't even know about this is until recently, this is what, that he is not tied to OpenAI models. You can use others models from Codex. You know, if I put myself in place supplier, this maybe not very obviously, because, again, everyone else suppliers, on which I watch when they create CLI, use it with our models. Again, that forced you to decide to be so lenient towards use or permission use your harness with others models?

Respondent: It seemed quite natural if you are part of this communities and create a wonderful harness code, then you think: why do you need it to associate it with your model? It was enough disappointing decision. So it seemed wrong. And in general, I think, I'm trying make decisions that I think: yes, I can simply explain, is that correct? This is the very reasoning that this is from open code. It would be banal for any who to create a fork, and then add support for something another, but then you just encouraging people just use this fork, and suddenly you have overheads arise costs. And the only one the reason why you have fork, is that, what do you want to change about 10 lines code to add support for another supplier models. It seems very silly. So, you know why just not to support him with from the very beginning?

Respondent: Other thing, we get great benefit from opportunities simply to provide optionality. So, today you love use OpenAI models, and you're with them super productive, but a new one will be released tomorrow model, and you will want try it. Why to force you completely change your settings just so that try a new one model? And then we benefit from the feedback we received or did not receive, that is, maybe you liked something in this model, perhaps, she really doesn't worked well. But it's something like of what to be polite to our users and to communities–this seems correct. And then, we too use, we let's try others too models, right? So, we try them in one building, and it's all good. And this is also often it happens that this optionality very important for companies, with whom we work.

Interviewer: Yes. And that's what it's all about. we absolutely rely. This last point, I think, like any serious company, you do want to have optionality and use tool that gives you this option.

Respondent: But I am to some extent I appreciate this because it seems fair. This forces the whole company compete with the best in everything: at the model level, at object levels, with open source, with suitable for using models, and it's not allows you relax and say: "Okay, we finished. We can wait for now." Yes, I want us to conquered users, having the best models, most effective models, the best product. And then, if we do all this, it will be similar to that we will have a good time. If we seem we force you use product, because it's simple... then I don't think so will attract the best people to working on this product; and it looks like on what we do here's my best work. We are very important experience. He should be pleasant, regardless of model that it provides, he has to be pleasant. I like the idea of winning based on merit, not on based on attachment.

Host: And this is the perfect time to remember that our season sponsor Entire also plays the same rules. Do we like it or not, Git is becoming bottleneck for modern software development with using agents. Developers create more code using agents. These agents send more code. Many developers launch more parallel agents. They send more and more code. GitHub explicitly trying to keep up over time and has frequent failures. So, what is decision? Entire was founded by the last general director of GitHub, Thomas Domke, and he rebuilt hosting Git for processing agenda from zero. Entire was created to be very fast and place your repositories regionally close to you to reduce delay, allowing fleets of agents send data in parallel. Some numbers published: Entire can handle 418 sendings per second. That's up to 89 times faster than anyone else competitor in the market. When GitHub is down, you still can continue work, and you not even necessary migrate from GitHub. You just register in Entire, and the platform reflects your repository. And more one interesting thing. You ever wondered what query resulted before creating this specific code? I believe that the request and conversation with an agent carry more information than himself request for verification, at least for me. Entire captures the entire history of requests from your agent right in repositories, easy check and has quite innovative user interface to show all this. If you are looking for Git hosting which works even when GitHub is down, go to entire.io/pragmatic.

Host / Speaker：... install CLI and with one click mirror yours repository. I already did it. did. Oh, and is I remembered that he works with any agent, and he has open weekend code? I would also like to remember our season sponsor, Antisys. Thibaut told about as experience working with software the security you provide you use, has to be pleasant Delightful does not contain annoying errors. But when you use agents for writing your own code, how to avoid mistakes, related to delivery? Audit each line of code becomes difficult because the amount of code that generate agents, that's why Antisys is coming out far beyond the scope code checks. Antisys launches all your system in conditions hostile simulation. This simulation includes both targeted testing, so and vague testing. Running this simulation, it finds each error earlier than yours will do it users. AND because the simulation fully determined, it not only finds errors, but also gives perfect reproduction every problem that makes it much easier correction. First, What did I think when heard about Antithesis, that's it that is automated error detection and fully deterministic testing sounds as a scientist fantasy, but actually it is hardcore engineering. Jane Street, fly.io and etcd community agent wrote code with full with confidence, because know that his checked Antithesis. That view more thematic research and details, go to antithesis.com/pragmatic.

Host / Questioner：And with that let's go back to Tibo and why competition between the tools are wonderful. Yes, and I think that as engineer, I always I see that when there is competition, like that, who uses tools, this always amazing. I I remember how Microsoft fought with JetBrains for the IDE, and then there are the clouds that fighting each other for all functions. AND now, of course, we have there are bindings, we have models, and how for user this great, because now we have more choice. They just developing rather. I think ours the voice is heard a little better, so that's great hear.

Host / Questioner：By the way about strapping, can you tell me how she is working today, that is, when I run Quota task, or it always is is running on my car? Does he choose Is it a cloud? Or it uses sandbox? And how can I to control it, or to know about it, or How much do I owe? to know about it as engineer?

Thibaut：Yes. So, for it is the default works, uh, in sandboxes. Uh, that's all, if there is a team that must be performed with additional outside permits sandbox, she asks you how user permission. But every performance tool takes place in sandboxes for default. And it is fully implemented in the car, uh, on your local car. And this is already happening, you know, more than a year. But this is something that develops and changes, for example, you can choose to run this, hmm, in the cloud, which then works as a managed virtual machine where this is the same virtual one the car you you get through GPT- chat process. And you can you can supposedly, check this out , but it works like container Kata. It's like safe environment. And that's why everything works inside this virtual machine, and not on your car. AND then the only thing that happening in your car, this is your entrance, and then streaming output transmission. THERE ARE -eh, and this, you know, obviously, much better for you processor and your machines, and you can to scale much, much, much more. And this just a step. Uh, this is there will be much smoother in in the future, hmm, use cloud machines, and then, you know, maybe have combination partial fulfillment on your laptop, partial fulfillment on cloud machines. AND really, what we're talking about we think, quite naturally, consists in because since models become better and, um, more powerful, they can use much more computational resources than available on your local computer. Therefore, in a certain moment execution restrictions on your local computer can become restriction.

Host / Questioner：But there is one wonderful thing in that he works locally, and I think that's why I I like it when he works locally. Of course, this is inconvenient. because if I do some work, then, You know, I have several agents, and this consumes processor. If I want to close my laptop, I don't I can leave it. half open , isn't it? When I was in one of the offices companies with artificial intellect, I held half of it open, and they ask: "You Are you launching agents?" I say, "Yes, I have one is running". He says: "I understand." But the reason why I do this I do is, that I have local tools. I have local database Postgres. I do this, this, this and this. What do you think? about the cloud? She wonderful, but in it there is no such thing configuration, or its it's just hard to adjust. You Do you think? experiment, you know, creating these configuration? And me too this reminds me of a topic that we discussed before the emergence of AI, namely cloud environments developments. AND, for example, in 2022, 2023 years they were popular, and then we talked more about AI. But yes, I think so outside the big technological companies, cloud developers never acquired popularity, because are very large initial costs. THERE ARE- oh, and then you too need to pay for service, and you just, you know, not you get from this benefits as a soloist developer or small team. WITH at this level opportunities, which in We have agents now, settings almost free, right? Yes? So, the cost settings and service–this that if your agent capable of doing this, he should just do it for you. For example, if you you say something like : "Hey, you know, I have is local SQLite or local server, MCPS etc. As far as difficult to set up exactly the same system and synchronize it on cloud developer? » Well, maybe it's not so difficult if the model just does This is for you. And that's why I I think we will see. revival, you know, partly cloudy orchestrated machines, who will then be released you from the laptop, or wrong? One of the things that which we saw huge success in working with GPT chat, this is what that she is simply available on your mobile phone. I start my day, just dictating to him a bunch of tasks next to coffee, and he just does this. He has access to my calendar. He has access to my email. He has access to Slack. And it's so wonderful. just have the opportunity to walk and, you know, doing things, not to carry with you laptop everywhere. And I I think it's the same thing, what...You know, we they released that remote Codox, where, you know, implementation happens as on your laptop, but It would be great if, you know, you don't it was necessary hold a laptop open.

Host / Questioner：Or Can you please do a little? tell me how earlier than you improved the Codex? Because I I remember when I for the first time used Codex, it was one of the early versions. You could with him to talk, he did various things. But, for example, I said something like: "Okay , make this change", and he made this change, and I conducted modular tests, and he doesn't started. And then, in a few months, I don't know exactly when, he just started to launch automatically. Were there these things that you improved, you know, script that starts, you know, instructions? I am not sure, exactly how you are You call it a script. initial download or something like this. This is an improvement models, as a developer, How can I imagine that? you improve each version between the strapping, and then between the model, and What is the connection between them?

Thibaut：Yes, that's...Good. question. So, uh, protection system in in a sense always a little ahead model.

Host / Questioner：Oh, really? How is that?

Thibaut：Uh, I mean, that you have a model; she capable of certain things, but then you you set it up, for example, for with the help of several crutches so that she could actually to perform tasks, hmm, on a certain level reliability and, um, in in a sense, hmm, that is, effective, and also with behavior, which you expect as user. And this, yes so to speak, the role protection

Codex Engineer: ...systems, or wrong? It consists in order to ensure protective barriers, to make it more efficient, more controlled; and then protection system usually also is responsible for what we call developer message, which seems to be injected in context, hmm, at the beginning of each turn.

And this obviously affects the goal I see, influencing agent behavior during, hmm, during the move. What you have is the result of using bindings and models.

At first, maybe you think: "Oh, it doesn't conduct tests. So, you know, you need to remind it about conducting tests." And then, you know, we teach the base model so that it is capable of better reflecting what you really want when you ask for something. Uh-huh, and then, you know, it doesn't need to have this explained anymore.

So sometimes we see that the system developer message decreases, and then the bootstrapping scaffolding also decreases.

Interviewer: In the Codex team, do you have specific goals? Are you saying something like: "Okay, now Codex as an agent and the model is not very good at coping with this, does it make any stupid mistakes or what?"

How can I now imagine how the team engineers work on the next Codex version? This is because, as a developer, I do not understand very well what this is: "Okay, there is a model that is magical to me, a thing that will become better."

Of course, I am sure you have feedback channels, but there are also tools that you create. Probably the team responds to this. How do you set your goals, right? As in traditional software development, you would say: "We will build this feature," and you build this feature because you know how to do it. But to me, it seems that this development process is a little fuzzy.

Codex Engineer: Yes, that's right. And exactly so, together we are developing most things. And this is the process where there is cooperation between research and engineering teams—for example, when creating a kernel, the main agent harness.

Always the question arises: "Okay, today we see that we are very good at this, we understand it, but not very good at that. And, you know, we have a desire to do something still, because that would be a very cool product feature."

And then we are always looking at this and thinking: "Okay, does this require a change in bindings or a change in the models?"

"And if it is a model change, for example, how soon can we implement it? Can we implement it in a month? Can we implement it, you know, in 3 months or 6 months?" And we somehow work on it.

And then, depending on how soon we will be able to fix it in the model and at what level of training, we might decide to do nothing in engineering for this process and just wait until the next model solves the problem. You know, that's it.

Codex Engineer: ...full-time agents, isn't that right? So, we use agents for analyzing many reviews, so that, you know, they come up with topics to help us lead these conversations and determine priorities.

But let's analyze this across all coding tasks. Let's analyze this in all other areas as well, such as finance, communications, marketing—you know, in all those areas where our users use these agents today.

And inside them, there are subcategories. That way we know approximately how well we are performing. And then we are always expanding the boundaries.

It's interesting that when we improve our model through foundational training—when we improve the general model—everything gets better. But sometimes there are specific areas where we pay a little more attention.

Interviewer: You mentioned that during this whole time we talked only with agents. Can we talk about the software development life cycle in Codex? In the sense that every time when a new engineer joins the team—any team—it's something like: "Okay, how is everything done here?"

And you know, before the appearance of artificial intelligence, when you joined companies like Uber or Google, they said to you: "Cool, here is how it works: we have an idea, or the project manager has an idea; we make a plan, we do some estimates, we break the work down into parts, we write the code, we run tests, do code reviews, we release, roll out features, and then we take turns on rotation. That's how it was before."

When someone joins the Codex team, even if they have contributed to open source code, what do you tell them? How is work done here, especially if they are completely new?

Codex Engineer: I introduce them to wonderful people, and then what is most often heard when they have a question is something like: "Did you ask Codex?"

Hmm, and Codex is connected by default inside OpenAI, so it has access to Slack, access to all documentation, and access to all code.

And newcomers are still surprised that you can ask it anything, and it very often gives a really beautiful response.

The easiest way to understand the project situation—who is in charge of working on something, or why a particular solution was adopted—is that Codex knows all about this internally.

Codex Engineer: Hmm, you too simply use all of it. We do a lot of work in public Slack channels. We open documents with broad permissions so that everyone has access to this information, and so that your agent can view things and reason over them.

And then, we have a few more things that are very useful for team productivity and collaboration, which we haven't released yet, but some of them will appear at DevDay.

All of this makes you very grounded and tuned to collaborate with the rest of the teams. It allows you to understand the state of affairs very quickly and independently create things.

The general recommendation is: "Hey, take care of the user, take care of product consistency, and pay attention to models and where they are heading. If you are building something and you find yourself writing 10,000 lines of code just to work around model limitations, you know what—you're probably doing something wrong."

So we have a set of principles, team culture, and engineering ethos. When people join, learning these basics is what happens first.

Interviewer: And then, when an idea arises—I think it's a great idea—I discuss it with Codex. Maybe I discuss it with some of my colleagues, for example: "Here is a new class or feature that I'm going to create as my first contribution or first major contribution to Codex. How do I do this?"

Of course, I write code using Codex. Of course, I test it and make sure that it is working. But what's next in the process? Do you still have concepts like code verification, AI-assisted code checks, testing, deployment verification, and phased deployment?

Because Codex itself is spreading among millions of people—I think it just crossed the 20 million active users mark. And if it is ChatGPT, it is being used by an even vastly larger number of users.

Codex Engineer: Yes. Hmm, but remarkably, it is a very similar process regardless of whether you are shipping for Codex or ChatGPT, even though ChatGPT is serving hundreds of millions or a billion active users and that number is steadily growing.

You can create a PR. You can make changes and ship them out the next day or even on the same day. And the change goes live across a vast user base—this is standard practice.

We instill a deep sense of ownership, responsibility, and diligence. So people have abundant opportunities to ship changes, even major structural changes. In general, what is required is solid evidence that the change will be well-received and verified to work correctly...

Guest: Like, a worthy addition; uh, evidence that, you know, it seems worth it to support over time, but also, for example, cost of service, as you know, a lot decreased. So let's think about these things a little differently than, say, 2 or 3 years ago. And we too, we automate the process as much as possible. So a lot of things, for example, process code checks and deployment, and, you know, detection regressions, all of this is almost automated.

And so, you know, you can concentrate on the idea and on how it will help our users. And you are interested in the consistency of everything this, general agent power, which helps to improve things. And we have a long, long list of things we want to do, but have not yet achieved this.

And then there is something like "Polar Stars"—a wonderful, simple in use personal artificial intelligence (AI), which knows everything about you (that it needs to know), has access to the necessary resources, can take, sometimes risky actions on your behalf, but then you get a push notification, and then you can check it.

And this is something that deeply understands you as a user, but also it knows about your schedule, it knows about your goals, it may be proactive, and it must be extremely natural. This has to be something you can manage, for example, with the help of natural language, voice, you know, maybe it has to understand your emotions, for example, if it was in a video call. This has to be the most natural thing on Earth. This must not be something out of 10 different buttons and configurations. The UI should be simple in use.

Host: We briefly mentioned the review, code review, but I would like to come back to that. You were at Google working on a product that is used by hundreds of millions of people, namely Google Maps, and Google is very famous for its very cultural rigorous code review. They have, I think, two code review levels: there is an audit of correctness of language, and they, I think, improved it in the industry for a long time, and they believe that this works, and they use it.

How, in your opinion, is this part changing, in particular human review? A long time ago, maybe a year or two years ago, I would say that code review has all these benefits: knowledge exchange, a second pair of eyes, elimination of the bus factor because now someone else understands it so that when this person is missing, they can step in. Conversations are happening about architecture, not only about the code.

But now, you know, there is much more code. And what is the value of code review? In what cases? And so, on your team, because you guys are so ahead of it, where do you see that people or developers still participate in the stage checks that are valuable, and where is it normal—did you think it is the norm to pass this on to an agent?

Guest: Yes, the role of code verification is changing. One of the early projects which I did on Coding Smith was associated with research model development for code check, which should have been at a level where it could detect errors in logic and reasoning, uh, to such an extent that for people it would be useful, as you know, taking potentially several hours to discover the same level of mistakes. Because it requires real digging, as you know, three or four levels deep studying dependencies, and, you know, just understanding that, perhaps, the documentation actually was incorrect, and the implementation, such as outside dependencies, differed from what you expected, and therefore your invariants are not supported. And these things, as you know, if you are not an expert in this library, you wouldn't know, and that's why you have a mistake.

And here it is: we developed these code verification models, and, you know, we released them, and now they have one of the highest-level opportunities and abilities to detect these mistakes, performing, you know, deep verification or just as part of the main models. When we compare them, it's like superhumans in checking code, and this applies not only to correctness; this also applies to security, for example, where they are able to think about very complicated things, and then, you know, come to the conclusion: "Hey, you know, you have a critical security vulnerability here."

What is now a must for every OpenAI code pull request: we block merging pull requests if, you know, we mark them as a security problem. And that's all automated.

And the role of code checks now—I think it always consists of correctness. It was always about making sure it works, but it was also something like a small ritual of information exchange and, you know, bringing people onto the same wavelength and, you know, encouraging discussions, which ideally should have happened earlier. But sometimes it only happens around the code because after the merger it is just starting in production and doing something. Yes. And then you need to support it. So there is also a social aspect.

I think that all this is changing. Correctness, cybersecurity—I think it will be automated. In fact, what we see, and I see, is that what is happening is some discussion of intentions, um, around the request for implementation. What are you even trying to make? Um, and is that the right thing to try to do?

I think you can have this discussion beyond the request for implementation. It does not need to be related to the code. So, maybe this helps clearly understand, you know, where we need to discuss, and where we did it simply because, maybe, we didn't have such tools that we have now.

Yes, I think that will change. It was similar to a forced execution function, because, you know, you need to have this discussion before you merge it and it becomes working code. But I think there are other methods of conducting these discussions, and, you know, developing things together and making sure that the intention is good. And then the code does not have such a big value.

Host: This is interesting because, when I remember all code reviews, of course, I have memories of when it was wonderful. We had a good discussion, or I learned something really interesting. But many times, honestly speaking, it was so officious. I tried to get your attention, saying: "Hey, could you please review my code?" And you're like: "No, I'm busy right now." "No, I really need this to unblock me." And then you switch context. And then I feel that there were always good and bad parts, right?

Guest: Yes. So I feel that no matter what we do, there will always be advantages and disadvantages, but now they are just shifting. So, I think one of the advantages is that, as an engineer, you will not have to pay attention to such basic things that don't need your intervention as such. Yes, it saves time.

And then, gradually, we will also see that you have an agreement about what it has to do, and what is located inside the box—if you have strict guarantees regarding resource usage, data access, security, such things. It's like what is happening inside the box can literally be whatever. You feel indifferent. And, actually, you need to get to an agreement on what this box actually does and what invariants have to be satisfied.

And I think it's worth having good conversations about that. Perhaps with the help of your favorite agent. But once you understand it, any changes inside the box do not need further discussion, and it simply attracts your attention if needed.

Host: Cost of service decreased. Service and maintenance is always like this hot topic. Every time when we created something inside companies like Google, Uber, even startups, construction was the most interesting part, but service was painful, and that's when we understood: "Okay, it's not worth it to build it, etc." What do you think can lead to the reduction in the price and cost of service within Codex and OpenAI, and what changes in...

Host: In terms of what you create, which ambitions, perhaps custom tools etc.? Technical service is something like the tax you pay over time so that everything works, and it always was necessary and will be necessary in the future. But in my opinion, many changes occur through automation. So, you know, if you have third-party dependencies and you need to update the version number, then this is like: "Ah, yes." You can completely automate this if you have a good changelog and your code is well-documented, and the model can simply analyze this. It can just pass through your codebase in a couple of hours, whereas you would have gotten stuck on this earlier because it's not the most interesting. But actually it is very important for your business. So this is really important for your project, especially regarding security vulnerabilities; you want to be up to date, right? Do you want to apply all these patches? I think so, it will be completely automated. So a significant part, for example technical service, is provided for free, right?

OpenAI Engineer: Yes, and I think so too. It's great to think about what was before: when you wanted completely to react, you had to create a new architecture, because you tried to make room for new, different compromises, or you have a new understanding of the workload, or you tried to introduce a new function, and suddenly you realized that your current system is very limiting and you need completely to rebuild its architecture. It was really a very expensive beginning, right? Sometimes it lasted for several years, and I think that now it is also extremely accelerated. So the cost of mistakes decreases, but at the same time, I would say that the old good software engineering rules, like the presence of good abstractions, indeed help. Well, you know, this is a comeback to before this, for example, the presence of a box of invariants. If you draw the correct form, you can change things much faster within this box and, for example, not influence the rest of the services or the rest of your infrastructure. And I think that this is important. It is important to design considering very rapid iterations and changes.

OpenAI Engineer: I remember how I talked to Peter Steinberger. It was before he joined OpenAI, about OpenClaw and how he is thinking about it. As you know, he told me he doesn't read code, but he was constantly thinking about what he saw, holding the architecture in his head; and he told me how he often does redesigns, and he thinks about how to make it modular, how to allow 100 participants to create their own thing without stepping on each other's toes.

Host: So I hear what you have in mind: this concern, this planning, this structuring has become perhaps much more important. In the past, it was as if an architect, a staff engineer, or experienced people did this, and other engineers were engaged in creating these smaller parts. But it seems that now all engineers need to know this when creating their own software provision, right? And plan for it.

OpenAI Engineer: Yes, and GPT models are getting better and better at this, you know, thinking about long-term maintenance and beautiful architecture, and it's like a natural next step, right? Yes. It's not just about code quality, like does this code make sense and is it clean in this file, but also about whether the architecture is really correct so that it reduces the load on maintenance over time and makes room for future expansions, whether product changes or functions. And just this act of engineering over time is something that models start to be able to ponder very well.

Host: I think it's simply fascinating to understand that in the software provision we create, we pass through the lifecycle much faster. As you know, before you scaled it, you launched it with maybe a small team of several engineers, and then you slowly added engineers, and maybe in a year, if it was very successful, you would have 50 or 100 engineers. You would have time to see it, time to onboard people, and start thinking about documentation and everything like this. But now it is just an explosion: suddenly you have about 100 agents contributing to this matter. This can happen over a weekend. And therefore, you just go through this with great speed compared to before.

Host: Okay, but how are people from OpenAI coping with this? Doesn't it bother you? You understand what I mean: when you've been in this business for quite a long time, decades or more, there was a pace we were used to, and obviously now it is much faster. But how do you realize the fact that what you did a year ago you don't do now, because now the model understands this well? And how do you deal with that? Because I'm sure there are things related to software development that you really understand well, which now you can hand over to the agent. Aren't you a little annoyed? You know, we talked about how it's annoying when your functions are implemented with open source, but this can also be annoying when I'm really good at something like refactoring or architecture, but the model becomes really good at understanding it too. And now I'm like, "Okay, damn. I'm glad, but also it would be nice if I did it."

OpenAI Engineer: Yes, I think that there's something to this like a skill. Sometimes I still open the editor and write some code; it's just nice and I have such happy memories, like late nights when I was sitting there, drinking Coca-Cola Zero, and there was simply no need to think about anything else except the problems ahead of me. But actually I think it all depends on being in flow and solving problems. And I think people here, and everyone I talked to, just adapted very quickly. If you have a mindset where code is a tool to solve problems, you can solve much more problems.

OpenAI Engineer: It's like earlier, when you wanted to compare something and you weren't absolutely sure where it was going to go. Now you can just do it. You won't need more than 30 seconds to launch something in the background, get the right numbers, and find the best compromise. It will make you a better engineer if you really care about the result and that the system works well. So what this allows us to do in OpenAI is that it allows us to execute our inferences much more efficiently. It allows us to get much more efficient computation and then deploy it in the world. So everyone is very focused on this and on solving important problems at a speed that was previously impossible. And I haven't met anyone yet who would say: "Oh, this is not good, this is not fun."

Host: I understand what it looks like: if you have ambitious problems, if you have a lot more problems than you can decide today, tomorrow or next week, then this is not a problem because when you get more effective, you keep moving forward. And that's true for a lot of startups, right? Startups are always much more ambitious than what they are currently able to do.

OpenAI Engineer: We definitely haven't run out of problems, right? And I don't think we will finish them in the near future either. We have a long way ahead of us: a long way from the point of view of mathematical breakthroughs, scientific breakthroughs, and making the world better.

Speaker A: ...better place, simply create for people and decide most important problems with which everyone encounters, and just do it deeply human way. Exactly, that's why we're here. Also, just returning to coding, and how do you... you know, these late nights, I think that's also maybe glamorous version of this.

Speaker B: Yes, same as me too. There was a lot of late nights when I tried something refactoring. Um, and you know, it was like that, like I spent somewhere 3 hours in refactoring, and then I understood that it was deaf angle. Uh, me too had to restart everything from zero, and it was very unpleasant. Um, and there are such very, very funny moments, but there is also a time when something is not compiling, and you just think: "Why isn't this happening yet? Why it compiles?"

I'm sure there was a time when you were leaving sleep later. Now very late. You need to go to bed to sleep, because you need a little to sleep, and then you can't really fall asleep, and it happens that there is some task that half-finished, and it upsets you. Sometimes I remember that I dreamed of a code. AND, probably one thing that I actually have not these days when I am working on software provision for your business, that's what... What do I really have? There is nothing that half finished. Because I can just tell him: "Do this", and then leave him in this condition, what is it, you know, ready. Or ready, either it works or I have proof that it is failed.

Speaker A: But interesting, because, you know, everything has sped up, has it?

Speaker B: Yes, maybe I did. So, how do they do it? Many people, and myself, sometimes I ask myself more important issues, and I, you know, from conversations, which I had during the day, or even I just didn't have time figure this out. And that's why I, you know, I am sending the Code, so that he just looked at him at night, and then I'm very glad to wake up and look at results. And therefore, you know, it's always like a fascinating morning.

Speaker A: Well, I feel like there is a certain way in performing long tasks. And, of course, you can use the target {slash}, which will enter and will start. You know, this is also what recently was added a few months ago, right? Yes? Command {slash} objectives of the Code.

Speaker B: Yes, and returning to the point that, perhaps, this is a crutch, isn't it? The purpose of {slash} was necessary so that, you know, to keep model on the way to single goal during very long period of time. And this as if allowing the model literally work days or weeks, if it's really that bad a difficult problem. But with the new generation models, as we see, you no longer target {slash} is required. You don't need it. Around her is the goal. You can just tell the model: "Hey, go work for a week". And she really will do.

Speaker A: Speaking of complex problems and the fact that that you haven't seen them yet decided that one of interesting things that you offered from outside, I would say that this it was, you know, like engineer, enough interesting, is that what you you call it a merger, i.e. Codex appeared inside ChatGPT. AND the reason why I say it's like engineers, it's quite interesting, so what are we used Codex, that is, yes, he is there. Now you can open it in ChatGPT application. Perfectly.

I just went there and immediately went to Codex, because I'm not... I not really I use ChatGPT in the application as such. But I talked to people with OpenAI and people from your team, and you know, they they told me that long-lasting big preparation, a lot engineering challenges. Can you to tell, how big this project that you it was necessary to do, and why it it was difficult realize? And as Codex and other tools helped you to do it as would it have been difficult before? Because since the launch associations, numbers, which you constantly share about the number of people who use Codex, as if growing much faster than earlier. So, I I assume that you decided some a large-scale problem.

Speaker B: Many complex moments from the association were, according to first, completely different. ChatGPT is completely managed cloud platform, that is, you know, you launch all on our systems. We store things in the traditional way, such as creating things created for scaling, for efficiency. Codex completely local. So the union is how to get the same ones advantages and opportunities from this local agent coding, and then create a product around him and create it like this so that he could to benefit much wider circle of people, that's why we all joined in OpenAI to bring very broad benefit audiences all over the world.

And it was very an exciting journey where we tried understand how create a cloud version of this, which, according to essentially capable of very, very much the same things, but she also built like this in a way that we can to serve her tens and hundreds millions users, when remaining the same effective so that we could completely include it in the plan Plus. ChatGPT's work is, according to essentially, launching a full Codex bundle in the cloud together with the cloud computer. This is very powerful machine, actually. People picked up on it and showed that it is possible make. If you creative approach to the tips, you can force ChatGPT to teach another model, and you say: "Wow. There are some pretty strange ones things. You can install Blender and to do 3D-modeling. This is very everything is allowed. He has Internet access. It's like a powerful machine, and Codex simply works on her, and that's what we implemented through ChatGPT."

So many systemic problems we with the team did very quickly. Obviously that Codex helped to make it more effective, to consider and to build a significant part infrastructure. AND then, you know, helped to solve many small ones differences that arose between Codex and ChatGPT, such as association plugin architecture, association libraries, etc., i.e., really, really working on by creating a single system, the purpose of which is what you shouldn't do to feel that you can to do something in the Codex, what can't you do in ChatGPT, or vice versa. We we are trying to create one single product, which gives you access to the same intellectual data, but as you do you want them use.

AND so it was very fun because Codex throughout this travel too worked as a journalist, who documented all the steps, debates and discussions that led teams. And these were very lively debates, you know, about how we to do it as we please to name a thing and, do you know when this is to enter, and how, and what to combine with what. It was considered many different options. So in this is very interesting journalistic element where we, for example, fully we list that Codex did over time. Uh, and yes, it became known, and also an arc switching OpenAI, where we presented working switch, around which also there were many discussions, is that correct?

Speaker B: Then, you know, we have this simply very liked it. Uh, but over time we will gather to put things together further. So, we really we are moving in the direction full unification, and, you know, we we consider this as temporary state when you have better and stronger capabilities in working mode, but over time we we implement this, you know, for everyone who uses ChatGPT.

Speaker A: How about you personally? Do you use Codex? What your work structure from point agents' view, tasks, do you manage it? AND regarding this I asked Peter Steinberger, what do I need should ask about you, and he said that, as well as you have, you need ask him how you are cope with it, that you are involved in all these projects, your calendar is similar on Tetris, but usually you you look pretty cheerful.

Speaker B: My calendar decent. Hmm, hmm, and simple, I can do it much more things now, because in I have these technologies like Codex, and I actually significantly switched to his own work on mobile phone, using Touch of the Work.

Thibaut: Whenever I have something I want to write down, I just launch this. I often use dictation. Every time when I have a question, instead of recording it to consider later or delegate to someone, I just launch it, tell it about the work, and get something like a report. It has a whole bunch of special skills and special instructions that are now very adapted to create reports, slides, and code research in a style that I can effectively use.

Thibaut: And so every time I am between meetings, you see me as I just dictate to my phone. As I already said, we work in public channels. We have a lot in Slack, we have a lot in Notion and Google Docs. So practically there are no doubts that Codex will be able to handle at least the first stage of deliberation.

Thibaut: Whether it is public opinion regarding a feature, viewing production logs, the usage intensity of certain things, compiling a list of items we should abandon support for because they no longer receive support, or understanding what a specific person on the team is doing. Whatever my question is, I can get a response within 30 minutes. That is how I use it. I use it for everything. This is like my personal agent for everything.

Thibaut: And then often on weekends, I do code research or create prototypes, and it is fun for me to imagine the future of the product in a certain sense. I do it with others across teams; it is not always the same team. And simply, in one day, I can create things that feel like they were already in my system.

Thibaut: If I wake up one day and think that something should be investigated, that means building it. And then I can just express all of this and present something to people in a day so they can think about it, critique it, and hopefully be inspired by it. It does not mean this needs to be released to the market, but it is more like: okay, I will get this out of my system, and then I will continue doing other things. It is such a magical time, and it is so inspiring.

Host: And finally, what would you advise a software engineer or AI engineer who wants to acquire the skill set and experience to work in a place like the Codex team at OpenAI, or in AI startups—to become a truly wonderful builder with the help of these tools? A question that often arises is: "Should I start with theory? How important are the fundamentals? Or do I just need to learn how to use the tools?"

Thibaut: Yes. I think there are two important things: deep curiosity about how everything works, and the ability to learn and understand things very quickly. Things will continue to change. But the people who cope extremely well are those who are able to quickly decipher a system, immerse themselves in a new codebase, and understand it.

Thibaut: Modern agents obviously help with this, right? But there is so much information that needs to be learned, understood, and analyzed. A lot of this comes down to asking good questions about how these things work and delving into the "five whys," which allow you to keep digging and digging. You learn very quickly thanks to that.

Thibaut: The other thing is to be in harmony with the community or the people for whom you are trying to solve problems. Not everything looks like a straight line to problem-solving. Sometimes you solve a problem that becomes useful for different groups of people. But having clarity in taste, understanding needs and requirements, thinking clearly, and executing through that clarity of thought seems very important to me.

Thibaut: For example, if you cannot explain what you are trying to achieve, if you cannot explain your intention, if you lack connection with the community, or if you do not have taste, it will be much harder to do wonderful work.

Host: Excellent, Thibaut. Well, thank you very much for this conversation. It was great.

Thibaut: Thank you for having me.

Host: I always wanted to meet Thibaut, and I am glad we finally made it work. I liked how Thibaut talked not only about the advantages of open source, but also about the disadvantages. Most notably, how competitors can copy features that you are developing publicly and release them right before your release, and how striking that dynamic is. In addition, you get a lot of low-quality contributions that you have to figure out how to handle.

Host: Another interesting moment was regarding the model side. Internally, the Codex team sees its task as creating scaffolding for models using guardrails, tools, and custom instructions. Then the next version of the model will be trained such that it requires fewer of these scaffolds. Honestly, as a developer, it sounds a bit demotivating that the things I build for the current model might simply be natively understood by the next version, making our work redundant.

Host: Furthermore, I suspect that it is not just about creating these temporary connections, but about creating tools that models will use; it is unlikely that the next version of the model will completely reinvent the MCP protocol, tool scaling, or plugins. At least, I hope not.

Host: I also liked hearing what the merger between ChatGPT and Codex looked like from the inside. It was the integration of what used to be a completely local coding agent, Codex, into a managed cloud application. And it was done efficiently enough to be included in the $20 monthly OpenAI plan, which is impressive considering the computational resources involved.

Host: It was quite interesting to hear how Codex acted as a chronicler for the entire project, being present across all Slack conversations and documents to record all important debates and decisions. I won't lie, this part reminded me a bit of Big Brother, with AI always watching, but it could very well become the new norm for startups in the future. I haven't yet decided what I think about this.

Host: And finally, I appreciated Thibaut's advice for engineers: to succeed, be curious, quick to understand things, and in harmony with the group you are building for. It is also reassuring to hear from Thibaut how important the fundamentals remain. Check out the show notes below to learn more about how Codex, Claude Code, and Cursor were built, as well as other related topics. If you liked what you heard, please leave a rating on whichever podcast player you use. It means a lot to me and to the show. Thank you, and see you in the next episode!