Head of ChatGPT & Codex: AI Agents for Everyone Are HERE | Thibault Sottiaux OpenAI

0:00Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the Unreal College Deal. Everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 Premium and a year of Xbox Game Pass Ultimate with a custom color Xbox wireless controller. Learn more at windows.com slash student offer. While supplies last, ends June 30th. Terms at aka.ms slash college pc. Today we helped a Latte for Sam coffee shop get an insurance quote simply and easily and made sure a floral delivery van was able to make someone's day.

0:41We're the Hartford, with decades of experience insuring millions of unique small businesses. When it comes to your small business insurance Thank you. one size absolutely does not fit all. Get a quote or find an agent today at thehartford.com slash smallbusiness.

1:00Thibault Sottiaux:There is going to be dramatic change. Everyone is going to get their own little personal assistant on their computer that allows you to do more than whatever was possible three months ago, six months ago. This is Thibault. He runs ChatGPT, Codex, and the APIs at OpenAI, the tech powering most of the AI you already use every day. And he just told me that in a few months, people who don't use AI at all will get the same benefits as the ones who've spent the last two years figuring it out. Today, you have to actively prompt and be creative about what you're asking. In the future, that's not the case anymore.

1:31What's going to make people brilliant, I guess?

1:34Thibault Sottiaux:If you're a creative person with good ideas, it's a great time to experiment. Can you show me a workflow everyone should be deploying today? One thing that I think this will mean is that... My audience is, some of them are developers, but a lot of them are knowledge workers. And they're very excited about AI. And what we've been seeing with software engineering, right? Google says 75 % of their code is now AI-ridden. So it has been really transformed by AI. And you said something that this transformation will extend to all knowledge work over the next six months. So what do you think is going to be changing for people who are knowledge workers?

2:11Thibault Sottiaux:I think it's not that much that people are going to change as the technology has matured. And now agents are reliable over long horizons. They are capable of using many different tools, computer use being one of them, browser use. we have added over like 100 different plugins that can tap into every little tool that you already use in your life. And the agent in GPT-5-5 is like extremely reliable at it. So it's more that the technology has matured and is ready. And so everyone will be able to get benefits from an agent. Whereas before you had to be a little bit technical and get in there if it wasn't reliable, if it was like struggling five minutes in, you maybe have to go into the configuration.

2:52Thibault Sottiaux:And obviously that requires like a technical background. That's not the case anymore. So that's why I think it's going to become like very widespread. What do you think is going to change the most? And say, like, if you take a marketer, what's going to change their day to day with agents? So you're going to look at what can you automate? Oh, maybe I spent like one hour, you know, doing market research. Maybe I spend one hour, like, summarizing a whole bunch of inbound and emails. Maybe I spend like, you know, two hours going through like prospects. And so a lot of the concepts that were like, you know, advanced concepts, maybe six months ago, such as running something on a cron schedule.

3:26Thibault Sottiaux:Nowadays, it's just in the app. You can just say, run this every 12 hours, do some market research, send me a PDF. And it's going to send it to you over email? Yes, and I can send it to you over email or you can just consume it in the app. Maybe you can print it. I did this a couple of weeks ago where I would summarize all the news from Slack and then print it on my printer physically like every day. So I would have it with the coffee. Oh, wow. This is like very executive way. It's kind of like, you know, maybe old school. I just kind of want my newspaper. This kind of thing is just going to be mainstream.

4:02You know what I was really inspired by? I was watching one of your podcasts where you were talking with Greg about having this dashboard. Basically, when you wake up in the morning, you have your coffee and it just tells you, like, can you approve this? And I'm going to run it. So you just basically read through whatever your agent did while you were sleeping and you're approving and it's continuing its work. When do you think it's going to be a reality?

4:24Thibault Sottiaux:It's not that far off. The technology already exists. It's about packaging it. For example, last week we released auto review, which is a concept. It definitely blew my mind when we did the research on it. So you have the main agent doing actions. And then you have a second agent that is verifying all the actions of the first agent and verifying that they are not doing anything that could be harmful to you and are low risk. And so that's an innovation that came out of our safety team and alignment team, primarily like alignment research. And it allows you to run an agent much more like for longer periods of time autonomously, even handling quite sensitive data without risking, for example, your agent like sending an email to a stranger with some of your personal information in there.

5:11Now that you have these pieces to make it more secure, more safe, and you're willing to give access to a lot of things in your life. How do we organize our data? Imagine we're all getting ready to deploy our first agent, right? Right. And the thing is, agent needs good data. And sometimes, you know, I've been working with mostly my tone of voice, like personal dossier. But then I realized that my agent would really benefit from knowing my strategy when it creates content for me. So what would you say to someone who's just starting out? How do we organize this data? What are the must-have files that they have to create?

5:46Thibault Sottiaux:I organize everything on local files on my computer. So I have a folder. I have my notes in there. I keep it fairly tidy. And I use my agent to organize it. And then it grows over time. That's my question. People work with local computers. I have a laptop that I travel with. I have two Macs, one in my studio, one in my office. Yeah. So what we're going to see over the next three months is all of this is going to come to the cloud. And you're not going to have to manage local files on just your laptop. And then if you happen to travel and you're on your phone, now suddenly it's not the same agent.

6:21Thibault Sottiaux:And that's annoying because you have two separate entities that you need to map to different things in your life. So a lot of this is going to come to the cloud. It's going to manage its own memory. It's going to help you manage files that are just hosted somewhere. There's no good solution yet, which is sort of surprising. Yeah, I'm just putting everything into my Google Drive, and I'm connecting a Google Drive folder. Yeah, that works. That works. What are the files that everyone should come up with for their agents? If we're talking about productivity for knowledge workers. Yes, tone of voice is a good one.

6:56Thibault Sottiaux:And for that, maybe surprisingly, I don't recommend actually trying to explain your tone of voice. I recommend just including examples of past newsletters maybe that you published or snippets from recordings or messages that you sent to friends and in different professional and personal settings. I have a lot of notes around projects that I have. Each project has its own folder with various files in it. Yeah. And then contacts. But you don't have to maintain everything in files. You can also just rely on all the productivity apps that you already have. And then Codex, for example, will be able to go and pull the right information.

7:41STEPHANIE WONG - How do you know when you stay inside a project or you want to Codex to actually build something, like a tool or an app?

7:47Thibault Sottiaux:JOHN MCCURRY - Sometimes I feel like I need a dedicated app for a new use case. And then I go and I create a new folder. And then I just experiment in there. But nowadays, more and more, I feel like I need less of static apps that just can do one thing. And I rely just more on my agent to do everything for me. If somebody starts using agents and somebody doesn't, what's the difference in their productivity in, say, one or three years? One to three years is so far off. The technology itself will change so much by then. And what we're going to see is, you know, like people who are willing to adapt and discover these things are, you know, going to be quite a bit more productive.

8:28Thibault Sottiaux:And it's going to enable to do like, you know, all these things that you sort of like, you know, maybe put off. Everyone is going to, you know, get their own like little personal assistant on their computer, doing all of the, you know, maybe filing your taxes, setting up like email filters, telling you and helping you like get more in touch with like, you know, your loved ones. That is something that will connect a lot of people. When I started using more AI, I felt this responsibility dilemma. You just mentioned like taxes, right? Right. Yeah, AI can totally do taxes for me. And I have this automation that tells me every month how much tax I'm going to owe and what tax strategies I could be deploying.

9:05Do I want to be responsible for that? I don't know. Like when it comes to these tasks, also as an entrepreneur, I totally realize if you're just starting out today, you're much more productive. But if you're an existing entrepreneur, I wouldn't really fire anyone on my team because an AI agent can run it. I would just ask that person to use that AI agent. Where do you see this go? Because I feel like when I started running so many agents, I feel responsible that I need to check on them. I need to verify their output. I'm like, oh, I'll just let somebody else handle it.

9:34Thibault Sottiaux:That's interesting. So at the end of the day, humans remain responsible. And then the way that we think about it is actually just really about augmenting your own capabilities and giving you a tool that allows you to do more than whatever was possible three months ago, six months ago. And take a lot of the boring parts of the work as well and just do that automatically for you. It is important as we see this for coding. For us, it's like if you produce a piece of code, you are responsible for it. If it breaks, it's not the agent's fault. It's your fault. Same for code review. And you cannot outsource understanding.

10:10Thibault Sottiaux:At the end of the day, humans are still, remain in control and remain the people who need to understand the entire system and how things work. And so if you view it under the lens of, you know, augmenting your own productivity as a human, it's like, you know, obviously, you know, you need humans. After all, we're building for ourselves to better our own lives. Yeah, it's just I feel like I've become much more productive first, but then I'm like, because I want to optimize so many things, my brain just started exploding. There is a there is a pitfall of like, no, maybe, you know, falling too much into, you know, using using it for everything and feeling like maybe you can do things, but you're a little bit too early on the capability curve.

10:47Thibault Sottiaux:And so like maybe things, you know, will be possible with future models. That's good. That means that you're pushing and you're sort of like, you know, discovering, you know, where it's reliable, where maybe it needs a little bit more work. And maybe this will be something that you can do in three months, six months. When you finally find your thing, you want the whole world to know about that thing. So you use a thing called Canva to make it an even bigger and better thing. Whether you want to create flyers for that thing, make presentations for that thing, or design merch for that thing, you can do anything.

11:18So people can see your thing, feel your thing, love your thing. The next thing you know, it's a thing. Canva, the thing that makes anything a thing. You can't reason with the sun. Trust us, we've tried. This summer, it's time to put that angry ball of fire on mute. Columbia's OmniShade technology is engineered to protect you from the sun's harsh rays that can burn and damage your skin. The sun is relentless, but so is our gear. Level up your summer at Columbia.com to spend more time outside and less time slathering on aloe lotion. You're welcome. Columbia. Engineered for whatever.

12:01Okay, let's talk about VipeCoding in general. So I shared before the podcast that in LanguageRiff, we're actually VipeCoding a lot of different apps. And we have a technical person who can actually write code and he's really good at it. But then it's also us who can't. And when we vibe code something, it works great, but then the technical person comes in and says, oh, I would have structured it in a different way because if we wanted to grow it, like, for example, we have this app that is 300 most useful words in English, right? And it was vibe coded in a perfect way, so it's usable. But if we want to extend it and make it like a thousand useful words, then the architecture is a little bit off.

12:39What would you say, like, how would you approach vibe coding these days? If you're thinking of building something like it's going to be a big company, should you work with a technical person or still try and bi-code it?

12:49Thibault Sottiaux:If the goal is to share it with, you know, a couple of people and like try something out and, you know, build it for the joy of it, you know, definitely like just coding it by yourself with the help of an agent is like that's perfectly fine. if the goal is to scale it and, you know, make something that is successful, you know, for like, you know, hundreds of thousands of people or more. Having a technical person like in the mix is still useful. I do expect that, you know, over time to go away where, you know, you will have agents which are able to like, you know, understand the long-term maintenance aspects, you know, the right structure and be able to like, you know, handhold you all the way to scaling to a very large and successful product.

13:31What do you think is the timeline? Because it's been changing so fast.

13:35Thibault Sottiaux:We will see significant improvements towards long-term maintainability of code over the next six to nine months. In order to not need the help of someone more technical at all, that feels quite far off. Yeah. What do you think in general about the future of software engineering? I was just moderating a panel around it with, you know, deep learning AI. And we're talking about like how we all had a consensus of six people on the panel that you just you still need technical knowledge, but you also need to become a generalist. And because everyone is vibe coding apps, like as a creator, I would have never thought of building an app 10 years ago because I knew how expensive it was.

14:19Now I would build an app, but then I will see thousands of people using it and I will experience some bugs and I will need a technical person. Are we going to need more software engineers or less with this?

14:28Thibault Sottiaux:Yeah, I think we're going to see an explosion in the amount of infrastructure and apps that we built. And nowadays, it's awesome because if you just have an idea, you can build it, or at least in some shape you can build it. If you're a creative person with good ideas, it's a great time to experiment. and then you can quickly prototype, iterate, and get to something that you feel, you know, like might be successful at there. And then it's hard to like put a limit and, you know, how much software is enough? How much demand is there for more software in the world? Like how many more problems can we solve?

15:07Thibault Sottiaux:It feels like we're at the very beginning. It feels like we keep coming up with problems we can solve. Exactly, right? So it's like, you know, as long as that continues to be through and, you know, we're able to continue technological advancement, It feels like there will be demand for a lot of technical folks. Yeah. Can you show me a workflow that everyone should be deploying today to maximize their productivity? Can you think of something? It really varies. It depends. One of the things that I do is I use it to stay in tune with what the world thinks, for example, about Codex. And so I receive a daily summary of the highlights.

15:42Can we build something for a museum like that? So let's do a new chat. For example, I am on a podcast talking about AI and codecs. Find relevant emails in my inbox and prepare draft replies based on what you know about my priorities from memory. So you haven't built memories yet. But if you use Codex a lot, it will learn about yourself. But it's going to pull memory from ChatGPT as well, right? Yeah. Another one is go through popular internet sources and give me a breakdown of what Codex shipped over the last two weeks. we can go back to email um i have a messy inbox in for gmail um set up filters for it to keep you know remember i think it was like 2019 or something i spent a couple days just organizing my emails and folders and then i realized it's such a waste of time because i I want to plan a trip to where should we go?

17:09Based on my calendar availability, right? Yeah. Or something. Oh, yeah, yeah, yeah, yeah. Because it can connect to my calendar. A trip plan and use my calendar availability to plan it, for example. And so it's going to go like here. You can see it's gone through your email. So this one is like I'm a podcast talking about AI codex. It's already gone through your email, found several live threads related to today's thing. It went through the calendar. It's all still going through email. So it's found a bunch of things. Apparently, you have a night with Codex. Yeah, and I with Codex, which is today.

17:57And then here you have another one. So you can see we're doing a lot of things in parallel. So you can just run little agentic threads. Here, you could say, I am making a presentation. Put this in a Google slide. And it will go, and it will just prepare a Google slide for all of the content. STEPHANIE WONG - The editing is going to be within the app, or I will have to leave and go to Google slides. So there is an in-app browser as well. I don't know if you've set up your pet. No, not yet. So we also shipped this cute little feature. Oh, I just saw it on your X. So now you can move this across, and you can see which threads it's still working on.

18:41So for example, it's still summarizing Codex updates. It's currently setting up some email filters. I don't want to make changes to your inbox, so I'm just going to stop this one. But yeah, this is sort of like a brief. Can you talk to me about this dashboard that I was trying? You have a trip to Crete's plant now. Nice. So it went through my calendar. Okay, so this morning I actually asked my chat GPT, based on my queries, what app should I build? And it gave me five ideas. And this was one of the ideas where it's content repurposing. And can you rate it? Like my, so it was basically a very basic prompt.

19:23I gave it a link to my MD files. Oh, that's cool. I created this app. But the only thing, I wanted to be able to dictate into the app so it can convert it into an actual email newsletter. But I don't think it worked. Okay, so let's have a look. So this was content repurposing engine, you know, from all the posts. Okay, and then it just built all of that for you. Well, it built something too fancy. See, I'm like, let's just make it very simple. You were like, okay, this is like you're being a little bit too ambitious here. Yeah. Yeah. And then it has, oh, I see, I see. A-B testing. It's like already a newsletter draft.

20:04Yeah, and actually I like that it didn't wait for my prompt. It just went ahead and generated. It just did a lot of these things. By itself, it was like, you know, this is probably what you need. Yeah. And then the only thing that doesn't work is like a voice input. I want to be able to dictate into the app. Yeah. Yeah. Is it something that another prompt is going to fix? Yes, yes. So we have voice dictation here, right? No, I mean into the app. Yes, yes, yes. So let's use it. Let's do a nest egg. Okay. So in the app, we don't have dictation working. So can you integrate it with the OpenAI speech-to-text API?

20:44For that, you will need an API key, but you can just make it up for now.

20:47Thibault Sottiaux:And you can go through the latest documentation online to figure out how to integrate the OpenAI speech-to-text API into this app. That was a very sophisticated prompt. How do I know that I needed that prompt? Because I would never think about API keys. That's why sometimes you still need someone who is a little bit technical. But a lot of these things that right now maybe are still a little bit in the realm of like, oh, you need more of a technical background are going to come and become mainstream. So maybe you need me today, but in three months, it will be able to set all of this up. It will just connect it to your OpenAI account.

21:25Thibault Sottiaux:And you don't need to set up a separate API key. And it's going to be able to just connect to everything. Okay, while it's building, can you talk about the most sophisticated use case or maybe the use cases that are very sophisticated in terms of workflow, but that change productivity tremendously for people day to day? Something that everyone could get inspired by. So you can say, like, you know, go through, you can tag a lot of different plugins. So you can say Gmail, Calendar, if you have, like, Docs. Yeah.

22:02And be my chief of staff and give me a breakdown of my day, a summary of what is important and prepare me and then you just run this at the start of the day