Codex for Everyday Work: AI Agents Beyond Coding

Hinzugefügt: 2026-05-21

[00:00:09]Good afternoon. Welcome to OpenAI Forum.

[00:00:13]Uh my name is Chris Nicholson. I'm with the global affairs team and I'm glad to be here with all of you. So the forum, as some of you know, is a place where we talk with experts about how AI is being used in the world. Today's conversation is about codecs and why it matters beyond software engineering. So more and more people are using codeex to help with knowledge work and personal work.

[00:00:36]They're using it to reduce friction, to perform tedious tasks, or just to understand a problem, organize information, and plan a document that they can share.

[00:00:47]So that makes codeex helpful for researchers, educators, operators in industry, small business owners, leaders, knowledge workers of of all kinds. Um, and today I'm joined by Tibo Sio.

[00:01:01]>> Thank you for coming.

[00:01:02]>> Of course. Um, Tibo is the head of Codex at OpenAI and he's going to help us understand where Codex came from, what the team is learning from people using this tool, especially non-technical users, and how folks can begin to use it outside of coding. So, TBO, great to have you here.

[00:01:19]>> Great to be here.

[00:01:20]>> Yeah. Um, well, let's get started. So, to ground this, Codex was created as a tool for developers. Um, maybe you could tell us when did Codex first launch and what problems were you and the team trying to solve at the beginning? Yes.

[00:01:32]So the journey started a long time ago at open. We've always been fascinated with uh helping just figuring out like how we can create very useful models that would then in turn help us you know also like accelerate the development and you know one of the grand challenges for a while has always been uh coding and being able to code at the level of a prolific software engineer. That was like a grand challenge. um like roughly like two years ago when we started on this journey. The first public version of Codex was something that we now refer to as Codex web. It was this idea that we would have u an entity running in the cloud accessible over a web interface.

[00:02:18]You would put your task in there. It would go and look at the repository like a code repository that you have. it would figure out what the right set of changes were in order to fulfill the task and it would open a pull request on uh GitHub. So it was like fully isolated fully packaged. You would just state your intent and then you would receive like a code change at the end and that was like the entire thing that we shipped. That was like roughly a year ago. What we found was um it was a little bit too high friction. um it was too hard to set up and then also a lot of developers just have you know a beautiful setup on their own computer and you know you needed to go and like reproduce all of that on our own cloud environment setup. So the friction was too high. Um the models were also not at the level where they would you know get it right all the time and so it was kind of like hard to iterate there with the model and we shifted our approach after that which I'm glad to go into.

[00:03:09]>> Yeah. Great. So, so you were building a tool for yourselves and it sounds like there was a pivot to where you realized it it should work on everybody's machine locally.

[00:03:17]>> That's right. We realized that doing this on our own cloud machines and asking everyone to just configure it there was too difficult and because the models were not yet at the level where it was reliable for long-term horizon tasks. Yeah, >> it was just not working.

[00:03:31]>> When did you first see people using it for tasks outside of coding? Oh, that's that's been um more of a recent shift in the list last uh six six months I would say. Ever since we released GPD5, we've had that step change in like generality and reliability for general purpose tasks. I would say around 52 as well where it became more reliable for long horizon tasks. And then one thing to realize as well is like even for software engineers and technical people like the majority of the day is not actually coding. Majority of the day is like looking at you know tickets prioritizing having discussions around you know what how how we should actually solve problems deciding on like the architecture >> solving um you know like bug reports investigating bug reports and figuring out you know whether it's actually true whether it's a user problem where the problem actually is dealing with outages and you know being on call and then you know just like general like having an understanding of the systems and doing a lot of information gathering I would say like maybe you know software engineers spend like 20 30% of their time actually coding uh and so what we saw is like a lot of the people who had adopted Codex were already you know the technical ones were just using it >> for their day-to-day work >> uh and then you know it was obvious that this we were holding sort of like a technology that was like much more powerful and that we wanted to you know was yearning to be useful for you know >> at scale you know for for the entirety of the society. M so it sounds like to be a software engineer you have to do a lot of things that aren't coding and and people were dog fooding they were using their own tool to these to do these non-coding tasks >> and in order to be useful at coding we were okay you need access to more context maybe if you have access to you know all of the information that we have in notion you know information that you have in documents um then actually we saw that the agent would become more useful >> at solving the task >> and then you know we for trying to just make that more and more reliable up until the point where now like the majority of the tasks that are being performed in Codex are actually non-coding tasks.

[00:05:40]>> That's really interesting. So it first of all Codex was searching code and then you realized it's also super useful to search a bunch of documents and come back with information and obviously that's something that all of us doing knowledge work have to do.

[00:05:54]>> That's right. Um so so maybe what is do do you remember the first moment when you were like oh this is actually going to be for everybody right when when that when the ambition changed. Yes, we were in the middle of a launch for Codex and uh Alexander Amirus who is like the the lead uh product manager on Codex was was using Codex to keep track of the state of all of the changes that we were trying to land >> um in in in in the run-up to the launch.

[00:06:27]And I had I had never seen someone be as productive as Alexander at that moment where it seemed like he had many little codeex agents, you know, doing his work on his behalf and chasing people up >> and then updating a document with the latest state of things of like, oh, you know, it's like we actually still need to polish like this specific feature. We have this user feedback, we have that.

[00:06:48]And it was sort of coalesing all of this information from all the user feedback that we have, the information from the developers and keeping like a plan, you know, neat and up to date while Alex was having a discussion with me. Um, and you know, I felt like, wow, you know, it's like we're really changing everything here, not just software engineering. So, so in the world before Codeex, Alexander would be either himself digging through Slack channels or documents or GitHub uh PRs and just like everybody like spend so much time with coordination and then you saw he had somehow delegated that really time consuming work to these tools and they're doing it while he's in a meeting with you.

[00:07:25]>> That's right. Um our models are very good at generate uh gathering the right context and summarizing as well. And so that's a strong use case that we see and he was doing that and then also using it to chase up information. So we have it connected to Slack. It's able to send in to send messages. The Codex agent can send messages to people to you know also ask like what's still his status on you know this thing. So doing all the chasing on behalf of Alexander.

[00:07:52]>> Chasing takes so much time. So so there's Alexander. He's one of the most productive people you've ever seen. He's using codeex in this way.

[00:07:59]>> That's right.

[00:07:59]>> What's the ripple effect like? what's the in concentric circles around the people around him, the team, what gets shipped?

[00:08:06]>> What the ripple effects that we've seen as uh engineers have been greatly accelerated and are now building faster than ever. um adjacent jobs are are changing like the role of the designer is changing, the role of like the product manager is changing and then you know we looked at how can we accelerate and and um give them also the means to you know just be more productive. And then we saw bottlenecks shift to um communications and marketing um where we were only producing so much work um and you know it was a challenge to sort of like tell the world and you know keep the story coherent and um so the bottleneck was shifting to other departments uh at the company.

[00:08:51]>> So OpenAI is shipping really fast. Do you would it be possible without codeex?

[00:08:57]>> Oh I don't think so at this point CEX is critical for us. Yes.

[00:09:00]>> And do you think other companies >> maybe with 10 times more engineers?

[00:09:04]>> Do you think other companies are similar enough to open AAI and like the problems they're trying to solve for the world that they can follow a similar playbook?

[00:09:12]>> Yes, I think now we the state of the technology is uh has reached a point where these agents are capable of very general work.

[00:09:22]um there isn't really many things you know that you can do behind your computer that the agent isn't able to also help with and so you know a lot of like you know the work that needs to go into like preparing a presentation to you know align you know stakeholders for example gathering context on public perception doing some marketing research organizing it's used a lot in our finance department Sarah Friars talks about this a lot but she organized the latest fund raise with Codex with the help of Codex and it was a great accelerant there as well. And so there's this beautiful generality to it and it's evolved like you know much more to like it's not for generating code it's just to perform general purpose tasks really.

[00:10:03]M so you mentioned that roles are changing.

[00:10:05]>> Mhm.

[00:10:06]>> What kind of changes do you see?

[00:10:11]>> The change that I see is that everyone has to deal with things moving much more quickly uh and adapt much more quickly as well.

[00:10:26]um hard hard problems things that used to take days now maybe take like a couple of hours across the field we're seeing this in science we're seeing this in engineering but we're also seeing this of like if you know I would have expected like doing an in-depth marketing research or you know analyzing the public sentiment around a new feature you know that would have taken a lot of time to go and like find all the sources and summarize and then uh condense it in a way where you know like different people could consume that now that's that's just really being compressed to like a couple of hours of work and you know being completely automated away. It does mean that everything is suddenly like moving much faster as well and you need to adapt to things you know just the the pace have been picked up people are able to do much more by themselves as well and you know there's something delightful about that too it's like it's very empowering and so maybe before you were like oh I need to go and talk to someone about this no you can just do it um we're seeing this a lot on uh data questions you know within within companies where it's like you have a question about oh you know how successful are we in this market um are we growing uh in India? Um you know what what's happening in Korea?

[00:11:37]And then now everyone is just empowered can ask Codex a question of like hey can you pull up the dashboard I don't really know where it is but like go figure it out and then you can go and like dig into um you know all the details of the business and you don't need to go and talk to like the data analyst team. It's like you can leave them alone. They can do more interesting stuff than answering your you know little questions.

[00:11:55]>> Yeah. I I remember the days when data analyst teams had long cues because they're a shared resource of things they owe to all the other teams. And as long as those other teams have access to the underlying data, very often they can extract insights by themselves. I feel like we're in this historical moment where it used to be that the people who had problems and the people who built solutions were two separate groups of people and the conversations they had about the product were really long and you you eventually got to something that was kind of okay, but you couldn't push it further because nobody had time. And now now like this the person who has the problem can build the solution very quickly and iterate through the changes they need.

[00:12:29]>> Yeah. That's why I see with designers as well on the team they're pushing code.

[00:12:33]>> Uh they're making changes. They they're in there with the developers uh shaping the product as they want to like know to kind of bring the best out of it and create an amazing experience. they don't have to go and like convince you know the engineering team you know to prioritize like you know what the engineering team probably feels like you know are fairly minor changes but you know for our amazer designers you know they're like oh no this is this is like just really elevating the craft and the experience you know for millions of users out there and so it's very empowering to them too.

[00:13:03]>> It feels like we're kind of in this era of home-cooked personalized software on so many levels. Yes, it feels that this is the wave that's coming where everyone will be able to have their own personal software and maintain it and it just does exactly what you needed to do.

[00:13:19]>> Yeah. You you were telling me before we started about something you built for yourself that visualizes some data. Do do you want to do you want to share that with us?

[00:13:27]>> Yeah, sure. This is just really a fun uh exercise. I have it on my screen here and I'm in I'm in I live in San Francisco. So do you, I think. And I think the price of bread is uh outrageous in San Francisco. I moved from Europe. Um and um so I was like, "Hey, Codex, can you find like the best loaves in the city and create me a spreadsheet uh so that you know I can understand like where are those loaves?

[00:13:54]Where should I buy them? And uh you know, what's the associated price?" And so it went and it worked for five minutes and it created this uh this spreadsheet which you can see here.

[00:14:03]>> Can we see that? Can everybody see that?

[00:14:04]Great. Um, and so you can see you have Jane the bakery, you have arzico, you have tartine.

[00:14:10]>> It has found like the different loaves, a little description and the price. So it seems like if I want to go and find a good loaf of bread, the sardo loaf, I should go to neighbor uh big house which is in Dock Patch.

[00:14:22]>> And then I was like, okay, it's great. I have the spreadsheet. I could have iterated on the spreadsheet and asked for more information. But instead, I wanted to consume it in a bit of a more of a visual way. So I said, hey, how about you create me a web page?

[00:14:35]Um, and so it did and it just put all of these on uh on a map here. So I can uh just and this is just really personal software that was created for me. This took four minutes for Codex to just generate.

[00:14:46]>> Crazy.

[00:14:47]>> I can go and like, you know, click around and uh I could probably ask it to show me a picture of these low spreads as well. And this is just, you know, this was just done like in in in in 10 minutes. I didn't even look at it. It was just a single prompt for me. And then the way that I like to engage uh with Codex, I would like to um you know show you I can I do everything over voice. So hey, I'm in San Francisco. I'm really into bread. Uh I want like a map of all the loafers uh the loaves that I can buy and um just prestructure it with a little explanation of where I can buy it and like the price of the loaf. And then I don't even need to type anymore and it's just going to go and do that.

[00:15:27]Um, and then I can do go and and and then and and then just like do other stuff on my computer. So this is not a thing where you know I'm even you know actively spending attention on it.

[00:15:37]>> So you're saying anybody can do that if they care about some data they can b and they have access to it. They can basically make a website, analyze it, visualize it, share with their friends, family, co-workers.

[00:15:49]>> Uh and and we used to talk about writing code. It seems like such an anacronism now. Now you're you're kind of speaking code, right? It's it's using code under the hood, but it's more of a tool. Yeah.

[00:15:59]For the agent. You don't even need to know about it. Um it uh it's just, you know, it's editing and creating spreadsheets and documents and slide decks for you. It's creating websites for you. You don't need to understand what's um you know, how how actually performs its work under the hood. It's just using code as a tool. And this is why there's this beautiful generality to it.

[00:16:20]>> Yeah. So, so in the old days, if you had the skills, this might have taken you a weekend to to create something.

[00:16:27]>> It would have taken you quite a bit of time to find all that information. Yeah.

[00:16:30]>> Right. And and now it's just an instant.

[00:16:32]And then if you don't quite like it, you can just tell it to change it. Right.

[00:16:36]>> That's right. If I'm like, oh, actually, you know, I wanted to, you know, do the same but for coffee.

[00:16:42]>> Um, hey, do the same analysis for coffee. Um, not just bread.

[00:16:49]That's it. That's all I need to do. It will go and work for eight minutes and it'll end up with likely will end up with um you know a website just about coffee in San Francisco.

[00:16:58]>> And you're looking for you're optimizing for price and quality, right? It's actually more complex than just give me cheap bread.

[00:17:04]>> Yes. Yes. So you you can express your own preferences. Um you know Codex will just go and like take that into account and um just really work to satisfy >> you know your your your prompt, >> right? So it it gets the data, it visualizes the data, gets you to an insight about the world, then you can make a decision to suit your needs, which is kind of like the universal loop we're all in in our lives. How do I actually decide make a better decision?

[00:17:29]Yes.

[00:17:29]>> To reach my goals.

[00:17:30]>> Yes. And this can be very simple. This can be like as complicated as you know like the same process was used for the latest you know fundra at OpenAI. So it's it's it's just codeex is just incapable of like the smallest things and then the most complicated things as well. Yeah. Yeah. Amazing. So, so are you now using it like Alexander or are you kind of coordinating >> all the time? You can see here in my uh sidebar is this is before we started to talk. I fire uh off like uh you know little tasks like all the time like maybe over a hundred different tasks like per day that I just like hand off to Codex and it just goes and like does it for me. This goes from like organizing my desktop files uh you know managing the compute fleet um helping me understand uh how the on call rotation is doing and like how the engineers are doing helping me understand you know the launch schedule that's coming up uh and you know flagging anything that's like at risk that maybe I need to look into um I use it as a little chief of staff where uh and I I run this every day. So for example, you can do a little automation and like go through Gmail uh notion calendar and uh flag um and give me give me a summary of my day and flag anything uh that is at risk. And then I can say, okay, like do this daily. Um, do this daily at 9:00 am. And then, you know, I can just go ahead and schedule that and then it will run it and then I will find it in my inbox um, you know, at 9:00 a.m. every day.

[00:19:07]>> So, it's like you've got your own chief of staff. Just, hey, don't forget to do these things. These It's It's even helping you prioritize.

[00:19:15]>> And that's right. It's helping me prioritize. It's helping me spend my attention where it matters the most. Uh it's and it's helping me also by you know I don't have to go and like do the small things um you know that would otherwise just take me time and maybe I would never get to.

[00:19:31]>> What were the things that annoyed you like the most tedious things in the pre-codex time that you spent hours every week on? Um the thing that annoyed me most was like I would not do certain things because I felt like oh I don't have the time to do it myself but I need to go and annoy someone you know to ask for like know this information. Um but you know I felt like it was you know maybe not important enough to go and you know put this on someone's desk and be like no hey can you can you help me with this question? Um so now I feel like you know I can just you know get the information that I want have the little reports that I want. um I can build all sorts of personal software that you know I didn't have the time to do. So you know it's just like a really uh it gives me joy because um it it is doing all the little tedious you know things that were like a very manual on my computer before and now I'm like okay I'm just in the CEX app and um it's doing all these things for me and I can just focus on you know the things that I feel I actually want to think about.

[00:20:30]>> Yeah. Uh, so I know people often talk about it as like this would have taken weeks, now it takes seconds, but it actually feels like this never would have happened.

[00:20:39]>> It would have taken infinity time because we never would have gotten to it. And now and now it's actually >> there's a lot of that. There's a lot of that.

[00:20:44]>> It also sounds like are you enjoying your job more because these things are getting done that used to annoy you?

[00:20:51]>> Yes. Yes, I am. there's like so much that I don't um you know like for example every day I'm like oh you know what's the latest news um and there's like it has very personal instructions for the kind of news that I care about and so one that is an example like I would have never received such a report about the news um and it gives me great joy like to just like read this in the morning and you know feel a little bit more informed a lot of things on uh when it comes to like investigating for example we have like a bug report uh from a user I see it like on Twitter and I was like oh you know that that is a bummer because I would have usually put this as like this is like a low priority thing this is probably affecting like three users on the planet this doesn't feel like a thing like now I'm like oh let's just put this in codeex and it will get it done and you know I can feel like you know just much better because I don't feel like you know things like falling through the cracks >> anymore.

[00:21:54]>> So, it's reducing the co cognitive load for me.

[00:21:57]>> I It makes me think that this is actually a tool to address burnout and over information overwhelm. Like people we we're all kind of surrounded by tools that are supposed to help us, but in the end we're kind of trapped within our tools and and this is a thing that's kind of liberating us as a tool. um and and maybe decre like so many people, teachers, doctors, so many people feel burned out and overwhelmed because they're trapped in those tools, manual data entry, that sort of thing. Do do you feel like >> it almost feels like our relationship to tools is changing. This is a fundamental shift in how we relate to tools.

[00:22:30]>> Yeah. To me, that's the promise. The promise is you have a trustworthy partner >> almost that you can go through and you know we'll do a lot of work on your behalf to a level of trustworthiness where you're like hey like if I delegate something to you um you know you'll you'll get it done if you don't you know get it done to the level of satisfaction you know you'll flag it to me but I know that it will be done well I know that it will done be done you know maybe to a level of quality that's superior to like know if I were doing it myself. I can also, you know, trust this partner to um shield me from like a lot of the noise and, you know, to flag important things, you know, when in a timely fashion. And so I don't have to go and, you know, consume like information in like, you know, seven different apps and be bombarded by things that, you know, I don't actually care about. Um, what I imagine the future to be like is, you know, I don't even have to read my email. Um, maybe I just have, you know, I have like my little personal agent that is just like reading my inbox for me, flagging me like whenever there's something really important, asking me for input and then just, you know, doing the work.

[00:23:45]>> So, you don't have to search for needles in a dozen different haststacks. The needles are kind of arranged in a newsletter.

[00:23:52]>> I can I can say like, here are my goals for today. help me out manage everything else and trust that that will happen.

[00:23:59]>> That's amazing. You must have seen over the past few months like the what you can trust this tool with is changing like it's like you mentioned time horizon like the time it takes to do a complex test that's that's changing.

[00:24:11]>> Yes. So we we actually launched a new thing there that is like more of an advanced feature uh which is uh slashgo.

[00:24:20]So in in Codex you have slash commands uh which allow you to uh for example enter like different modes. Slash goal allows you to enter a mode where you give a long-term goal to Codex and it goes and is really relentless at it. And so for example, you can give it the goal of solving a very hard mathematical problem and it will go at it for hours, maybe days, maybe weeks until um it has decided that it has satisfied the goal.

[00:24:50]Um we're seeing it being used to do like amazing things on like improving the performance of programs, rewriting entire programs from one language to another. Um it's used in science problems as well. I've seen like really cool breakthroughs on uh mathematics and and and physics breakthroughs with it.

[00:25:08]Um and uh this is fascinating to me because maybe like a couple of months ago we were oh you know we were this is exciting. It's working for 10 minutes.

[00:25:18]Now we're talking about >> you know agents working for like weeks >> on like the hardest tasks.

[00:25:24]>> Right. Right. Sometimes I think genius is just the ability to think about the same thing longer.

[00:25:29]>> Yes. Maybe if you know if you have infinite time like probably like uh you know a sufficiently smart human would be able to come up with everything that we have done you know in humanity so far.

[00:25:38]So that feels right to me.

[00:25:39]>> So so goal you can use that in the in the CLI. Can you use that in this codeex app that you're showing us right now?

[00:25:45]>> Yeah it's it's it's um it's coming. It's not yet uh launched but it will be available.

[00:25:50]>> Okay. We'll we'll let the community know when when goal is available on all surfaces. Okay. Awesome. So, so it's running for days and it's just going going going in the background and it comes and then it comes back and reports to you, hey, I'm I'm done or I'm stuck on something.

[00:26:04]>> Yeah, >> that's right. Um, and where the future is headed, I think, is where it just it's not going to stop. Um, where you're just going to have an agent that runs like 247 and, you know, does useful things for you and gets steered along the way.

[00:26:19]>> Yeah.

[00:26:19]>> Right now, it's like very turnbased.

[00:26:21]>> Yep. uh where you know you go and you have a specific task and you're like oh you know it's like I want you to go and solve that task like go find all the loaves you know in San Francisco and like tell me the price that's great solving tasks is super important um goal oriented task solving is also a big unlock um in terms of like the value that you it can create for you but the next step I think is just it just runs continuously >> and it does useful things whether you instruct it or not and then maybe at some point it's like I've done all the things that I think are useful. I'm just going to sleep for a little bit until I talk to you.

[00:26:55]>> Um, but it's not like um it's not like it's just working when you have like a very clear instruction.

[00:27:02]>> Is there any trick to specifying the goal you want it? Like how do you make goals succeed?

[00:27:09]>> Yeah. So, one of the one of the shifts with agents um compared to you know maybe if if you haven't experienced you know if you haven't installed the codex app and you know it's like the first time you're you're using it just just explore a little bit you can engage um in a very chatty way about you know what it what it's capable of like it it it is it understands its own capabilities so you can just you know ask it like what it can do but a good trick is to be precise on helping it evaluate its own success. And so if you're able to describe what good looks like and you know what salt looks like and what you want to see you know at the end of the completion of the task then you know codex will be able to sort of like understand whether it's doing well and whether it has completed the task or not.

[00:28:00]>> Can you put a number on it? Can you be precise?

[00:28:03]>> Yeah. So for example, you know, you can say like, hey, I want uh I'm presenting my work. Um I would like a slide deck, you know, I would like it to have like 10 slides. The first two slides I wanted to contain this kind of information.

[00:28:18]>> The next six slides I wanted to, you know, just really go into the meat of the problem and do a technical breakdown and then I want, you know, two slides at the end with open questions and Q&A. And if you are clear about it and you know you're you're specific about what you want as the output, then it's much more likely to succeed.

[00:28:36]>> So very similar to what you might do with an assistant or a lieutenant.

[00:28:39]>> That's right.

[00:28:40]>> Yeah, that's right.

[00:28:40]>> Okay. So we take questions from the community >> and I'm going to read some of those. So we've got one here um from and excuse me if I mispronounce this Galing Karia uh principal engineer in financial services um is asking what would make non-coders move away from chat GPT to codeex.

[00:29:01]I think it's going to be an evolution and it's really I I don't expect that um you know everyone will want to move to critics but it is a nice complimentary to chat um what I recommend to use it for is to do things for you um so anything that you know where it's like manipulation of files on your computer. Um, running things on like automation and doing things in the background every couple of hours, you know, you can do that codeex.

[00:29:41]Um, you know, and and Chad GPT for me is still my go-to for like, you know, quick answers.

[00:29:47]>> Yeah. I remember the old days for me, I would be copying and pasting code from chat GPT into a file or a terminal. But you don't have to do that anymore. The copy and paste era is over. Yes. Codeex can just like work on the code and the data, right? Yeah. And if you have um you know if you have like a a file on your computer, you have pictures on your computer, you know, just tell Codex like hey you know use this file, read it and you can just do it.

[00:30:10]>> You don't need to go and like know click on it manually.

[00:30:13]>> Yeah.

[00:30:14]>> Okay. We've got one from Anastasia.

[00:30:18]What do you think becomes the biggest bottleneck in enterprise adoption? Uh and then the choices are model capability, human trust or organizational process design.

[00:30:32]I don't think it's a capability thing.

[00:30:34]Uh the capability is there. That's not what that's holding back adoption uh in enterprise. I think it's primarily um trust and uh for trust it's it's about is it safe and is it secure. Um, you know, if if you have an agent, just run around uh in in your in your company and like, you know, maybe delete sensitive files or, you know, upload information and, you know, send an email that contains um, you know, and exfiltrates like information that it shouldn't.

[00:31:04]>> People aren't going to use it.

[00:31:05]>> That's that's that's terrible. Um, so we've been thinking a lot about that. By default, agents run inside of sandbox with like tight controls. you know exactly that they are able to only access um you know a specific portion of the file system and so you know like you can constrain it to a folder you can disable network access we're giving a lot of enterprise controls I think this is very key to continue to expand from there there's also investments that we've done it's on our alignment blog post uh we call this feature auto review where we have created an agent that is capable of reviewing the actions of the primary agent and the way it works is that it the primary agent CEX is incentivized to do the work for you. At times it might take an action that is you know maybe a little bit risky and so you have another agent that sort of reviews every action that it's taking and flags and stops it if it's like a high risk and you know I expect like more innovations like this will be necessary >> like a referee or an umpire that's against the rules.

[00:32:03]>> Yes. Yes. And be like hey stop right there you know this is like too risky of an action I would like you to take like a less risky action. So just to help the non-engineers in the audience, a sandbox is that just like a folder in like in a file system like do you say don't go outside of this folder just do things within this?

[00:32:19]>> Yeah. So that's that's right. It's um it's similar to how um you can think about you know if you're in a company like oh you know maybe you only get access to information for one part of the organization that you're in. Um and you know people have like different silos. This is like the equivalent of a sign sandbox for an agent. you can sort of like restrict what it can have access to and like the set of actions that you can take.

[00:32:41]>> Yeah. And and so a rule might be you can read this data that it gave you access to, but you can't write back to it. You can't delete it. You can't like change it. The data is preserved, but you can kind of extract and analyze it.

[00:32:52]>> Yeah. So for example, you could say, oh, you can only have read access to something.

[00:32:56]>> Yeah. Cool. Um, okay. We've got one, uh, Jason Duca, owner and principal consultant, crossing point IT solutions.

[00:33:03]What are non-developers doing that makes codec codecs work well for them? And what habits separate the people who are getting results from the ones who feel stuck?

[00:33:13]>> Um, I think what I've seen is engaging with creativity and talking to others who have um seen success agents.

[00:33:27]I I I am surprised myself at like some of the use cases. Um >> so being part of a community >> being part of a community is great.

[00:33:37]>> Um the second thing is trying to engage with it um with precise instructions instead of uh you know being being quite vague about what you want. And so putting some effort into like just really being like hey you know this is this is um this is precisely the outcome that I would like we talked about this a little bit but I would just really consider you know agents to be you know >> someone new who just joined the company has like no context about you doesn't know what you like doesn't know about your preferences like how would you engage with them you would probably be uh you know quite >> uh um you would probably give them quite a bit of information about like where to find things, how to think about, you know, relevant documents to read.

[00:34:25]>> Mh.

[00:34:26]>> And so often people like forget about that.

[00:34:28]>> Uh and then the the third thing is connect a lot of like different sources.

[00:34:33]So we have plugins for like almost everything now. We have more than 100 different plugins. You can collect you know your calendar docs, you know, notion, you can connect like your favorite tools. The more you give it access to like tools and information from your world, like the more useful it will be. Mhm.

[00:34:49]>> So the more and better context gives you more and better results and some of that context is in your head.

[00:34:55]>> And so it's your goals. It's things you've experienced that are not in the docs and you and and you you've got to be a good boss and actually share some of these things.

[00:35:02]>> That's actually interesting. So I I have taken now the habit of like writing everything down. Um, so a lot of like my own thoughts, my goals, it's like it's it's actually in files on my computer and uh, Codex has access to it as well so that it can like align itself a little bit more. So obviously, yes, if it doesn't have access to, you know, it doesn't have access to your brain, so it can't read your mind, so you have to verbalize that some.

[00:35:27]>> Yeah. Neat. Okay, we've got one more from the community. Daniel Green, uh, Kansas City AI collective. What are your favorite realworld uses of computer use so far, especially outside traditional coding?

[00:35:40]>> Um, I I use it for shopping um where it will will go and um order stuff for me. So, uh I use it what I do is on my personal computer, I have Codex do meal planning and then um and then it will go and actually like order the ingredients. So, that's the thing that I use it for. Um, another thing that I've seen uh people use uh it for is it's all it's it's almost impossible to find like which settings you know which settings pop up to open up you know like be it on Windows or or Mac OS to to do something on your computer and so just being like hey Codex um I'm trying to uh change this on my computer just can you show me how to get there and it will go and you know it will actually like you know take you around and like you know click on the right things. Um and then you know you learn something about your computer.

[00:36:34]Um sometimes as well um there's like something that you want just kind of want to adjust um you know some slides or integrate like an image and then you know computer uses that for that as well. For technical people it is very very useful for QA um where you know Codex can then like open the app and click around and test that it actually functions.

[00:36:56]>> Yeah. Amazing. Um okay let's see what do we got. Sai Tri Hamsini N at a IML engineer at adapt view. What are the biggest mistakes people make when prompting codeex? I want to know the answer to that one. Oh um as you progress in delegating to codeex it can be at some point I see it becoming very tempting to just delegate everything including your own understanding >> and not using codeex enough to elevate your own understanding of the problem.

[00:37:40]>> Um, if you just if you're just like purely like, oh, I'm going to delegate everything. At some point, you know, you might realize like you don't actually understand, you know, what's happening anymore. Um, and then, you know, you've kind of like, you know, you're not as grounded and you're, you know, maybe going to lose a little bit um of your your productivity and like that's not good. So, spending more time to also gather information, have it like explain things to you. It can draw diagrams like images. V2 is like amazing at this um because it's very good at rendering text too. And so I see people use it all the time to go and read um you know read like um our plans for launches or you know like things from marketing or like parts of the codebase and then create images to like know just explain a concept to you if you're trying to learn something. So I would say like the mistake I'm seeing is just like too much delegation and you know not enough like using it to just helping you understand stuff.

[00:38:37]>> I totally agree. Whoever does the work does the learning basically. So I I've built myself a little skill a tutor skill that flips the tables and it asks me to give answers about what it's tried to teach me already so that I can force myself to do that learn of active recall. Yeah. Super cool. Let's see what else do we have. um why does codeex matter beyond software and and where do you see these use cases going next? So we're really talking about the future.

[00:39:02]We've talked a lot about how things are changing up till now. I think we're the future for a lot of people, but there's a future for us too and for everybody together.

[00:39:10]>> Yes. So fundamentally what we're building is a very general and very powerful agent that if connected to the right sources of information and you know with increasing access to all the information in the world and given the ability to take actions upon the world we'll be able to do pretty much anything that um you know you allow it to do. And so the promise there is just incredible value creation. Um a lot of things that you know would have taken like prohibitive amounts of of time you would never have gotten to or like just like too hard to accomplish uh for humans will become possible. Everyone we will distribute this um the the intention is to distribute this as like a broadly broadly as possible and to give you know the entire world access to these capabilities uh and these agents. And so it's just really about, you know, just lifting, you know, the the the amount of things that people can just, you know, even dream about accomplishing.

[00:40:13]>> Yeah. Yeah. I've spoken everybody from prize-winning physicists to industry project managers, they've come to me and said, "I had a stack of good ideas and I've been able to start work on them and they never would have been realized in the world before. It feels like we're entering a golden era."

[00:40:29]>> It feels like it.

[00:40:31]>> Yeah. Amazing. Okay. Well, TBO, this was a great session. Thank you.

[00:40:35]>> Thank you.

[00:40:36]>> Um, I and I just want to follow up. Uh, Tibo mentioned that having a community will help you learn how to use codec better. And I just want to make the point that OpenAI forum is a community where people are using codecs and we can teach each other and support each other in learning that better. And I hope everybody in the session today will continue to engage because this is a place where skills are modeled and transferred so that you can bring this tool into your life in new ways. Um the question we tried to answer today is is why um why should people who don't code care about codeex? I hope we answered that. I'd also like to make the point that although the word code is in codeex, >> codeex actually means book which is something we all are used to holding, right? So so so codex is a more general word than just code. Um but I but I hope this conversation um made that answer concrete. Uh >> codeex is uh useful for developers. Um, also useful for people, anybody, knowledge workers, people working with information go, you know, searching for the information they need, the needles in the hay stacks, uh, struggling to make sense of it, data analysis, visualize it with kind of instant visualization like Tibo showed showed us um, and prioritization and performing complex tasks in in the background of their lives. So, we think that Codex is uh bringing those powers, those superpowers to lots more people, and we're really excited about what that's going to mean for our economy and society and their businesses. Um, as we close, I'd like to ask everybody listening to think of one workflow in their life or in their organization where Code Codex might help. Um, and that could be a research brief or a plan or an onboarding process or reporting or a decision memo. And just give it a try because uh you've heard a lot from us today, but the the the clearest message is just going to be your experience with the tool, right? You don't have to trust anybody else about it. You can just experience that tool. It's there available to you. So, thank you again everyone for being part of OpenAI forum.

[00:42:34]Uh we appreciate your time and your curiosity and the thought everybody brought here and especially the the folks who got their questions in. Those were great questions. So thank you. Um we have an upcoming event. It's Tuesday, May 26th. How AI could detect and prevent the next pandemic with the Gates Foundation. Super excited about that. So thank you everyone and thank you to