OpenAI Just Released ChatGPT Agent, Its Most Powerful Agent Yet

Isa Fulford, Casey Chu, and Edward Sun from OpenAI's ChatGPT agent team reveal how they combined Deep Research and Operator into a single, powerful AI agent that can perform complex, multi-step tasks lasting up to an hour. By giving the model access to a virtual computer with text browsing, visual browsing, terminal access, and API integrations—all with shared state—they've created what may be the first truly embodied AI assistant. The team discusses their reinforcement learning approach, safety mitigations for real-world actions, and how small teams can build transformative AI products through close research-applied collaboration. Hosted by Sonya Huang and Lauren Reeder, Sequoia Capital

• 0:00Speaker 00

  I think this model is actually very good at multi-turn conversations. And it's very nice to continue working on a task with. I think that's one of the deficiencies of deep research. A lot of people will do multiple deep research requests in a single conversation, but it doesn't always work so well. So I think we're really happy with this model's multi-turn ability and we just want to, you know, improve even further. And then I also think like personalization and memory for agents will also be very important. And right now, every agent task is initiated by the user, but in future, it should also be doing things for you without you having to even ask in the first place.

• 0:54Speaker 01

  Today, we're exploring the evolution of AI agents with Issa Fulford, Casey Chu, and Edward Sun, the open AI team behind the new ChatGPT agents. You'll learn how they got to a huge leap forward in capability by unifying the architecture across deep research and operator, allowing for multiple tools to share state, giving users fluid transitions between visual browsing, text analysis, and code execution, all within a single environment. We'll discuss their training approach. Rather than programming specific tool usage patterns, they let the models discover the optimal strategies through reinforcement learning across thousands of virtual machines. They've created an agent that can work alongside you for hours, asking clarifying questions and accepting mid-task corrections, expanding the ways that we can interact with AI agents.

• 1:39Speaker 01

  The team shares fascinating challenges around safety, guard relays around agent activities, and why things like date picking still remain mysteriously difficult for AI systems. They've revealed how a small, focused teams are achieving breakthrough capabilities through careful data curation, suggesting that we're now entering a new phase of AI development where product insights matter just as much as compute power. Enjoy the show.

• 2:04Speaker 02

  Isa, Casey, Edward, thank you for joining us today.

• 2:07Speaker 00

  Thank you so much for having us.

• 2:09Speaker 02

  So you're the team behind the ChatGPT Agent or Agent Mode. What is it?

• 2:13Speaker 00

  Yeah, so this has been a collaboration between the Deep Research, the former Deep Research and Operator teams. And we've created a new agent in ChatGPT that's able to carry out tasks that would take humans a long time. And we gave the agent access to a virtual computer. And through that, it has a few different, two different ways to access the internet, or actually more ways, but we'll get to that. It has a text browser, which is similar to the Deep Research tool. So it's able to efficiently access information online and search through things with this very fast text browsing tool. And then it also has a virtual browser, which is similar to the operator tool. So it actually has full access to the graphical user interface.

• 2:58Speaker 00

  And it's able to click and type things into forms and scroll and drag and all this kind of, you know, all these kinds of things. So together, it's much more powerful than either of those two tools, because one's more efficient and one's like much more flexible. And then we also gave it access to a terminal. So it's able to run code and analyze files and, you know, create artifacts for you, like spreadsheets or slides. We also, through the terminal, it's able to call APIs, so either public APIs or private APIs. If you allow, if you, if you sign in, it could access your GitHub or Google Drive, SharePoint, many other things. And the cool thing about this tool is all of the tools have shared state.

• 3:42Speaker 00

  So it's similar to if you're using a computer, like all of your different applications have access to the same file system and things like that. It's the same for the tool. So the model can do quite flexible things. And yeah, we'll talk more about this later, but I think it's just a very flexible way for the model to, to do very complex tasks on behalf of, of, of users.

• 4:03Speaker 02

  Tell us a little about the origin story. How did this get started?

• 4:05Speaker 03

  Well, our team worked on Operator.

• 4:08Speaker 00

  And our team worked on Deep Research.

• 4:10Speaker 03

  And so, back in January, we released our first agent, Operator. This is a product that can do internet tasks for you, like buy things on the internet shop for you, this kind of thing. And then, two weeks later.

• 4:23Speaker 00

  We released Deep Research, which is a different model that's, or a different product that's able to extensively browse the internet, synthesize information, and it creates a long research report with citations for you.

• 4:36Speaker 03

  And we were kind of thinking through our roadmap, and we were kind of like, hey, this is kind of a match made in heaven here. So Operator is really good at visual, you know, interacting with a webpage, but it's less good at kind of the text browser, like reading long articles. Whereas Deep Research is really good at reading long articles, but it's, has a tougher time with like interactive elements or like highly visual things.

• 5:00Speaker 00

  Because the tools are different. So Deep Research has a text browser, so it's able to really efficiently read information and search and synthesize information, but it's not able to like scroll and click in the same way, or fill out forms in the same way that Operator is, because it has actually full access to the GUI browser. And as Casey was mentioning, like, Deep Research has some things that Operator doesn't have. And then similarly, one of the biggest requests for Deep Research is for the model to be able to access like paywall sources or things that you have to pay a subscription for, and Operator is able to do that.

• 5:35Speaker 03

  And also, one of our members of our team, Eric, he was running an analysis on the types of prompts that people were trying on Operator. And we realized that it was a lot of Deep Research-type tasks, like, "Research this trip for me, then book it." So it really is a natural combination.

• 5:51Speaker 04

  In what way is 1 plus 1 equals 3?

• 5:54Speaker 05

  So in Deep Research, we always wanted to figure out how to, you know, let Deep Research have access to a real browser that can, you know, loading all the real contents that previous Deep Research cannot have access to.

• 6:06Speaker 03

  It's funny that you bring up the 1 plus 1 equals 3, because not only did we combine Deep Research and Operator, but we also threw in like a bunch of other tools that basically like everything you could think of. So like, the terminal tool is there, so it can use like, it can run commands to like do calculations.

• 6:25Speaker 03

  The image gen tool is a fun one, like if it wants to, you know, spruce up its slides by making an image, it can do that.

• 6:33Speaker 00

  Can call APIs.

• 6:34Speaker 02

  Can produce PowerPoints.

• 6:36Speaker 00

  Yes, yeah, it can do a lot of different things.

• 6:38Speaker 02

  Yeah, tell us a little bit about how people are using it, knowing it's still early days.

• 6:42Speaker 00

  So I think the cool thing about it is we have some ideas of how we think people are going to use it. But I think we intentionally kept it quite open ended. I mean, it's called agent, that's so vague, partially because we are excited to see how people end up using it. So I think some of the things that we specifically trained it for were, of course, deep research type tasks. So things where you want a long report on a topic, operator type tasks where you want it to do something for you, like book something or book a flight, buy something for you. And then also tasks to make slide decks. We also, you know, spent a lot of effort on making spreadsheets and doing data analysis. But I think there are also just so many other things a model can do.

• 7:24Speaker 00

  So we're just excited to see how people use it, kind of similarly to how when we launched Deep Research, we saw a lot of people using it for code search, which was really surprising to us. We're hoping to see a lot of new use cases that we didn't even think of ourselves.

• 7:37Speaker 04

  Would you guess it'd be more consumer or kind of B two B type use cases? Or is that a false dichotomy?

• 7:42Speaker 00

  Hopefully both.

• 7:43Speaker 04

  Okay.

• 7:44Speaker 03

  Yeah, I think we're kind of aiming for like the prosumer, like someone who's willing to wait like 30 minutes for like a detailed report, but that can be in the like, yeah, in the consumer case or the, you know, at your job.

• 7:53Speaker 04

  Yeah.

• 7:54Speaker 03

  I think it could be both, good for both.

• 7:56Speaker 02

  Do any of you have favorite things you've used it for?

• 7:58Speaker 05

  For me, it's more like, you know, pulling data from our like spreadsheet or Google Docs, like a document in our like an experiment log and then make some slides to present the data or like organize the data. It's pretty useful.

• 8:13Speaker 03

  I've been doing a deep dive into ancient DNA. It's just one of my, you know, one of my interests. And there's actually a lot of exciting work going on, like these past like five years. They're like sequencing all this DNA and like discovering all these facts about like, oh, where did like this group of people come from and like, you know, historical stuff? The problem is that everything is so new that.

• 8:37Speaker 03

  There isn't like a reference like source material for like to summarize like a survey of these materials, but you know, agent can go out and pull together all these sources and synthesize it into a report that I can read or slides that I can read. And I think it's kind of made for this topic.

• 8:55Speaker 00

  Yeah, I like it for consumer use cases. Like I've used it for online shopping. I think especially because a lot of websites require using visual browser because it will have a search filter or something that it needs to go through or the model like actually needs to be able to see what the item looks like. And then also for planning events, it's been pretty useful.

• 9:16Speaker 04

  What's your favorite shopping query?

• 9:18Speaker 00

  I think I was using I was using it for clothes shopping.

• 9:24Speaker 04

  Love it. And you guys also showed us a really cool use case right before we filmed this episode. Do you want to share that one?

• 9:30Speaker 00

  Yeah, sure. So that was actually something that one of our coworkers, Tejal, shared with us. She asked the agent to estimate OpenAI's valuation and create, based on things that it found online, create a financial model with projections. So create a spreadsheet, also create a summary analysis, and then also create a slide deck presenting the results. And so hopefully the model is correct because it was it had quite an ambitious projection for us.

• 9:59Speaker 04

  It was an impressive slide deck.

• 10:01Speaker 03

  One thing I want to point out about this trajectory was that it reasoned for, I think, 28 minutes. And yeah, I think this is kind of opening up a new like paradigm where, you know, you ask the agent for a task and then you step away and it comes back with a report. And yeah, I think I think as agents become more agentic, you know, it'll be like longer and longer tasks. And this is a good example of one.

• 10:26Speaker 04

  Are these the longest running tasks you guys have have launched so far?

• 10:29Speaker 03

  I would say so. Like I just did one that was an hour long and I don't think I've ever seen that.

• 10:33Speaker 00

  I didn't know how long Codex can run for.

• 10:35Speaker 03

  Hmm, that's true.

• 10:37Speaker 04

  Is there anything special that makes, that goes into making an agent run for so long without kind of flying off the rails?

• 10:43Speaker 05

  We have some, like some tools to enable the model to be able to further extend its context lens beyond what like what's, what's the original, like the hard limit. So that the model is able to perform task by like, you know, documenting what's, what it's doing. And step by step, like a kind of like increase the time like it can do, the task, the horizon of the task it can do without the human's interruption.

• 11:12Speaker 04

  Yeah.

• 11:13Speaker 02

  It's also the, the flow to go back and forth between the model and a human also is very nice. So I can correct it as it's going, right?

• 11:19Speaker 00

  Yeah. So this model is very flexible and collaborative and that was very important to us. So it's modeled after how you would interact with someone if you asked them to do a task for you. So imagine you're asking someone on Slack to do something for you. You'd probably give them instructions. And then they'd ask you some questions and then maybe start doing the task. And then maybe in the middle of the task, they'll say, "Oh, actually, can you clarify this for me?" Or, "Can you sign into this thing for me?" Or what, "Am I allowed to do this for you?" And similarly, you might remember something that you forgot to say when you first gave them the task. You might want to interrupt them and just say, "Oh, hey, please also do this."

• 12:00Speaker 00

  Or you might want a, a status update if they're taking a long time to do it. Or you might want to redirect them if they're getting, you know, going on the wrong path. So that's kind of what we modeled it after. And I think it's very important that the user and agent are both able to initiate communication with each other. So I think what we have now is probably the most basic version of what, of what this could be. But it's better than anything we've released before in this area. Because at first, the model can, or the agent can ask you clarifying questions, similar to deep research. But it's more flexible, so it doesn't always ask you clarifying questions.

• 12:37Speaker 00

  And then you can interrupt the model. So you can say, "Oh, can you summarize what you've done so far?" Or, "Oh, I forgot to say, I actually only want like blue sneakers."

• 12:48Speaker 00

  And then if the model is going to take some kind of destructive action, or if it needs you to log into something, it will also ask the user if it's allowed to do that before doing anything.

• 12:58Speaker 03

  On this topic. We kind of built this like kind of computer interface, you guys saw it, where you can kind of watch along with what the agent is doing. And that actually persists for beyond the conversation. So like, once it's done with the task, you can actually go back and ask it follow-up questions and ask it to like fix something or do another task. And you can also take over that computer. So you can click in and then now you have access to its environment and you can like click for it or like login for it or insert your credit card information or things like that. And so, yeah, I like to think of it as like looking over your coworker's shoulders and like

• 13:42Speaker 03

  and like being able to take over if necessary.

• 13:45Speaker 04

  Thank you for enabling the micromanager in me.

• 13:50Speaker 04

  Just kidding.

• 13:51Speaker 02

  So we'd love to talk a little about how this works, to the extent that you can share.

• 13:55Speaker 05

  Yeah, so this agent is chained with the same technique as the one that we're using for small learning. So we give this agent model the all the tools we have, like implemented in the same virtual machine, like a text browser, like a GUI browser, a terminal and the imaging tool. And then the model will try to solve the task like we curated, like which are pretty hard task that the model has to complete with using these tools. And then kind of like we give, we reward the model if the model completed the task efficiently and correctly. And for example, like after this training, the model is able to, it should learn to, you know, switch between these tools like fluently.

• 14:39Speaker 05

  Like for example, if for a task of, you know, the like you ask the model to research some restaurants and maybe booking a spot for you, it will first do like a deep research style text-based browsing. And then we'll probably also use the like GUI browser to view the like images of the food and also view the availability, like which is usually like, you know, written in JavaScript that you have to use a real GUI browser. And then, or for example, if you ask it to create an artifacts, it's usually, you know, can pull sources from a website and then use them in the terminal.

• 15:16Speaker 00

  Yeah, I think the cool thing about this tool compared to tool use implementation in the past is that all of the tools have shared state. So it's all, it's like you're using, when you're using your computer and you have many different applications, you know, like if you download something, it's going to be accessible to other applications. It's very similar. So the model can open a page in the text browser, which is more efficient, but then maybe it realizes it needs the visual browser. So it can just seamlessly switch, or it could download something using the, using the browser and then in terminal it manipulates it or something like that. It can run something in terminal and then open it in the browser. It's very flexible. And so it's just giving the model a more powerful way of interacting with the internet and files in its file system and code and things like that.

• 16:01Speaker 03

  Yeah. One interesting thing to emphasize is that, like, we essentially give the model all these tools and then like lock it in the room and then like it experiments. You know, we don't really tell it when to use what tool. It kind of figures that out by itself. It's kind of almost magic.

• 16:15Speaker 04

  Is the technique, it sounds very similar to, you know, deep research. We had you on the podcast before. Should we think about this as the standard technique of how OpenAI thinks that agents will be trained going forward?

• 16:27Speaker 00

  I think we, we can take this really far.

• 16:33Speaker 00

  You know, this was, we haven't, our teams haven't been collaborating for that long. We even framed this model run as

• 16:42Speaker 00

  kind of minimum shippable de-risk. That was mostly for PR reasons internally, but this is really like the most basic version we could make together. And I think we have so, so much further we could push this with these methods. For example, the slides capability is a new capability. It's like very,

• 17:03Speaker 00

  you know, already impressive. It's a great work from Aiden, Paloma, Martin, a bunch of other people. But, you know, there's so much further we can, we can push that and improve using the same techniques. But I think we can take it further, but we probably need other things too.

• 17:19Speaker 05

  Yeah, I feel so far it's pretty magical, like the same IO algorithm just works.

• 17:23Speaker 02

  Yeah.

• 17:24Speaker 05

  Um, like O1 reasoning, like deep research with tool car. And then now like a more advanced, you know, computer use, browser use agents.

• 17:32Speaker 02

  Where does it run into the limits with this strategy and with this model specifically as well?

• 17:37Speaker 00

  I think the interesting thing with this model is that because it's taking, it's able to take actions with external side effects.

• 17:47Speaker 00

  There's a lot more risk. So for deep research, it was read only. So there's kind of a limit to what the model could do.

• 17:56Speaker 00

  In terms of like data exfiltration and other things, but with this.

• 18:01Speaker 00

  In theory, the model could successfully complete a task, but take a lot of harmful actions along the way. Like you could ask it to buy you something and it decides to buy just like a hundred different options.

• 18:14Speaker 02

  Makes on the shopping spree.

• 18:16Speaker 00

  Exactly. Or, you know, you can think of many examples like that. So I think that safety and safety training and mitigations was kind of one of the really cool parts of our process with this model. And yeah.

• 18:28Speaker 02

  Maybe Casey can talk more about it.

• 18:29Speaker 03

  I was gonna mention that kind of along the same lines, it's like this contact with the real world, that makes things difficult. You know, we have to train this on like a bunch of VMs, like it's like thousands of VMs maybe. And, uh, you know, things break and, you know, as soon as you're hitting a real website, like the website's down or like you're hitting like all these capacity limits and like load testing and this kind of thing.

• 18:56Speaker 03

  Yeah, it's really the very beginning and, you know, we're gonna iron out all these details and continue, but that's a major limitation.

• 19:02Speaker 04

  How do you think about from a safety perspective, building in the right guardrails and like how do I make sure the model's not, you know, logging into my bank account and sending it all off to a Nigerian prince?

• 19:13Speaker 03

  Yeah, that's a, that's a very good question. Yeah, this is definitely, definitely an emerging risk where, you know, the internet's a scary place, you know, there are a lot of like attackers and like scammers and this kind of thing, phishing attacks, like this goes on and on. And yeah, our model is a bit like, it can, it can reason about these things, like if you tell it to be careful, you know, we've done some safety training to make this like more robust, but sometimes it can get fooled and sometimes it, you know, is a bit too over eager to complete your task. We have a long list of mitigations and the team has worked really hard to like stack

• 19:53Speaker 03

  together a bunch of techniques to really try to make the model as safe as possible. So, you know, one example that I'll call out is that we have a monitor that looks, kind of looks over its shoulder and just sees if anything looks funny, like whether it's going on a weird website or anything like this, kind of like antivirus for your computer. It's like just kind of persistently watching and then if it looks like there's anything suspicious, then it'll stop the trajectory and stop there. You know, of course, we can't catch everything and this is a major area that we'll continue on, continue to iterate on.

• 20:33Speaker 03

  We do have like a protocol for if there are new attacks in the wild that we discover or we encounter, then we can rapidly respond and update these monitors, kind of like you would update your antivirus software, like it would like kind of pick up on these new attacks and hopefully keep you safe.

• 20:51Speaker 00

  Yeah, I think the cool thing about the safety training is that it's been a really cross org effort from the safety team, governance team, legal team, research team, engineering team, like so many others, and we have so many mitigations at every single level. We did a lot of external red teaming, internal red teaming, but yeah, as Casey mentioned, there's more. Surely when we release the model, there will be new things we uncover. So we just need to make sure we also have robust ways of detecting those and then mitigating those.

• 21:25Speaker 02

  For some of these models, there's a risk of what you can do with the models, whether it's creating biohazards or otherwise, how do you guys manage some of that?

• 21:34Speaker 03

  Yeah, it's actually, bio has been heavily on our mind. Yeah, the team has been really thoughtful about, you know, yeah, this agent, you know, we think it's very powerful, it can do research, it can like really speed up your work, but that also means that it could speed up harm. And kind of one of the top things that our team has been looking into is the risk of bio, bio risk, so like creating bio weapons, this kind of thing. And yeah, the team has been really thoughtful about how to mitigate against this and generally being very cautious. We did like many weeks of red teaming to make sure that this model could not be used for those harms and a bunch of other mitigations in place.

• 22:19Speaker 03

  Shout out Karen, who spearheaded this effort. And yeah, in general, I think we're very like aware of this and just trying to be very cautious.

• 22:31Speaker 02

  Yeah, makes sense. Tell us a little about the team that came together to build this.

• 22:34Speaker 00

  So as Casey mentioned earlier, we had deep research, research team, and then deep research applied team and operator research team, well, computer using agent research team and operator applied team. And we effectively merged everybody. We all work really closely together, both the research team and the applied team.

• 22:56Speaker 03

  And the vibes have been great.

• 22:57Speaker 00

  It's been so fun.

• 22:59Speaker 03

  Keith and I have been friends for a long time.

• 23:00Speaker 00

  Yeah.

• 23:00Speaker 03

  So like, it was a natural, like, it was great.

• 23:04Speaker 00

  How many of you are there? On deep research, for the majority of the time, three or four. Now we have some new people, which is very exciting. And then on Kua.

• 23:14Speaker 03

  On Kua, I think around six to eight, somewhere around there.

• 23:18Speaker 00

  On the research side. And then we have an amazing applied team, so like engineering, product, design,

• 23:26Speaker 00

  led by Yash Kumar, and then he has just a really correct engineering team. So it's been very fun to work really closely. I think that's one thing that's made this collaboration really special is that. The research and applied teams work so closely. And even from the beginning when we're defining what the product should be able to do, it's very much a collaboration between research and product and design. So we go backwards from the use cases we want to be able to solve to training the model and building the product. And obviously, it's able to do, it's not able to do all of those things fully yet, and it can do some things that we didn't plan. But I think it's a good framework for us when we're starting, starting a, starting a project that's very grounded in how we want people to use it in the real world.

• 24:11Speaker 04

  It's a way smaller team than I was expecting. Small teams can do amazing things. Yes, you've built a lot.

• 24:16Speaker 00

  Yeah, and we haven't been working together for very long.

• 24:19Speaker 03

  It's been a few months.

• 24:19Speaker 00

  Yeah.

• 24:20Speaker 05

  And actually, the boundary between the research team and the applied team are not, you know, like a very, like a decisive, because like, you know, during the model training, like lots of applied engineers, they are helping us training the model. And also, after we, you know, change the model, we, like some research team members are also working on the, you know, like a new setup as model in the deploy the model to the real users.

• 24:43Speaker 04

  What was the hardest part about training this agent?

• 24:46Speaker 05

  Yeah, I think one of the biggest challenge we have is how to, how to make training stable, especially given, you know, like when we train deep research, it's only using browsing and Python, it's like, it's, it's already, like, it's pretty, you know, mature tools there, like we've been using it for a while. But when training the agent model, it's, you know, it have some new tools like a computer and also the terminal like bundled in the same container, in same virtual machine as a computer. So it's actually quite hard to train because we are literally, you know, set up, you know, hundreds of thousands of virtual machines at the same time.

• 25:25Speaker 05

  And then they all like, you know, visit the internet and we, so it's, it's one of the biggest challenge, like we, we see the, actually, the training sometimes we feel, but like finally, we are very happy that we, we get this model trained.

• 25:40Speaker 04

  The VMs. Yes. All back to the engineering. A lot of pain.

• 25:46Speaker 02

  Tell us about what's next. More sources, more tools, better model. How do you think about it?

• 25:52Speaker 03

  Well, I think one thing I like about our agent framing is that

• 25:59Speaker 03

  You can ask it to do whatever you want. And you can ask it to do like every possible task you can imagine. It just might not do it well.

• 26:08Speaker 04

  Could you tell it like, go make me money on the internet?

• 26:10Speaker 03

  You can tell it that.

• 26:11Speaker 00

  It'll try.

• 26:12Speaker 03

  It'll try.

• 26:13Speaker 04

  Should we try that right after this?

• 26:15Speaker 03

  Let's do it. But yeah, I think it's really a matter of like improving the accuracy, like the performance of tasks, of like the whole distribution of tasks.

• 26:25Speaker 00

  That anyone does on a computer.

• 26:27Speaker 03

  Right.

• 26:28Speaker 00

  Which is a lot of tasks.

• 26:30Speaker 05

  And yeah, and through this like iterative deployment, we are very excited to see like what's the new capabilities that our user will, you know, find in our agent, like, you know, like the coding ability in deep research or, you know, deep research ability in operator.

• 26:44Speaker 00

  You were using the agent mode for coding?

• 26:48Speaker 05

  Yeah, I use it for coding a lot because I feel it's, you know, it's actually not, you know, like a very, like I always try to rewrite my whole code base. It just actually have some small editing. And also it actually read the, you know, original docs of different functions pretty well. Like, so it, you know, I think, I feel it hallucinate less on the function coding.

• 27:11Speaker 02

  Oh, interesting. How do you choose when to go to Codex versus when to go to Agent for that?

• 27:15Speaker 05

  For the Agent, it's, I use it more similarly to how I use 403. So it's more like an interactive experience. For the Codex, it's more like, you know, you have some, you know, well-designed problem that you want, you know, a coworker to solve, and then it will make a PR for you. But for the Agent, it's more like just give you a function or give you a suggestion.

• 27:37Speaker 00

  Cool. And it can do code search because it can access GitHub through the API connector. So code search kind of things.

• 27:43Speaker 04

  Yeah. It almost feels like, you know, the agent roadmap up until now, you've, you've built the different, almost like appendages of what it would take to have an agent. And by combining them all, this really is like the first, like fully embodied agent on a computer. I think it's very exciting.

• 28:00Speaker 00

  Yeah, I think another area that we're excited to push on is the experience of collaborating with the agent. I think this model is actually very good at multi-turn conversations. And it's very nice to, to continue working on a task with. I think that's one of the deficiencies of deep research. A lot of people will do multiple deep research requests in a single conversation, but it doesn't always work so well. So I think we're really happy with this model's multi-ternability and we just want to, you know, improve even further. And then I also think like personalization and memory for agents will also be very important. And also, um, right now, every agent task is initiated by the user,

• 28:42Speaker 00

  but in future, it should also be doing things for you without you having to even ask in the first place.

• 28:47Speaker 03

  Yeah, I'm also pretty excited about the UI and UX surrounding the agent. Because right now, um, I think, you know, obviously, we're working in a chat GPT world. Like, it's like you start a conversation and it goes. But you can imagine a lot of different modes of traction, interacting with an agent. And I'm really excited to explore different ways of, yeah, different ways of interacting with the, with the agent.

• 29:09Speaker 04

  Do you see this as always being a kind of single omniscient super agent? Or will there be the, you know, financial analyst sub agents and the, you know, personal party planner sub agents? Like, what's your vision for how that kind of plays out?

• 29:25Speaker 00

  I think people have different opinions on this. I think in the limit, if you could just ask one thing and it can figure out what it needs to do to finish the thing that you want it to do for you, that seems like it would be easiest. Like, if you just had a really amazing chief of staff who knows how to route things correctly and basically can do anything you need, that seems like it would be, um, pretty easy.

• 29:52Speaker 03

  I think I agree with that take. And like even in some of our trajectories where, like, I don't know, you're asking about,

• 30:00Speaker 03

  I don't know, maybe like a shopping task. Like, sometimes it'll go into terminal and like do some like calculations, you know, budget. And like, I think the model should be free to use all the tools it wants. It doesn't need to be a financial analyst to like. Have the financial analyst like toolset.

• 30:15Speaker 05

  Yeah, I I feel like like when you launch the product, it's sometimes makes sense to, you know, have some GPTS like, you know, like a customized model or customized instruction to, you know, put the model into a specific role. But in general, like when training the model, there are lots of, you know, positive positive transfer between, you know, deep search cooperations.

• 30:37Speaker 05

  Also, you know, a slide generation, like all of these skills are transferable. So it makes much more sense to, you know, just to have a single agent, like, you know, as an underlying base model.

• 30:47Speaker 04

  Totally. I guess, even though, you know, people do different types of work, we're all fundamentally, we're sending emails, we're making slide decks, we're doing a lot of the same work in front of a computer. I'd love to understand some of the learnings from the reinforcement learning perspective. Like, it seems like that's the method that seems to really be working for you guys with agents. Was it like very data intensive to kind of get to this point of having an agent that's so good at, you know, such a wide variety of tasks? Or like, what were some of the learnings from an RL perspective?

• 31:16Speaker 05

  Yeah, so, you know, we actually, we create a bunch of very diverse set of tasks, like some tasks to find some, you know, like a very niche, like topic or very niche answer in the internet, or some tasks, you know, like, like just very similar to deep research, like you need to write a whole forty-length article. And also lots of, you know, tasks, like just all the tasks that we want the model to be good at.

• 31:44Speaker 05

  And, yeah, and so far, we we we think that as long as, you know, like you can, you know, grade this task, we're like, after the model give you a result, you you can judge whether the, you know, the the role or the the model's performance good or not, you can kind of like reliably train the model to be like even better on this task.

• 32:03Speaker 02

  Was there anything special you need to do to make sure it had good turn-by-turn interaction with users when doing that training? Or was it just about the type of trajectories you collected?

• 32:12Speaker 05

  Hmm, yeah, so, like, I think most, most of the time, we focus on end-to-end performance, like, you know, like, from, from the, from a well-specified prompt, how to complete the task.

• 32:26Speaker 02

  And somehow it's very good at working with users.

• 32:29Speaker 00

  To your question, the reinforcement learning is very data efficient. So that means that we're able to curate a much smaller set of very high-quality data. The scale of the data just so minuscule compared to the scale of pre-training data. So we're able to teach the model new capabilities by just curating these much smaller, high-quality data sets.

• 32:52Speaker 03

  I will say, to get the operator piece to work well, you know, before we do RL, the model has to be good enough to have a basic completion of tasks. And our team has spent a lot of time in the past, like over the past two, maybe three years, getting the model to that point where it's able to actually reason about a page and like kind of like understand the visual elements really well. So this model is built on all that as well.

• 33:22Speaker 04

  Actually, could you say a little bit more about that? Because I remember early days of OpenAI, this was always part of the world of bits stuff, and you're trying to RL the mouse paths, and it was just like way too unbounded of a problem. What's changed now for that to be kind of solvable?

• 33:38Speaker 03

  Yeah, that's great that you point out the world of bits. This project does have a very long lineage dating back to like 2017 or so. Actually, like our code name is like World of Bits 2 for the computer use part.

• 33:54Speaker 04

  That's awesome.

• 33:55Speaker 03

  And yeah, what's changed? I think essentially the scale of the training has changed. We have, I don't know the multiplier, but it must be like 100,000x or something, like in terms of compute, the amount of training data we've done, both in pre-training and RL. So yeah, I really think it's just scale and the scale catching up to our ambition, I guess.

• 34:26Speaker 04

  Wow, scale's all you need.

• 34:27Speaker 03

  I believe it.

• 34:28Speaker 00

  And some good data.

• 34:31Speaker 04

  Are there particular capabilities or functionality that you're especially excited about in agent mode?

• 34:36Speaker 05

  Yeah, so this model is actually pretty good at doing some real research like data science. And also summarize the reports or like the findings in a spreadsheet. So we have some evaluation like data science bench. We evaluate the model and it's actually outperform the human baseline. So in some sense, it's actually superhuman in some research tasks. We can rely on the model to perform some basic analysis for us.

• 35:04Speaker 00

  And this is an area that John Blackman on our team was really pushing on, like spreadsheets and data science. So shout out, John.

• 35:10Speaker 04

  Spreadsheets and data science, you are automating us out of a job over here. Elevating us out of a job.

• 35:16Speaker 00

  Enhancing.

• 35:18Speaker 03

  Another thing I'm excited about is, you know, we released Operator in January. And it was decent at clicking around, but I think we substantially improved that capability where it's like much more accurate. And just kind of getting the basic things right is what I'm actually excited about, where it can reliably fill out a form and, you know, do those kinds of things.

• 35:41Speaker 00

  Date picking?

• 35:41Speaker 03

  Date picking? Date picking still needs a bit of work, but.

• 35:45Speaker 00

  For some reason, date picking is just the hottest toss.

• 35:49Speaker 04

  It's hard for a human, too.

• 35:50Speaker 00

  Like picking a date in the.

• 35:52Speaker 04

  The calendar drop down?

• 35:53Speaker 00

  Yes.

• 35:55Speaker 04

  Okay, last question. It seems like you guys have the overall framework and structure in place for something really interesting here. What's ahead? Where do you go from here?

• 36:04Speaker 00

  I think the thing that we're really excited about is that this tool that we've given the model access to is very general. It's basically most of what you could do on a computer. And if you think about all of the tasks that a human can do on a computer, it's very extensive. And so now we kind of feel like it's a matter of us making the model good at all those tasks too and figuring out a way of training on as diverse of tasks as possible with this very general tool. So I think there's a lot of hard work ahead of us, but we're very excited about it. I think what we're also excited about pushing different forms,

• 36:45Speaker 00

  ways of interacting with the agent. I think there'll be a lot of new interaction paradigms between like users and use virtual assistants or agents. So a lot of exciting times ahead.

• 36:58Speaker 04

  I can't wait to see it. Thank you. Thanks for joining us. Congratulations on the launch.

• 37:03Speaker 03

  Thank you so much.

• 37:04Speaker 04

  Thank you for having us.