OpenAI's Alexander Embiricos on Codex: API revenue doubled in a week, GPT-5.5 computer use, and the 'Lord Bottleneck' automation story

Speaker 1: Remember this whole thing? You didn't

Speaker 2: I just wanted a filter to be able to buy only shop for things on Amazon.

Speaker 1: Well now instead of a filter, you can listen to an hour long podcast about every possible SKU, and then you can make the most determined decision available. Well, we have Alex from OpenAI. He's a member of product staff here to talk about Codex and GPT 5.5. I believe he's in the waiting room, so let's bring him to the TBPN Ultra. Alex, how are you doing?

Speaker 4: Hey. Doing great. How are you guys?

Speaker 1: We're doing fantastic.

Speaker 2: Close to the camera. You make me wanna you make me wanna scoot up. Some people come

Speaker 1: on and it's standoffish.

Speaker 2: Yeah. It's just like a little bit of a There

Speaker 4: we go.

Speaker 2: No. You're locked in. You're locked in.

Speaker 1: Okay. So give us some updates on Codex. I mean, we've seen some crazy numbers. Things seem like they're going well. But what I'm most interested in is applications, usages, just as I'm talking to friends and folks that are outside of the most intimate tech community, what are the magical experiences? What are the prompts, the usages that are seeing adoption? How should people even be thinking about Codex and ChatGPT these days?

Speaker 4: Totally. I mean, it is such an exciting time. I feel like last year was the explosion of agents for coding.

Speaker 1: Yeah.

Speaker 4: You know, now we bring it to to everyone else. Codex now is becoming this, like, amazing tool for just, general work or, honestly, anything you can do on your computer.

Speaker 1: Yeah.

Speaker 4: So the model launch, was it was that only last week? I think it

Speaker 1: was last week. Think it's

Speaker 2: really fast. Really feels like a month ago.

Speaker 4: Yeah. Yeah. Honestly, I lose track of time but like, you know, API revenue for that model is growing two x faster than any prior release. Like Codex revenue actually doubled in the last week.

Speaker 1: I see.

Speaker 4: So it's just like the the growth is insane.

Speaker 1: Yeah.

Speaker 4: And what's cool is that, obviously, the growth in usage and the way people are using Codex for coding is is just like getting way better. But I think if this is kind of what you're asking for, there have been we're we're making Codex great for everyone. That's all the way from like what the agent can do and also just like how simple it is to use. Yeah. And so now we have, you know, Codex being used by like salespeople, marketing people, finance people, data science people, whatever you have whatever you're thinking. Yeah. It's like 85 of the company uses codecs and we're seeing this happen outside as well. So, yeah. I'm happy to jump into some fun use cases if you want.

Speaker 2: I wanted to ask about like how you feel the the perception around computer use has been. Mhmm. Like, it felt like there were so many things that that got released at one time that that kind of was quite magical and you see some posts popping up here and there, but do you feel like that's getting the attention that it that it deserves now?

Speaker 4: Yeah. I mean, like, so broadly what we have is we have this amazing amazing agent that can write code and if wanted we to do more work than just coding work, it needs to be able to do anything you can do on your computer. And so computer use is this big step because it's like, hey, well, now it basically can do anything you can do on your computer. But the magical part of computer use and that the team really cooked here is that it can use applications in the background. Mhmm. So, you could be doing something in one app and the agent can be doing something in another app and that's important because

Speaker 1: it means

Speaker 4: you can actually delegate Yeah. Which means you can give the agent hard tasks that take us some time to do. Yeah, So, mean, reception around that was actually awesome. I I thought we got people noticed sort of just how hard and magical that is more so than I even expected. So, yeah, we've been really happy with that.

Speaker 1: Yeah. I I I've noticed that codex puppeteering a mouse cursor has gotten good. It feels like it crossed some sort of like touring test where when I see a video of codex moving a mouse around, it doesn't read to me like, oh, that's a jittery that like the AI is using It just looks like, oh, somebody just recorded themselves moving the mouse. Is there a benchmark? Is that is that something that just comes from the new model or work that's being done on Codex? Like, what is going on with the the progress

Speaker 2: to the Sky team Yeah. Of acquisition. Yeah.

Speaker 4: So there's kind like three things going on there. Like, first of all, GPT 5.5 is our best model ever for like general work or knowledge work. Yeah. And so it's really good at using computers. The next thing is, well, what harness do you give the model or, like, what information is the model seeing and, like, what tools does it have to use the computer? A lot of early versions of giving a model access to a computer were just giving it screenshots of the computer. But there's a lot of secret sauce in our implementation where the model actually gets text representations of what's on screen from like frameworks like accessibility. Sure. And so the model is like much more efficient when it has access to all this information. Oh, that's interesting. And then the last bit that I I honestly think I well, don't know if I would say it's underappreciated because I feel like people really appreciate it, but it's the level of craft that was put in to how it feels when the agent is using the computer. So, for instance, you could totally just have the agent, like, click around on your screen and just have that be invisible and and, you know, you wouldn't even it's just like the computer is updating as clicks happen. But the team put a ton of care into, like, exactly the animation that this, like, mouse cursor takes as it goes between the different click positions. And, yeah, it was it was really fun to talk about that and jam on that with them. Like, we actually made some interesting trade offs. Like, having this animation actually slows down how quickly the agent can work by just a tiny bit.

Speaker 3: Mhmm.

Speaker 4: But it means that it's so much easier as a human for me to understand the system and therefore to trust the system.

Speaker 1: Yeah. That's interesting. That that sort of goes back to I remember the the initial ChatGPT app launch on iOS had haptic feedback as the token streamed in. And so it felt like it was typing. And that could have easily just been generate the full response, give you a little waiting wheel and then boom, it loads like a web page. Like when I go to, you know, The New York Times, like the whole page loads. It doesn't stream in. But that streaming made it feel more like a conversation. Those little like UX queues help like create this more like interactive motion back and forth. That's very interesting.

Speaker 2: Gordy? Someone's trying Codex Yes. For the first time. What are you've got one minute to explain the three things that they should do to get the most out of it. What do you tell them?

Speaker 4: Okay. Cool. Is this person an engineer or are they like a knowledge worker?

Speaker 1: Let's say you're not an engineer because I feel like

Speaker 2: Well, no. Let's do let's do both.

Speaker 1: Sure. Yeah. Let's start with engineer.

Speaker 4: Okay. Start with engineer. If someone's using Codex for the first time, I just say download Codex, attach connect it to your project, whatever you've been working on Yeah. And then ask it a question about your code base, like a hard architecture question, and it's just gonna give you an amazing answer. Yeah. Then the next thing you do is you say, cool. Like, give it a bug that you've been trying to track down and ask it to solve the bug. And, like, I hear this all the time, like, on Twitter, etcetera, like, people will be like, hey, I was trying to solve this bug myself, couldn't do it. I gave it to, like, all the other coding agents, they couldn't do it, but Codex could do it. Yeah. And then maybe as a final thing to, like, experience a little bit of the magic that has shipped in the last week or two, if you're working on, like, something that's, like, has a web view, like, maybe a website or something like that, ask Codex to make a change and then, like, watch it just, like, iterate on that change by opening the in app browser and sort of observing its outputs, clicking around, and then, like, naturally just, fix and improve the change that it made. It's, like, it's super magical because you just realize when you see it, like, work in that loop, you just realize how powerful it is now.

Speaker 1: Okay. Speaking of loops, explain Ralph loop, explain slash goal.

Speaker 4: Okay. So basically, we've had this interesting thing for a long time where people this is like months ago. People would tell us that they knew that Codex was super powerful, but it felt annoying to work with because they had to like constantly tell the agent, you can do this. Keep going. Right? So you have this like brilliant model, but you have to like encourage it constantly. We then shipped a feature called queuing, which you can use to give the agent a message that it will, like, receive whenever it thinks it's done. And people use that to say, do this, then do that, then do that. But actually, a lot of people started using that to just say, keep going. So, like, they would just queue, like, 10 messages, like, keep Keep going, keep going.

Speaker 2: Yeah. Exactly.

Speaker 4: And now at this point, we have these amazing models that if you know how to use in the right way, you can have them do hours of work or even days of work, just like independently, autonomously. However, the average person doesn't necessarily know how to do that. You have to do like all this contrived harness setup, and so with Goal, which is a feature that we shipped into the command line interface and will come to the app soon, we wanted to make that super easy. So now you could basically describe to like, hey, like, I want you to keep going until you achieve this goal, and then Codex will just take care of working for however long you need until you can do that. And so internally, this is something that people have been really excited for for a long time when we do a lot of like long running work. So for example, when you're like babysitting like a training run for a model, you don't want to just like tell Codex to do something and then have it return in five minutes. You actually want it to pay attention like all night. Mhmm. So that's what that's what goal is. Do you want yeah. Go ahead.

Speaker 1: I was wondering about the like, what is for for the non engineer use case, the thing that gets you to move from ChatGPT to Codex, like the first the first the first, like, more complicated query, more complicated project that you would recommend someone say, oh, yeah. Like ChatGPT can do a decent amount of research, but for this, you should spin up Codex and and and start working with that as your primary workflow. Like, what is the entry point? What is the the appetizer into the Codex workflow for someone who's nontechnical, not writing code, wouldn't mind if code was written behind the scenes, but really just is interacting with typical knowledge work suites. So email, spreadsheets, Word docs, generating graphics, research. There's a lot of stuff and ChatGPT satisfies a ton of that. And so when are they jumping over to Codex?

Speaker 4: Yeah. So I kind of think about the way that you can do general work with Codex. There's maybe three categories of work that's like, just very easy tasks. Mhmm. And that's not to dismiss those. Actually, most of my usage of Codex is like easy tasks. Mhmm. And those are the first things you should try. Then there's like hard tasks that are like really cool demos, but actually, if you start to do a hard demo, you might you might just, like, waste a lot of time. Yeah. And then there's automating. So you everyone should kinda go through this progression. It's, like, easy, hard, automated. Yeah. And if you think about it, it's kinda like a human teammate. Right? Like, you hire someone onto your team, you know, you don't usually well, maybe at OpenAI we do it. You don't just say, like, hey, just, like, figure out what you wanna do. Maybe what you do is you say, hey, like, why don't you do this, like, small starter task or something, and then from there, give them a harder and harder task, and then eventually you say, okay, just like go figure it out and work automatically. Mhmm. So, okay. Easy tasks. The thing I always recommend people do is, like, wherever your company works at OpenAI, that's Slack, maybe for you it's Teams or it's email, connect Codex to that tool and then just ask it like, hey, like, am I missing any urgent like, do I need to reply to anything urgently? Just, like, draft some answers for me. Or maybe, like, I get tagged on these really long threads all the time. It's like, hey, just summarize this thread. What am I supposed what am I being asked, and what should I what what should I answer? You know, just ask these, like, these questions. Or maybe someone mentions something you don't know what it is, and you're just like, hey, search all my company information and tell me what is x y z, you know, the thing. So those are really basic, easy queries that Codex is amazing at, and I have found that even though those may not sound that hard, people just get hooked on doing that all the time, and then from there, once you're hooked, then you kind of become fluent in using this tool to answer any even small question. Use it for harder things. So, for example, one of the use cases that I do a lot that tends to, like, really get other people excited when they see it is I you know, we we don't like meetings. I try to keep meeting attendees to a minimum here. So, often, instead of adding people to a meeting, I'll post in a channel and say, if anyone wants to talk live about this, just reply. And This is something I used to do, you know, before, but now what I do is I I tell Codex, post in this channel, see who wants to join, and add anyone who wants to join to the meeting. Sure. Super basic thing. Right? Yeah. But the fact that you can ask an agent to do a thing and then monitor that post. Mhmm. And then, like, just keep it up to date Sure. Is actually really powerful. And then when people reply to the the post and then the the agent replies back saying, like, cool, you're in the meeting. It's at this time. Here's the link. Yeah. That always mind blows people. Yeah. And then you start people start trolling, you know, people like add spam to the meeting. Oh, yeah. The agent knows not to.

Speaker 2: Yeah. That's just like functionally what what if you're lucky enough to have an EA, you get something like that where it's like, hey, I need to get this thing done. I'm gonna kinda set it. I'm gonna I I wanna set off the process, but then someone else monitoring it, kinda backfilling it.

Speaker 4: Yeah. I think any any work that requires, like, a lot of work, like, a lot of time from you but is not hard, those are good starter tasks. Same for, like, managing tasks, you know, tracking issues coming out of a bug bash. For, like, pre launch, we do this thing where I'll have Codex monitor a bunch of channels I'm in with internal and external users, and I'll just say, hey, if anything comes up, put this in in linear, which is where we track our bugs Mhmm. And, like, make sure it's deduplicated, and then let people know that we're tracking the bug. So, you know, very very simple task, but actually incredibly useful time wise. Last thing I'll say on this Automation.

Speaker 11: We have executive

Speaker 4: assistant kind of like plug in internally that I would love to ship externally. Mhmm. We should get on that. And it's really taken off internally. Like, people basically have this thing that is, like, keeping track of all their information and just, like, helping them organize their day and, like, stay on top of things. It's it's pretty cool.

Speaker 1: So then, yeah, talk to me about automation because I I I understand from integrations and computer use how a single prompt could go into Slack or go into your email or pull some documents together, write code, do whatever it needs to on the Internet, APIs. I understand all that. But then what is the workflow to get Codex to do something every morning at 7AM to prepare me a digest or take the same actions to really get rid of that rote work that I've demoed and I've seen the results. It's working. And now I just don't want to think about it ever again forever.

Speaker 4: Yeah. I mean, it's it's honestly super simple. You just tell Codex, hey, do this every morning. There

Speaker 1: we go.

Speaker 4: Right. Well, like, in my case Yeah.

Speaker 2: Mean, that's kind of an interesting thing with these kind of products that have effectively unlimited potential. It's like, where's the line between like a feature and just a prompt? Yeah. And when when should as a company, when should you just launch something and make a big moment out of it? Or when should you just let people just understand the full potential of it? And then when they want something, they can just ask it. Right? They don't need to, like, request a feature. Yeah. It's just, like, literally just request it Totally. With the agent.

Speaker 4: It's really interesting as we design the product. Like, I like to think that we need to keep sort of all the rules and, like, sort of the heavy UI and stuff around the agent to a minimum because it kind of constraints Mhmm. How much better the product can get when we have a smarter model. So the more things we can kind like, let the model decide, you know, it's like it's AI soon, AGI, the better. And then sometimes you just need UI so that people can learn that this thing exists. But even that, the model can suggest things. So I'll I'll give you the the most powerful, I think, example of a use case I heard recently was this person on the growth team needs to figure out, like, what experiments to run, and then they need to write code to run the experiment, and then they need to analyze the experiment. And it turns out they were use they started using Codecs individually for each separate thing. So they would have it, like, run a bunch of analyses, they would sort of interrogate the data. They would just talk to Codex about the data. And then they would pick an experiment and then they would, like, write ask Codex to write the code. Then they would run the experiment and then they would ask Codex what the results of the experiment were and then they would, like, produce a deck. Right? So all steps they were doing individually, and they they didn't start by saying, I'm gonna automate this entire thing because that's like hard and scary. Right? They just started with using codecs, like, accelerate themselves, individual productivity for each task. Then they started basically connecting all these things together into a giant skill and then one day, they just said, hey, why don't you do this every morning? And they gave it a name, it's called Lord Bottleneck because it's like solving the Mhmm. Bottlenecks of, like, friction for new users. And basically, this is the thing the team does now, every morning, Lord Bottleneck evaluates past experiments, looks at data, looks at new things, proposes some experiments, and offers to the team like, hey, like, let's run these experiments. The team, like, picks, like, let's do this one or that one. Then Lord bottleneck is like, okay, cool. Here's some code or whatever config that needs to be done, runs the experiment, and at the end, they, like, they go and do the same loop the next day and analyze the results. And so that's actually, like, really serious value in automation that's, like I forgot the numbers, but it's produced, like, you know, significant company value just automatically through codecs.

Speaker 1: How are you thinking about game development? It feels like all of these tools are super useful and could really accelerate game development.

Speaker 2: Especially with images too.

Speaker 1: Totally. I think there was a time when like everyone had an idea for a meme but not everyone knew Photoshop. Now everyone can go and use images in ChatGPT to sort of make the exact meme that they want very quickly, very easily And there's this takeoff in like everyone can create the thing. Now everyone can vibe code a Python script pretty easily. And it feels like the next thing is like going and shipping something. And we're starting seeing these like simulators and these little one off like web games. It's still a little bit like prosumer, I would say. But is this something that you think just is like a natural outgrowth or something where you might actually spend some cycles thinking about that as a particular, like, workflow?

Speaker 4: Yeah. I mean, I think I'm really just excited about how easy it's becoming for everyone to bring their ideas to life. ImageGen you you bring up ImageGen. That was actually a really big deal Yeah. In a way that I still think people don't really realize. Like, having an agent able to just generate images as it's working is massive. Yeah. I mean, ImageGen two in itself is actually really popular and is great even outside of Codex. Like, I think let me see here. Yeah. Like, I think we are like, people are making, like, 50% more images with ImageGen two just, like, a few weeks after launch than they were before. So it's, like, really taking off and it was already popular. But then, with ImageGen, we see all these cool use cases where people will have the agent let's say, we'll take a development and a non development use case. For development, they'll have it, like, build a bunch of sprites. Sprites is like a term for, like, the image that you use in your game.

Speaker 1: Yep.

Speaker 4: So for so, like, one cool example of this is, like, in the GPT 5.4 blog, we have some example prompts to try, and one of them is, like I forget. It's it's, like, build, like, an art roller coaster simulator or arcade simulator or something, and you can just copy paste that prompt into Codex now and it it's it's just like even better because Codex will like create all the assets for the game using ImageGen and then like build a game using it and then play the game for you to test it in the in app browser. So, yes, this is happening very naturally and I suspect it would be quite interesting to lean into it to make it even better. Yeah. A cool, like, non non coding use cases, we're seeing people use ImageGen to create slides or, like, slide templates

Speaker 3: Yeah.

Speaker 4: Or assets go on slides and then produce the slide deck, which again, you can iterate on just inside the Codex app. Interesting. Yeah. Image images are massive for us.

Speaker 1: Yeah. Yeah. That makes tons sense. Jordan, anything else?

Speaker 2: No. This is great. Yeah. Appreciate the update. I'm sure you'll there will be more news this time next week.

Speaker 3: Yeah. We're we're

Speaker 4: thinking about like how many Thursdays can we ship in a row, you know, like

Speaker 3: Yeah. Well,

Speaker 2: the more the more you ship, the easier it is to ship. So it's a it's a good feedback loop.

Speaker 1: That's fantastic. Thank you

Speaker 2: so much for taking Great to see

Speaker 4: Thank you.

Speaker 1: We'll talk to soon. A great weekend. Goodbye. Should we show show everyone the game that we made? Should we show everyone TBPN simulator? Do we have that pulled up? Can we play that? Would would now be a good time? It's a late might

Speaker 2: not have enough time.

Speaker 1: We we can have them pull it up while we while we read through

Speaker 2: the Let's pull up this packaging of the neo robot over at one x.

Speaker 1: The Neo robot over at one x.

Speaker 2: Have you seen this?

Speaker 1: Oh, yes. They got

Speaker 2: a suitcase

Speaker 1: got a for suitcase is remarkable. Let's pull this up.

Speaker 2: This packaging is so cool.

Speaker 1: Aesthetic. This warm tone. It's it's welcome.

Speaker 2: I gotta say it's really cool until it's 2AM and you hear a shaking in your house and you go and you turn on the lights and this little like, you know, suitcase thing shaking and then Neo pops out and I and goes Terminator mode. I wouldn't know what that what that actually is like, but I'm

Speaker 1: gonna watch Terminator this weekend and report back. Ridiculous. Wow. Amazon, Google, Microsoft, Meta collectively are spending more money than the Manhattan Project every single month. That's insane. Wow. Wild. Anyway, in the mansion section, there is a record breaking home in Los Angeles with a question. Will a Los Angeles area mega mansion become America's most expensive residential property? $400,000,000, they're asking. 70,000 square feet. The the guesthouse is 30,000 square feet. It's absolutely insane. 39 bedrooms, eight acres, three pools. It has an x-ray machine. This is in the mansion section of the Wall Street Journal. Fascinating because Tyler was unimpressed. And I was wondering like what could possibly be better than a $400,000,000 mega mansion in Los Angeles but you found something. What what absolutely

Speaker 5: Well, I was comparing it to, you know, some kind of palatial manner. Right? So clearly it's like Versailles. Yes. Versailles, you know, 10 x bigger. It's like it's like 700,000 square

Speaker 1: feet. 700,000

Speaker 5: indoor space. Obviously with the gardens, it's like millions of square footage.