OpenAI launches GPT-6 to 1.2 billion users with on-the-fly intelligent UI generation and streaming reasoning

Speaker 3: Great to meet you.

Speaker 1: I'm glad

Speaker 3: to be back on

Speaker 1: soon.... Raising and we'll talk to you

Speaker 7: Thank you.

Speaker 3: Great story.

Speaker 1: Have a good one. Goodbye. Let me tell you about Console. Console builds AI agents that automate 70% of ITHR and finance support giving employees instant resolution for access requests and password resets. Our last guest of the show is not a mystery guest. We have Alarush from Welcome to the show. Here he This is Chad GPT, product manager, and we're very excited to be joined by him. Welcome to the show. How's your day going?

Speaker 2: Good. Good. It's good to see you guys again.

Speaker 1: Yes. Welcome back. Been... Long. What'd you launch? Take us through it.

Speaker 2: Yeah. So today, we're launching GPT six and Chat GPT for all of our 1,200,000,000 users. Yeah. So a few weeks ago, we announced Astra, which is sort of our frontier model. Yep. And then a couple weeks ago, we walked through g p d six Sol and Luna. And so today, we're actually bringing GPT six now to everybody Mhmm. In ChatGPT. And it's coming with real... Two really cool new capability improvements. One, which we're calling intelligent UI

Speaker 1: Mhmm.

Speaker 2: Which is now ChatGPT can control not just what to say, but how to show Yeah. Its answers and can kind of control its interface. And the second is that ChatTPN can now answer while thinking. So historically, when you use something like the reasoning model, you've had to wait for a really great answer. Right? It can think and do a lot of work, but then you... The trade off comes with latency.

Speaker 1: Yep.

Speaker 2: And, now, we've enabled ChatTPT to actually start answering while it's still continuing to work on the problem. Yeah. And so, that really changes... It's a subtle change, but it's... It really changes how people experience chattypty for more complex questions.

Speaker 1: Yeah. There's subtle...

Speaker 3: I can't tell you how many times I'm like creating a new chat because I've fired off

Speaker 1: like two... Two two six. It's pro to understand what the capital is.

Speaker 3: It's really it's really not that......

Speaker 1: State is

Speaker 3: or You really don't. You're doing way too much.

Speaker 1: Yeah. Exactly. It happens all the time. One thing that I've always noticed on the product development side and the UX side is the use of haptics. Like very early on in the ChatGPT app, as the tokens were streaming in, you'd hear you'd feel the little haptics. And I ran into it recently when I asked to pull temperature data over a hundred year period. And I was just trying to get an idea of like has it gotten hotter in this town more than the global average or whatever. It pulled all the data. I expected that, but it generated a chart and the chart was interactive. And as I moved my thumb along the chart, there was a haptic feedback and I was like, wow, a really delightful experience. And I'm interested in how you have puzzled with when do you want to build a widget, when do you want to actually set the application up for success to use basically like a harness, something that... A pattern that you've decided versus how much are you giving over to the model over time to say, hey. There's just a whole bunch of APIs. The haptic one is right there. You can pick from it as you please.

Speaker 2: Yeah. That's a great question. So historically, we've built these as sort of... You used the word widget, is kind of exactly right. Like, we've built these predefined cards. Yeah. And the content inside can sometimes be dynamic like the charts you show we talked about. But the response has mainly been marked out. Right? These chatbots produce text and then we find ways to sort of hijack that markdown to allow it to reference additional artifacts like the... A weather widget or a chart or something like that. Yeah. With this change, like... So that approach is somewhat limiting. Right? Because actually, it should just create the perfect interface

Speaker 1: Yep.

Speaker 2: For whatever you're trying to do. Yep. Sometimes text is amazing. You're writing something, you're looking for advice, But other times, ChatGPT should really have a lot more control for how it constructs the entire layout and interactivity of your answer. And so, with this change, we've actually trained ChatGPT to have full control over what it shows you. Yeah. And so, it can generate a UI on the fly and do it really fast. It's not something that'll take a minute or two. It's really for every chat in the same latency. Yeah. And it can figure out what kind of functionality it wants to add on top of that. So some of that will be like interactivity within the chat. Mhmm. Let's say you're trying to understand a topic. It'll actually give you an interactive diagram for you to play with to better understand that. But it can also add functionality. Right? Like, often now, if you ask for like a recipe, it'll give you a shopping list, but it'll add a copy button. And that's not something we've like built in a pre made way. The model just realizes adding a copy functionality when I give you a shopping list is the most useful thing for me to

Speaker 1: do. I imagine that there's a there's an interesting flywheel here where, obviously, giving this out to, 1,200,000,000 users that are free and maybe not fully monetized with ads just yet is expensive. But there's a lot of value that comes from a free user iterating and saying, no, you should actually have the slider this way. You should add haptic feedbacks, add that copy button. And then that goes into the rollout. The model gets better at knowing which tool to pull off the shelf at the right time. Is generally how you think about this?

Speaker 2: Yeah. So for this launch Yeah. We really wanted it such that you don't have to ask what you want.

Speaker 1: Okay.

Speaker 2: ChatGPT is gonna try and figure out Yeah. Okay. You're trying to understand the offside rule in soccer. Yeah. The most... The best way for me to explain that is not via an instruction manual Yeah. But let me just show you a diagram that you can play with Yeah. That I create on the fly. And so, you can always ask with ChatGPT what you want. Yeah. But one of the big changes is we've actually trained the model to figure out when does visuals and interactivity

Speaker 1: Yeah.

Speaker 2: Really make the answer better. And then, the other part is exactly to your point of how do we scale this to 1,200,000,000 users. You might have seen Astra can produce super cool visuals and websites.

Speaker 1: Yep.

Speaker 2: Right? But that takes a lot of compute and it's slow. Yeah. So we've actually done a lot of work. We've actually built our own sort of library in house that allows model... The model to generate UI but in a very token efficient manner and to do it in a streamable way. So it feels fast, it works cross platform, and it's very, very efficient. And so a lot of that is what went into enabling us to scale this to 1,200,000,000 users without it melting all of our GPUs or us having to charge users. Right? Yeah. So everyone will get this in a sort of unlimited fashion Yeah. To sort of build much richer answers in chat.

Speaker 3: In the twenty tens, it felt like like product management practices were getting really cemented. Product managers were oftentimes like the superstars of their organization just under under the founders today. Obviously, researchers tend to capture a little more attention than than you. Do you feel like there's like new norms being... Like, just feels like it's like a... It's 10 times as hard to be a great product manager now because you're not just working between the customers and engineering, but you're adding research and then the the the capabilities of the models are advancing. But then you also have this issue of like, you know, historically, if you had a SaaS product and like nine times out of 10 it did what the user wanted, that would be like a failing grade because people would be getting really frustrated. Now, you could have a product where nine times out of 10 it nails it be... And that's actually a good thing. Right? But but you're still... There's still gonna be some percentage of users that are frustrated. So it just feels like all the all the sort of the approach is being rewritten and the the sort of in the twenty tens or late twenty tens, you had you had, you know, books and and podcasts. It feels... It felt like product management was really like being cemented around best practices that worked in the SaaS era.

Speaker 2: Yeah. It's it's definitely changing a lot. I think the... You sort of have to start from scratch a little bit and be comfortable with nondeterminism and figure out what... Like, does product management work in a area of nondeterminism. And and even within that, like, getting very comfortable with evals and and how do we measure progress is something that's, like, a really important skill. One thing that's truly different is actually, like, we were used to nondeterminism in text. Now we're entering nondeterminism in the interface itself. Like, now it's not just the words are different. It's even the way it presents the answers are super different. And so that's also actually gonna start changing how we do product development at OpenAI because no longer can a designer build a weather card or a sports card and be like, this is how it's gonna appear when it triggers because the model can just generate that on the fly. And so we're actually now also figuring out how do we do design in the age of nondeterministic UI in ChatGPT. And, we've actually had to build our own sort of design tools and development stack such that design can now work a lot closer with research on actually steering. Here's how we're gonna teach better and better design to the model. And so, to your point, like, product management is definitely changing and I think design is also now changing a lot where design is gonna sit right next to research Mhmm. In shaping model outputs.

Speaker 1: Sure.

Speaker 3: Yeah. And that's such a such a such a big update Yeah. Versus the the prior era of like Mhmm. Being able to obsess over a certain like element of the product for for months and and then knowing knowing that consistency versus now. Yeah, just introduces like, well, how do you test this? How do you collect user feedback? All these different challenges. How

Speaker 1: do you think about interfacing with other product teams? We had Cath from Chattypuji Sites on the show last week at Dev Day. And a lot of the things that you just described as instantiating UI on the fly, my previous workflow was, hey, generate a site of this and I'll share that around. But this doesn't feel like it obsolesces sites. It's more like I was I was driving to work today and I was thinking like, wait, am I going to be able to like add OAuth to these sites and build a database? Like sites is going to grow in its own direction and you're gonna grow. And how do you think about working together with other teams? Because there's so many projects like computer use. Like operator was like a side project that like dropped and then came back with a vengeance and these things sort of morph. But what is it like interfacing with the other teams to make sure that you're not stepping on each other's toes but also setting yourself for success? That feels like a critical piece of the role these days.

Speaker 2: Yeah. It's a it's a great question. I think the first is we run super lean. So, like, there is very little sort of running into other people's toes just because we just don't have enough

Speaker 1: Sure.

Speaker 2: PMs. I think the the one thing that's really cool is with models like g p six, they're getting better. And because the model is better at front end design and building sites, we've also therefore been able to teach it to do really good UI generation in chat GBT

Speaker 1: Sure.

Speaker 2: And make it really good at using our own in house component libraries. Yeah. Right? And so one thing is we we see really nice transfer capabilities as sort of different teams make improvements and and try to stretch the model in one way that then goes and benefits all these other teams. Yeah. And this separate thing is really focusing on our users. Right? So, Cath and others are really focusing on how do we enable developers and then enterprise users to have... To be, like, maximally productive. Right? And so, the concept of a site is actually a place where everybody can collaborate and do work is amazing. And we're going to keep pushing that and making the model better at that. And then the separate thing is... Which was what I spent a lot of time on is like, hey. How do you get everything... Everyday things done? You're asking about a recipe. You're trying to learn a concept on your phone on the go and really think about, okay, how do we bring a lot of these innovations but make it accessible to our billion users. Right?

Speaker 1: Yeah.

Speaker 2: And so, generally, the ability to improve... Like, have everybody improve the model but then optimize our features and kind of the experience we're delivering for our very different user cohorts. Right? Because, like, you building sites is just so different to someone who's sort of, like, out there trying to like get to work and catch up on the latest news. Right? And you might be doing both of those, you know, throughout the course of your day. Should give you the best site building experience and we're gonna give you the best sort of everyday HATTPN experience.

Speaker 3: Last question. How how dynamic can... Do you think the the UI in a in a thread can get? Because, like, the peak peak, you know, chaos would be like you're... You effectively can generate an MMO and invite your friends into the thread and and and play it. But how how far does it even make sense?

Speaker 1: That's how I like my recipes delivered.

Speaker 3: Yeah.

Speaker 1: Yeah. I want to cook a virtual meal to how to cook a real meal.

Speaker 3: Well, I wanna turn my my recipe into a first person shooter.

Speaker 1: Ideally. Yeah.

Speaker 3: That that we can both Yes.

Speaker 1: That's that's the feature.

Speaker 2: You guys you guys should play with what we launched today. It's it's... I don't... It can't make an MMO recipe. But it can make it in... Moving the goalposts.

Speaker 1: We're moving the goalposts. Yeah.

Speaker 2: Wait wait wait and set me up for success, guys. The... It it can't it can't be that, but you want your entire conversation to look iMessage blue or look Windows '95 style, it can do that. It can create sort of a ton of interactivity. Sure. You know, it can it can really build anything and it's gonna stream really, really fast. Yeah. I think the sort of logical extension of this is that chat is going to be this operating system that helps you get whatever you want done and sort of update itself and be that perfect UI. And, yeah, it should be that if you want a recipe MMO game, we're gonna give it to you and we're gonna give it to you really, really fast as well. And Yeah. And that's sort of where we're we're heading. Yeah.

Speaker 3: Very cool. I'm gonna test it on on my my favorite personal benchmark, SurfBench. How well does it give me the surf report? It's wild. Because that's that's like the kind of that's the kind of like prompt that can actually deliver like, you know, I I do want... You want

Speaker 1: it over time surf level... Yeah.

Speaker 3: Yeah. Yeah. I might want UI that's more dynamic than just like a chart that shows me different numbers. Yeah. If it

Speaker 2: if it doesn't work, ping me on Slack.

Speaker 3: I will. And Perfect.

Speaker 2: We will make it work.

Speaker 3: Yeah. Awesome. Well, thank you so much for coming. Great update.

Speaker 1: Congrats on the launch.

Speaker 3: Of course.

Speaker 1: And we'll talk to you soon.

Speaker 2: Thank you guys.

Speaker 1: Have a Good good bye. Good man. Well, that's our show. We gotta get out of here because we have another show to do. We've been invited to guest on another show. I'm not gonna leak it because there might be some roll out strategy there. But it'll be a lot of fun. We'll promote it when it goes live. But thank you for tuning in to TBPN today.

Speaker 3: It's been an honor.

Speaker 1: It's... We're we're so much closer to seventy seven days away from Christmas. Start counting the hours.

Speaker 3: That's right.

Speaker 1: Because we are now

Speaker 3: We gotta start doing a countdown to the next.

Speaker 1: We are now less than eleven hours away from being only seventy seven days away from

Speaker 3: Countdown to the countdown.

Speaker 1: Countdown to the countdown.

Speaker 3: We gotta That's what we got, folks.

Speaker 1: Leave us five stars on our podcast on Spotify. Sign up for a newsletter at tbpn.com, and we will see you tomorrow. Goodbye. Flashback.