Transcription Q&A Sam and GDB OpenAI Dev Day 2025-10-06

gistfile1.txt

https://x.com/swyx/status/1975424358672908745


Sam:
... personalize intelligence with them everywhere. And whether that's inside of ChatGPT, or we make it easy to build agents, or we make it easy to, like, someday log in with ChatGPT and use it elsewhere, like that, that idea is something that people clearly want. I, I don't expect another move, like, from... The, the move from the operating system to the browser was, like, a sort of crazy technological platform shift. I don't think you'll see something like that from the browser to the chat bot. Uh, browsers will still be important in a lot of ways. But, but I think you may see something like AI sort of... You, your personal AI sort of seeps everywhere. 

Audience
Thank you. What's the interface of the future? 

GDB:
I mean, the thing that is starting to happen is, like, generative UI in the sense of generating, like, JavaScript and HTML. The thing I think will be very cool is the full generative UI that's, like, more Sora, right? Where it's just, like... It's even hard to picture it, right? It's like the interface fully molds to whatever thing is happening, and it's fully generated on the fly in real time. Um, I think it's gonna be super cool. Like, will there even be buttons anymore? Like, where we're going, maybe we don't need buttons. 

Audience
Or roads. Or roads. (laughs) (laughs) 
Uh, you guys have taken a, a couple different bites at the apple at app stores. So there's plug-ins, there's- 
But, but at Apples? 
Yes. (laughs) 
(laughs) 

Speaker 4
Um, and so with Access DK, like, what do you think was the limiting thing in previous iteration of the sort of app store apps concept inside of ChatGPT? And what do you think plug-ins or a- Access DK is solving for that, you know, because GPT-2 plug-ins didn't have? 

Sam:
So, each successive version has been used, like, much more than the previous one, and we've, we've tried to build things that were at the limit of the then current technology. So, you know, plug-ins could do some stuff. GPTs are, in some contexts, like, still used a huge amount, but now we can build something with smarter models and all of the other things we built around the model intelligence to offer a better integration. And I sort of think, like, this trend will keep going. You know, on the point about AI sort of seeping everywhere, we'll be able to build deeper and sort of more useful integrations as the models keep getting smarter. Uh, yeah. 

GDB:
By the way, plug-ins was one of the coolest projects at the time, and there's so much promise and potential there. It's just the models were not smart enough. Like, we struggled so much to get it to be able to use, like, three plug-ins. That's why we limited it to, to that. Um, and it's just amazing to see how much smarter the models are now. And so, you know, what, what is, what is old is starting to work. 

Audience
Um, following up on this, what's the plan of discoverability of apps within ChatGPT? Because find it's the... It took time until we built, like, a plugging app store a little bit, and then monetization also is something that was never launched ... with plug-ins, or maybe a little bit. And then for custom GPTs also, like, very limited people. So what's the plan on discoverability and monetization for apps? 

Sam:
On, on discoverability, we're gonna start with having people sort of name the one they want, and then we'll try to suggest it, and then we'll do a directory. Like with every other product, we assume our rate of learning will be very high, and there may be... You know, maybe people do wanna select their things once. Maybe they want them, like, entirely suggested. We'll have to try it as we go. On monetization, it really depends on what people build. I have a feeling this is gonna look very different than apps have on other platforms. Um, one area we know there's a lot of excitement for is sort of commerce, so we're gonna have to start with supporting the commerce protocol. But depending on what people build and how they wanna make money for their businesses, we'll try, we'll try different things. 

GDB:
For what it's worth, I do know at least one company that is successful now and got its start as a ChatGPT plug-in. And so, um, even in that, like, baby infancy stage of the models and the whole discoverability story and all those things, there was still value, uh, that people were able to, to, to squeeze from it. And so, I think there's gonna be a lot more here. 

Audience:
Um, in a recent podcast, you mentioned that this is one of the best times to be 22. What is your ideal vision for how student developers can make the most advantage of tools like Codex as the team makes them better and better? 

Sam:
I, I was watching Ramón do his demo today, and I was like, "Man, this would have taken, not that long ago, this would have taken so much work and so much time to create this thing." And you probably just never would have bothered to prototype it, because it's like, you know, kind of silly whatever. But maybe, maybe, maybe the fact that you can do it at all points is some other interesting better idea there. So I would just be building stuff super fast and launching it and trying stuff super fast. And it is... I don't, I don't think the world is, like, yet caught up with what is now possible relative to what was possible in terms of the f- the output that one 22-year-old developer can have. And probably that's gonna, like, be an earthquake for how startups work. 

Audience:
Um, yeah, on the startup side, it looks like ChatGPT apps, it's almost you're bringing along the existing products. If I'm starting a new company, should I be a context engineering platform for ChatGPT to build on top of and just bring to the data? Because ChatGPT Canvas could in theory generate a lot of these UIs. So what's the trade-off there? 

GDB:
Well, I, I tend to think that when building a company, if you're looking to what will be durable as the models get better, um, I think it is important to really think about, okay, let's say that the models just get smarter. Like, what things will they just do versus what things will they still struggle with? And I, I always think of being vertical-specific and going deep on specific domains. Like, education is a good one that I like to talk about, of like thinking about you've got a parent, you've got students, you've got teachers. Um, that's a lot of stakeholders, and that you need them to interact in very specific ways. And that doesn't necessarily just get solved by having more intelligence. It's really about the application layer and really understanding those users. And so, I guess I would think of it as doing pure technology that's just the models are kind of there, and they'll, like, probably just be there in 6 months, 12 months. Um, that's probably less durable than, than kinda the thing I just described.

Audience:
What's it, (clears throat) what's it like emotionally to be building one of the most important companies, building one of the probably most important technologies? What was that like? Especially you guys have a balanced family and, yeah, I mean, I can't imagine. 

Sam:
You go first. (laughs) 

GDB:
Well, I mean, I guess I don't think of it that way in some ways. Like, I think that the one thing I really believe is the moment you believe you've won, you've lost. Um, and I think that for us it's just always been, it's the same day to day activities, right? It's just like, I don't know, just like debugging some low-level CUDA deadlock. Like, that's still the thing you're doing, right? And of course the impact is massively different. Um, and that's, that's, that's amazing. Um, but I guess just really just trying to do the grind and focus on the work is a lot of, of what I try to do. Uh, definitely it's a lot of work. Uh, there's no question about that, and it always has been, but I guess I would just say that you just gotta tune out the noise. Like, you're never as good as they say you are, you're never as bad as they say you are, and you just gotta keep going. 

Sam:
Yeah, I, I think the subjective experience is nothing like that. Like, it's nice for you to say, I wish I felt that way sometimes. It does just feel like a sort of exhausting grind. I, like, I have a lot of gratitude for it, it's incredible to get to work on it, but, uh, there's no subjective experience like that, fortunately or unfortunately. 

Audience:
Can, can I ask a small follow-up on this? Uh, especially around the amazing slide. You guys just talked on stage that you have 800 plus million weekly users. That's a lot of people for whom, like, you guys are basically not fully in charge, but you can, like, inject thoughts into. Kind of like based world imagining, you know, there's about and algorithmic stuff. Um, so how do, how do you guys contend with that responsibility, like being in the life of population? 

Sam:
So, this, you know, this is not the, like, excitement of, "Oh, we're doing hopefully building something." I mean, this is just, like, stress of the responsibility. This we think about a lot. Um, the fact that, you know, 10% of the world is talking to kind of one brain I- is a strange thing, and there is a lot of responsibility and easy to imagine ways that can go very wrong. Uh, when we started OpenAI we thought about a lot of ways the AI could go wrong. This one, we didn't. I mean, people talk about it a little bit but not very much, and now it's this, like, real thing we have to make decisions with it. You don't get to put them off there as these, like, theoretical things you can think about later. We've clearly had some scrubs, we tried to fix them quickly. I think we've done a lot of good things. But this is an issue that I, that the world is starting to take seriously. Probably not seriously enough yet. And kind of no matter how, even if we don't have any, like, safety scripts, and of course we'll have some things we do wrong, just that the impact this has on society is no one's got a playbook for it. There's no good historical analog for this. And we have a lot of ideas, we have a lot of conviction on some things we're going to do, but we will be the first to admit no one knows exactly how this is gonna play out. 

Audience:
We'll toss it to the back of the room, Tim and Claire. 
Yeah, so what did, what did startups get wrong with, with working with OpenAI? 

GDB
Well, I j- sorry, I just wanted to quickly add something to, to the last question as well, and we can address th- that good question as well. Um, I'd say maybe the single biggest thing that I've updated on throughout OpenAI is just AI is far more surprising than I had anticipated. Right? I think all of these ways that Sam was, was mentioning that we thought about, "Hey, AI could be like this, it could go wrong this way." Um, there's echoes of that in what we see, but the way in which it plays out and the deep nuance of how these problems contact reality is something that I think no one had ans- anticipated. I think, like, ChatGPT and GPT-3, the ways in which they're super human but also completely sort of, you know, limited in other dimensions, like, sci-fi I don't think really anticipated any of it. So, I think that we have learned that iterative deployment and really trying to say this, these tools need to contact reality, we need to learn, um, but that's been a really key part of, of us being able to continue to make progress. Um, so the question was what, what do startups get wrong I- with, with building on OpenAI? Um, I think it comes back to what I said earlier of like I think that it can be tempting to say the, the thing to build is to fill a small gap on the technology side. Right? It's like, "Oh, the model's almost good at this thing. It's not quite reliable at it. Let's do with a ton of context engineering and, like, really painstakingly like squeeze this, this last performance out." It can make sense, right? But it's hard for that thing to then still be the way that you're gonna generate value in six months. And so I, I think that if you... As long as you're aware that the exponential is going to continue, as Sam said, we're going to have much smarter models in six months. Um, and as long as you're prepared for that then I think it's all good. 

Audience:
Okay, uh, so I'll ask another startup question which is giving your view of where models are going, the capabilities, what do you feel like, what space do you feel like is really under-served by the startup ecosystem right now? I saw 7,000 moving startups just today. (laughs) I think we have a lot of people thinking about that problem. What are problems that you feel like startups aren't thinking about them enough given your unique view of where things are going? 

Audience:
I'd say boring enterprise problems are actually the most exciting thing in my life. 

GDB:
(laughs) Yeah. No, it's really true. Like, I- you know, so before OpenAI I was, I was doing Stripe which was very much the, like, let's just go solve a problem that is just, like, painful, no one wants to deal with it, and that's what value is. You know, the sh- the fuck up problems I think are real. And so I think just, like, go find those 'cause they're everywhere in enterprise. Well, I would say boring enterprise problems for boring enterprise companies, very under-served. (laughs)  

Audience:
Yeah, what's something you, you've changed, uh, in your mind, maybe in the la- in the last 12 months, for example? Or something you, you are worried, worried about? 'Cause, uh, I know you've, you've answered to your mission a bit right before. So something maybe that changed in the last... Not easy.

Sam:
The overhang on what the models are currently capable of versus how most of the world is using them, I think is much bigger than I realized. You see glimpses of that with stuff like Codex but we are, we got so in the chat paradigm and other people building got so in the chat paradigm, that I, that the fact that these models can now do these much more complex, very intellectually difficult tasks over long horizons, there was more capability there than I thought and a real overhang built up. 

Audience:
Nice. Uh, uh, Dan has one. Oh, yeah, go for it. 
Speaker 12
I actually- 
Oh, go ahead. 
What have you learned about what intelligence is from your (...)? 

GDB:
Well, I think, yeah, the question of what is intelligence is a very deep, philosophical one. And I think we've seen a very practical implementation of intelligence arising from trying to solve goals. And I think it makes sense, right? It's like, if you think of what intelligence is, is like sort of accomplishing goals in a maybe complicated environment. Um, and we're starting to see the implementation of that in silicon. One thing that's been very interesting is like, I- I think that, you know, you could always really anger the classic neural net people by making biological analogies and be like, "Oh, it's just like the brain, it learns like the brain learns." Um, and I actually think that it's looking more and more like actually it is pretty much the algorithm, like at least like, you know, these multi-layer neural nets is like kind of the fundamental algorithm of intelligence. Probably your brain is trying to implement something like that. Again, apologies to any of the, uh, classic, uh, uh, you know, uh, uh, neural net people who, uh, are offended by that. But I just think that the fact that, like to me the most amazing fact is that neural nets, the idea for them, the math behind them was kind of derived in the '40s, right? And that as a, as a, uh, you know, sort of instantiation of the, uh, information processing of the brain. And despite 80 some years of people saying all the reasons this wouldn't work, it's still the thing we're doing, so there's something deeply fundamental. 

Audience:
In the '40s they were thinking about it as a symbolic processor though, so that's like slightly different, right? 

GDB:
Well, it's not, it's not 100% clear to me how much the, the, you know, is that a detail or is that fundamental? Right, just like the idea of just like multi- multiple layers of computation, like if you look at the McCulloch-Pitts five neuron paper, you can see these diagrams, like these like multilayer nets that like, I'm like, "Ah, those, those look a lot like the diagrams we draw right now." And one question you can kind of ask is, you know, would, is, yeah, is the case, is it the case that the brain by, you know, doing these more spiking patterns, uh, is that actually optimal? Or is that just like what's optimal on the available hardware? And I, and you know, silicon's a very different substrate, so maybe, uh, you know, maybe it's just, you know, just a question of what hardware you have available. 

Or just the prediction is extremely close to intelligence. 
Audience:
Who's? Yeah, go for it. I'll go after. 
Sorry, where do you think we are in the timeline for humanoid robots, particularly for consumers and at home, and what do you think needs to be true to get them at home effectively? 

Sam:
A few years, certainly not a decade. Uh, there's both hardware and software progress to make, but I'm confident we'll get there. 

Audience:
A few years. 

Sam:
Yeah. 

Audience:
So OpenAI has a forward deployed team now. Assuming the reason for that team is that enterprises struggled to deploy OpenAI and ChatGPT, what are some unexpected insights from that team so far? 

Sam:
Uh, I- I actually wouldn't say enterprises have struggled to deploy ChatGPT. I would say that some enterprises have completely figured out how to do it, not just deploy ChatGPT obviously, but the API and other services we have. And now that we sort of understand how to make companies successful there, we just want to scale it very rapidly. So we didn't have one for a while, because we weren't quite sure like what the repeatable playbook looked like. Now we think we know and we'd like to help people be very successful with it. 

Audience:
You talked about overhang. This room influences probably tens of millions of people on a daily basis, weekly basis. Clearly telling people to just start is not working. Um, what is the thing that we should be saying? And your next sentence is gonna be very influential. 

GDB:
It's tough because I actually think "just start" is like the fundamental thing. I think the thing that happens is like people have, and I've observed this so many times across different models, people like try two prompts and either they see the magic or they don't. Like if you get out a great result, you're hooked, you're gonna keep doing more and more and more and really see what these models can do. And if you just get a two, two bad rolls, then like you're just gonna move on and be like, "This is all, this, you know, none of this works." And so there is something about that that is a little bit irrational. You shouldn't just do two, you should do 10, you should do 100. Um, but I think that, that the reality is for people who aren't quite sure, that they need something that they can just try very quickly, um, see some value and then say, "Okay, I get it. Let me try to, to double down." 

Audience:
What are some experiences in-app that can help foster that aha moment similar to maybe Facebook's early days, you needed to have X number of users? Or Fortnite now uses AI and everybody basically will get wiped out by the day. So it's like what are some of those experiences you can put into ChatGPT? 

Sam:
We, we really haven't even done the basic stuff, like we throw you in there. We don't even like suggest good prompts, we don't make it easy to connect, uh, your, your data in any way. I, and, and we're just, you know, it's been like everything we can do to just kind of keep up with growth. But as we get more bandwidth and as we're able to build code faster, we're better, build product faster with our own tools, we do want to make it much easier to get started with ChatGPT and be successful. Um, but even with a prompt-... that's pretty difficult to learn how to use well. People have surprised us on the upside. A lot of people have really gotten good at using this. So, you will see from us a lot of work to make ChatGPT more mainstream accessible. Uh, but we've been really pleasantly surprised by how much people have been able to just figure it out. 

Audience:
We have time for one more question. 

What kind of pressure comes from working on potentially the most important product ever made, and what lessons have you learned from, uh, that kind of pressure? 

Sam:
Uh, it's nice of you to say. To the earlier point there, you certainly... there's no felt sense of like, "Oh, we're working on what will be the most important product ever." There's like, "Ah, it's like slow and buggy and we don't have enough GPUs, and users are complaining. They're unhappy with the host training, they don't like this and that." Like, you know, "Now we have this big competitor coming after us," and all, like... so maybe there's like someday where we get to really reflect. But right now it feels like the odds are still long and we have to work super hard. Uh, trying to think about this in real time, definitely that there is magic in a general purpose tool versus a bunch of specific tools, and the fact that if- if you can give users like a simple way to do a lot of different things, that really resonates with them. And then another is like user creativity is... that's been an amazing thing all the way through. Like, every time we put out a new model, people come up with things that we never discovered about the model, or new ways to use it, new ways to get value from it. And that has led to our culture of, you know, put these things out into the world even if we don't know exactly how people are going to use them. And the- the collective creativity of the world is, uh, is always like remarkable. Every single model update we've been surprised. 

Audience
Great. Sam and Greg, thank you for your time. Uh, everyone in the room, thank you so much for- for being here and for making this dev day incredibly special. We're really grateful that you all chose to spend your- your time and your day with us. You know, we have lots of friends who are local, but also lots of folks who travel from afar. So, thank you for being here. Um, Sam and Greg will be around for the next 15 minutes or so. So, feel free to grab a beverage, grab some light bites, and we would love to spend time with you. Thank you again. 

Sam:
Thank y'all. (clapping)