You're surrounded by journalists. Hold your position.
Overnight success.
Double leg, right?
That's misinformation.
Clearing order inbound. Let's just roll.
We are surrounded by journalists. Hold your position. Cop, get up. Trust the experts.
We are founder. Fine. Good.
I see multiple journalists on the horizon. Stand by. UAV online.
Blaze. Double blaze. Triple blaze. Double kill.
Cop team deathmatch. We are experts. Triple blade. Let's just roll. Right.
Hi. Clearing order inbound. Come get up.
We are surrounded by journalists. Hold your position. Strike one. Strike two. Activate gold. Trust.
Mark clearing order inbound. Five.
I see multiple journalists on the horizon. Stand by.
You're watching TBPN. Today is Tuesday, September 29th, 2026. We are live from OpenAI Dev Day 2026.
That's right. We're very excited. Beautiful day here. Beautiful day in San Francisco. We don't get up here nearly enough, but what a fantastic day to be here. Bunch of interesting launches. Bunch of guests from OpenAI joining us. We have Sam Altman joining us in just about 10 minutes.
Bunch of friends here. Amazing opportunity. If you're here, walk in the background and get a screenshot of it. We'll get a little picture together. I see Jeff from Stripe over here hanging out. We're live, Jeff. We're talking about you.
Great to see you. Come, come, come in the background. This is Mr. Agentic Payments over here. That's true. Captain Agentic Payments.
We're live. Jeff, you're live. We're live. We just started. Anyways, we'll get back to the show. There we go.
Anyways, having a lot of fun here. Yeah, great group. Before we get into a bunch of great guests, we should talk about the news. Anything going on?
Well, there's Ramp. Time is money. You say both. He's used corporate cards, bill pay, accounting, and a whole lot more all in one place. Always a good chance to tell people about Ramp. The headline on Dev Day, just to kind of set the table: 200 attendees. It's been four years in a row now. 20 launches today. I think the number might be higher depending on how you slice them up. And the big numbers that they put on stage: ChatGPT is now used by more than 1.2 billion people every week. Over 35 million people build with ChatGPT working Codex every week. And the big launch was personal AI agents, dots, but also ChatGPT Spaces, GPT 6.1, Soul, and a bunch of other stuff around sites, Codex, and work that we will get into. But let's take people through some of the other stories that we should cover. I'm sure we'll be covering this more tomorrow as well.
Start with Oura. John, you want to get straight there? Let's get straight into it. There's some unfortunate news. The Oura IPO seems to be taking a pause or a bit of a slowdown.
Yeah. The headline is smart ring maker Oura delays IPO due to market uncertainty. Joe Weisenthal isn't really buying it. He says the S&P 500 is up 12% this year and companies are delaying their IPOs due to market conditions. I saw some commentary on this saying that their IPO is oversubscribed by 4.4x. That sounds great. The historical average is I think like 2.5x, something like that. So why'd they pull it? That's confusing. I talked to some other folks who said no, more in more recent eras you've seen IPOs go out at 10x oversubscribed, 15x oversubscribed.
I think Figma is something like 30 or 40. It was a crazy, crazy multiple. Yeah. And also I believe this IPO was going to be pure secondary and so yeah, it's not like they need the cash. Yeah. And so it was more of like a marketing and excitement, a key moment for the company. It wasn't a necessary financing round right now. And so they can pull back and keep building.
So on one hand, amazing product. Yeah. Have managed to have relatively little competition in the ring category despite much broader competition in wearables.
It seems like it's a great business. I think investors generally are going to be a little apprehensive to own a name that in any other decade was at scale but now feels like subscale relative to some of the other action in the market. And so I'm sure they'll be back, but yeah, unfortunate news.
Are there any rumors of direct competition from the bigger companies like Apple or... I don't know, this always felt like something that Apple should have bought totally or competed with. I mean, a lot of they didn't buy Pebble, they just launched the Apple Watch, but they did very well in that category. But maybe it's just... No, but in many ways, like it feels like it would make a lot more sense than let's say like a Beats by Dre to just like fit into the whole Apple system. Yeah. Right away. They could, you know, it almost doesn't need to be even reskinned, right? It feels like it fits into the ecosystem already. So also they are facing a bunch of competition from ring makers and device makers that are more AI forward, like they have a microphone on them, but that feels like an easy... very unclear if people actually really need or want that yet. Oh true, true. You think that's why they're moving slower on it is that they... I don't think, I think a lot of people, I think if they came out and said we're adding a microphone to your ring, a lot of people would be like you're getting, you're collecting enough of my data. Don't need a microphone. Yeah, my heart rate and blood.
Yeah. Other commentary. Hero over on X says the NASDAQ hit an all-time high 7 days ago. So basically saying like look, it's probably less market uncertainty and more of a disconnect between what investors would pay and where they want to price.
They were looking to raise as much as 2.2 billion in the IPO which had drawn about four times as many orders as there were shares available or is the highest profile firm to date to delay its offering. And they list out a couple others. The stock market, meanwhile, has broadly held up over the past few weeks despite the 10-year Treasury yields reaching their highest levels since 2007. Now, it's up more today. There's a quote here from Joe Saluzzi: I wouldn't say we're in a market that's scary. So, this is a bit surprising. This isn't a market I would be scared of if I was IPOing.
I can handle the market uncertainty. This guy's just like I don't know what they're fearing. No, but again, like there's always a price that this clear and clearly they'd rather, you know, the bankers and the team would rather sit it out and just get it done at their price when they're ready.
So 4Erunner and Lifeline were going to offer 36.5 million shares. Oura is planning to sell 13.5 million shares and so those venture funds will have to wait to recycle the capital I suppose. In other news, Long Lake completed their acquisition of AMX Global Business Travel.
How much do they pay? 6.3 billion. And they have like thousands of employees now or something. Up to 30,000 employees across the company. Never have done a RIF at any of the companies. They're actually growing headcount. So again, more data that shows if you're actually applying AI well, you can actually grow your payroll.
Sure. Yeah. The other big story at least in the AI world is that AMD sets $8 billion deal for AI research company. Of course AMD bought World Labs, the neolab I suppose is the term from Fei-Fei. AMD agreed to acquire artificial intelligence research firm World Labs in an all-stock deal valued at about 8.2 billion. An acquisition that reflects chipmakers' ambitions to control more of the AI stack. AMD, a semiconductor company that recently touched a trillion dollar valuation, said the deal would bring leading AI research and development talent and help the company develop better AI hardware, software, and systems, especially for the growing field of physical AI and robotics. So, I think...
Big moment. Yeah, big moment for Fei-Fei and the whole team. Like I think it's very, I'm super excited to see what they do at AMD. I wonder how much they'll continue to focus on exactly what they've been doing to date. Massive moment, massive moment for Andreessen Horowitz. Another multi-billion dollar acquisition. Phil apology. Yeah, yeah, yeah. He filled out the apology form. Credit to him. But yeah, big moment. And yeah, it was always, Fei-Fei and the team are incredibly impressive. It always felt...
It was still, no matter how much I got explained the sort of near-term strategy, I never understood exactly how it would translate into billions of dollars of revenue today.
I mean, yeah, people would say robotic training data. You simulate a world and then the robot can interact with it and that's synthetic data for training humanoids or any robotic system. The more near-term one was you could sort of see the progression of world models to a point where you could create a video game on the fly and have all of the generation, all of the world creation happen on the fly with just a new technology. You would see, you know, companies potentially license that technology like people pay Unreal Engine to, you know, host or develop their games on top of. There was something that was happening there with consistency. Google has a product as well. But yeah, there still hasn't been that like ChatGPT moment for adoption of world models really. But yeah, and you can imagine this is like a future-facing bet as in hey let's make it a much bigger play into robotics forward. Yeah. I mean Jensen... Yeah. Yeah. I mean Jensen's been doing some more stuff with world models and open source.
We have Sam Altman joining for the third time. Welcome to the show of the hour. Good to see you. Congratulations. We're going to have you throw that headset on. Welcome to the show. How is Dev Day 2026 going?
I think they are. What were you doing 20 years ago? John knows if you... No, I was like working on a little startup. Didn't Loopt launch 20 years ago basically today. That's incredible if that's accurate. Yeah, I think September of 2006 was the launch. And I was thinking about it because as we enter the moment of personal agents, like a lot, there's obviously work applications but a lot of the demos that I've seen people talk about are things that feel aligned to the original Loopt vision. It's like if my friends are in town, let's plan a barbecue. Let's order the food. Let's go to the restaurant. Let's get the reservation. And I'm wondering is this a full circle moment for you?
It's interesting. It didn't feel like that until right when you said it. Now I guess it sort of is.
Yeah. And I'm wondering how you have been using dots, how you think this will actually instantiate itself into people's lives. It's so easy to lean into the work side, but this feels like a reinvigorating moment for OpenAI as a consumer company.
Look, sadly, all I do these days is work. So, I've mostly been using it for work and I hang out with my family. That's really the only two, but I have mostly been using it for work and it is amazing for that. And I feel like it has me back and ability to like explore and do things and I love it for that. I have not yet figured out how to like go use this to have more fun in my personal life. I'm inspired to go do that.
But you're optimistic that people will ultimately do that. I feel like, you know, OpenAI has been called like the accidental consumer company and at the same time the vision seems to be like actually get AI in the hands of people and maybe let's not end up in a future where that means everyone working 24/7. Right.
100%. I think the world's always confused when we talk about AI about whether we're worried people are going to have to work too much or too little. But I think it's great to give people like more time to hang out with their kids, do their hobbies, work on creative new projects. I think most people actually want that and I really feel it with dots. Yeah. So I think people are going to get that.
Yeah. What has the last... I mean, the last time we talked was a couple months ago or maybe more than six months ago on the show. It's been a wild time, like a lot of things have gone right it seems. Right. There's some numbers that leaked today, but it seems like some things changed. How do you tell the story of the last six months or so?
Um, I mean, I think the answer is it's all just one jagged exponential that keeps going and probably the next six months will be even crazier. Yeah. The technology has been capable for a while. It's clearly gotten much more capable, but it takes a while for technology like this to diffuse through society. And I think in the last 6 months, people are using it in much more. And now with these sort of always-on agents, I think people will again go through a period where they're like, 'Wow, I can really use this in all kinds of ways.' And this is like a significant new thing in my life that's helping me.
Yeah. So, is this a moment to sort of like pay down the capability overhang right now? You know, when you were talking about how, you know, you can use this to like do the stuff that we were thinking about 20 years ago and like book a restaurant and, you know, coordinate with my friends, I was kind of like laughing in my head because I was like, yes, you can do that, but it can also like solve the greatest unsolved problems. And so the delta in capability between, you know, some of the stuff we're talking about now and what these models are so smart can do is just going to take us a while to figure all that out.
Yeah. What have you learned this year about the sequence of events? I was shocked by if you had told me that there was going to be Hugging Face and then Navier-Stokes and then after that I saw the first person said I use ChatGPT to book me a haircut. I would have not predicted that those things happen in that order. Like I maybe would have predicted that they all happen in the future. Yeah. But I would have predicted we get the haircut first. And I'm wondering if you've learned anything about not just predicting the future, but the sequence of events.
I mean, one of the obvious things is just people want their lives to be a little easier. Yeah. And actually, I think more people are interested in just their lives being a little bit less hectic, a little easier than they are in whatever Navier-Stokes can do for them or at least their direct experience. Totally. So, but I think the cool thing about this is that models that are this capable are going to be applied to make people's lives better in big and small ways, in ways that we can't even really imagine today. And you know, I think when the fact that we all talk about haircuts and booking haircuts is like a failure of imagination, but the way it gets corrected is all these people here. Sure. Like as I've had a little bit of time today to talk to what people are building here, like I am amazed at the creative spirit of what people are building for each other with this technology. And I think it's been one of the greatest surprises of the last 6 months is seeing like just how far people are pushing the developer tools to build this amazing new stuff.
Yeah. Talk about some of the new functionality being able to log in with ChatGPT and basically bring compute with you anywhere across the internet because to me that was, there was a bunch of great moments earlier but to me the moment that stood out is imagining as an independent developer that has been sort of held back by their ability to scale cool products because hey I can't just start losing $20,000 a day serving free users.
I think it's so important to build an open ecosystem here. And we want people to bring whatever they want into ChatGPT, run other apps there, but also to take their subscription and go use it elsewhere. And any harness you want, any service you want. Like this is, I have been disappointed to see other companies with AI subscriptions go in the other direction. Like this is a thing you buy and you should be able to take it with you anywhere and you should be able to go use all of the wonderful things these people are building. And I think we're going to push that much further.
How do you think about the potential transition from OpenAI's products being very pull-based, as I mean like the app is there unless I come to it with a question, with a prompt, it doesn't do anything, to a world where my dot might message me and then you get into how much is it messaging me, how much is it pulling me in, what are the retention... there's a lot of potential risks with that, right?
It, yeah, I mean there's clearly ways that people could misuse or growth hack the fact that AIs can, you know, ping you a lot. Yeah, I'm sure we'll see some companies get that wrong. We're going to try very hard to not do that. But my own experience over the last few weeks is I actually want more. Like I think we've tuned the dial a little too conservatively for now. And the fact that everything my dot sends me is something I want means I think there's something still left on the table. But I can easily see how this goes wrong on the other direction and we'll have to figure out how we're going to navigate that.
Yeah, I'll often open ChatGPT and realize there's like a very important email that I need to respond to and like I didn't, it somehow like got mixed up in... and then you're like, you kind of wish it pinged you, right? What can you say, any advice for let's say public company CEOs that want to maintain a startup's culture at scale because that's something like, you know, OpenAI is still a private company but the impression of seeing how this company works is that it's not just like, yeah, we act like a startup at scale, like it's actually like that, it's like incredibly scrappy, it's very small groups of people trying to do the impossible on impossible timelines. And something that you're doing is working because it allows this like incredible velocity. It's not just the models.
Someone said to me a few days ago that joined OpenAI somewhat recently and had been at a big kind of slowish public company before was that the key is to have the work and expectations per person go up over time. And the way that big companies go wrong is there's too many people with not enough to do and they kind of like fight each other over like these fake levels of work. And if everybody is really busy with real impact and the amount of work per person and stuff to do per person keeps increasing, then you can fight the slow drag and politics of the big company.
On the flip side of that, I feel like there's a lot of companies that are going through sort of a slopification and they're fighting slop grenades. I've wondered if the correct KPI is maybe we should all be producing the same amount of words in our emails and it's okay if AI is involved but you don't just want to take every text message and turn it into a 20-page document, right?
We fight super hard against the slop grenade and the many slop grenades. And I think just like very strong cultural norms that if you are writing something and then you are having ChatGPT expand it and then somebody else is going to have ChatGPT summarize it back to your bullet points, the right answer is to just send the bullet points.
Yeah, I've told some people just send me the prompt and I can expand it myself in my mind. Sometimes. How are you thinking about compute these days? There's so many different battles and different markets across consumer and enterprise. What is the plan for the next couple of years?
To build a lot more compute. One thing that OpenAI has that doesn't get as much attention is our chip program. Yeah. Which will start to come online in the first half of 2027 and will scale a lot in future years. Yeah. I think it gives us a huge inference advantage per whatever, for whatever like watts we can find in the world and given the growth and demand we are seeing across consumer, developer, enterprise, really everywhere. And to say nothing of like people are going to love dots and then they're gonna say, 'Well, can I have 10 dots or 20 dots or 100 dots or whatever.' We will have to make more efficient models, more efficient chips and figure out how to build a lot more infrastructure.
How do you see the user experience across the different models, the different cost and speed parameters being vended to the actual end user? Is that something you want to maintain forever? Because developers love hearing about 6.1 Soul and ultra fast mode or are there consumers, you know, I noticed dots doesn't have a model picker.
I think people are really tired of the model picker and the like model, you know what, like this model, that model, like we want to find a way to really simplify that. Like the dream is that there is a single model and it kind of makes great decisions for you and you can maybe turn a speed dial when you really need it and a cost dial when you don't. But you know, we were talking about like models that can book haircuts versus solve Navier-Stokes. I think this model should be able to figure out a lot of the trade-offs and we should be able to do a good job with routing. I think we'll always offer choice to developers.
Yeah, sometimes you want to drive a Ferrari to the grocery store.
Sometimes you do. So, I think we'll always offer that choice, but most of the time you don't. You know, most of the time it's the minivan to the grocery store and the Ferrari at the track. And the model should like be able to make that smoother.
Yeah. Do you feel like you have more or less clarity around the future than you did a decade ago? Because when again I always encourage people to look back at your blog in the sort of 2015-2016 era because there was a bunch of things that you and other people that were writing at the time just like pretty much nailed entirely. And I go through life now knowing that there's people out there that have like very accurate predictions around the future. And sometimes there's an incentive to share, sometimes there isn't. But how do you think about your own forecasting ability?
It's nice of you to say that. There's like plenty of things I was wrong about, too. Like, you know, I try to leave the wrong predictions up obviously. Yeah. But I think the important point is like you get some things right, some things wrong. I think we'll be very right about some things. I think we'll be very right that intelligence continues. I think we'll be very right about the rate of scientific discovery. I think we'll be very right about how people kind of are going to have the same things that they care about and drives in their lives 10 years from now they have from today and even with this like amazing intelligence a lot of parts of life are going to stay the same but then a lot of the details and kind of what order science goes in or whether it's, you know, haircuts or Hugging Face or Navier-Stokes first, I'm sure we'll be wrong about.
It feels like that the leadership skill or basically having the mentality of a venture investor and somebody that's made a bunch of very long-term bets makes people very well suited for today in the sense that having that openness to if I get three things out of 10 right and I get them really right then that's great, like it sort of pays for all the other bets. One of the things that I think is coolest about this new generation of people building with AI is you can try something so quickly and so cheaply that if you are a developer and you have 30 ideas, the canonical advice that was correct two years ago is pick one and put all of your effort into it because you can't build 30 things. You can now build all 30 things and you can get people to use them all and then you can work more on the ones that work. So, I think this mindset of lots of ideas, quick feedback, and be open to a lot of things not working is great for this new era of how people build.
I have one more question, then we want to get you to hit the gong. On alignment, you've been deep in the weeds. I'm sure there was a debate I saw. Is it an engineering problem? Is it a science problem? Is there a different frame of mind that we should be approaching alignment with?
I think it's like all of those things, but it's certainly incorrect to talk about this only as an engineering problem. Anyone who says we have solved the science of alignment I believe is wrong in a very dangerous way. Like we need to make more research progress, I assume we will. We have been great at making this research progress. And of course there's a lot of engineering work to do too. You know, we need to continue to figure out how to build better sandboxes and better monitoring tools. But eventually, if we are going to create models that are like much, much smarter than all of us, we have to actually solve the science of alignment.
Have we considered moving away from the term sandbox? Because I have a 5-year-old. If I put him in a sandbox, he's not going to stay there. I think we might need a new term.
My kid, unfortunately, will just stay in.
Oh, okay. Okay. So, maybe. Yeah. But, you know, different alignment. But yes. No, I think this draw. How do I do this? Just however you want. Just with like bang. There we go. There you go. With authority. Love for you to sign this as big as possible. Big signature right there on the gong. Thank you. This is going in the Museum of Business at the TBPN Ultradome. Thank you so much for coming on the show. Great to see you. Thank you. Congratulations on a great Dev Day.
We will move on to our next guests. We have a few more hopping on the show. We have Alexander Mericost jumping on in just a minute. We will see with him on all things Codex. We need a slop grenade button. Being able to... That's bad in the chat. That's great. We could sort of simulate that if we have a remote guest on, we could say, 'Hey, check out, I have an idea. I have an idea. I have an idea for you. What do you think of this?' And then you throw like a six-page. Okay. I was thinking a slop grenade would be like you go full back to Nickelodeon slime and when you push the button, it just slams Tyler with it and he's covered in slop. I think that's the one we got to go. I don't know if you... Oh, wow. Look over here. You got a crowd chasing Sam around. They want more tokens. Give us... Got to be careful with this signed gong. Yes. Oh, yeah. That's a hot commodity.
Over in Washington, a CNBC reporter was trying and failing to interview tech CEOs at the White House. Alex Karp just took the microphone out of his hand and started asking them his own questions. Interesting. He's... No, he's literally, they're doing like a man on the street interview and just, Karp took the wheel. That's great. Well, there's more stories. We'll get to those in just a few minutes, but let's bring on Alexander.
Welcome to the show. We're going to have you throw that headset on. Do you need anything to drink? This is warmer water. Okay. Well, welcome back to the show. Always great to see you. How is this Dev Day for you? I mean apart from the outage. Excellent. Okay. Yeah. Wait. Outage, a global outage going on or... Yeah. Yeah. We had a bit of an outage during the demos. Like, you know, I haven't deeply read up on what exactly happened but like because we launched 6.1 Soul and just immediately got hammered. We did that. Oh yeah, that's a lesson. That's like the Meta Connect last year. Bos accidentally set off like the wake word but every device in the whole space started going off and then crashed everything. So positive framing is a champagne problem but broadly this Dev Day I'm so excited. Like when I joined the company like a little over two years ago, like this moment is the moment I've been looking forward to the entire time. Okay, you know, like we had this chat, we invented that paradigm, it was awesome, informational, couldn't do things for you. Okay, then we had coding agents, agents that could do things. Awesome. That's the next level. But you still have to like really babysit it and figure out like what am I going to ask this agent to do? Yeah. Yeah. And so we're finally at the stage where I can go to someone and say, 'Hey, if you want to do something manually, do it with Codex.' Yep. If you want to delegate it, sure. Go use a dot. Yeah. Right. And this dot, it has your back. It figures out how to help you. It just helps you get so much more value out of this. So for me, like I've been looking forward to this for years. Yeah. Does it feel like... Yeah. It's the same dynamic of like if you had a bunch of people on your team that would just sort of flag information but then not take any action. They're like, hey, this thing, it's happening. Like, okay, why don't you do something about it? Listen, I like that. But I would actually say like Codex is actually more like a principal engineer, but who doesn't initially, unless you tell them exactly to do it, doesn't read like Linear or doesn't check Slack. And so the super smart principal engineer, but you know, you have to make sure you sort of give it the right instructions. Yeah. Yeah. Whereas dot, it's like you have that principal engineer because your dot has its own computer. It can do anything that you can do on a computer in the cloud. It can also connect to your computer. So, if you have a really awesome local Codex setup, it can actually go use that except that it has its own identity in Slack. And so, it's available wherever you talk. And you know, I don't know if you guys have seen this, but it's, or for anyone listening, the way that dots show up in Slack isn't just like a random name like Wisp. We tried that. We had agents with random names and no one knew what was going on. It was a bit confusing. But they show up with like your username. So, like I'm AE. It's like AE-dot. Yep. In Slack. And so, you have all these people and their delegates just doing stuff together. Yeah. Principal engineer level capability. Yeah. Or if you're not an engineer, just like, you know, maximum capability. Yeah. I mean, the demo that I saw just randomly through a friend was like we were looking for Sam Altman's headshot and we were able to delegate that to a dot that was able to go find like a file in a file system. That's easy, but just that little retrieval would be a couple clicks on a web browser. And now it's just not even a couple clicks. It's like searching around folders. Sometimes it's like, yeah, digging through stuff. And so yeah, it feels like we're hitting this inflection point of like leaving just the developer workflows and getting into like everyone's workflows. And you always had the prosumer of the enterprise who was like, oh yeah, they're a sales rep but they also use Codex and now it's like there is a product that's approachable for everyone. Are you... It feels like there's the expansion of territory that's happening within businesses and then there's also like different types of businesses that are able to actually get a great value started with software development firms and software companies. Are video game developers the next leg up? Like who are you seeing here at DevDay that you're like, 'Oh, I haven't actually talked to that many folks in your industry.' And maybe the progress in Astra is flowing through. I mean that's what my experience was. I was modeling stuff in Blender. I haven't touched 3D rendering for a long time. And I'm building video games and stuff and I saw that, okay, this is, we've crossed this, the spike of intelligence has spiked beyond what's relevant in this particular category. And I'm wondering where else you're seeing it or what's been the most visceral to you. So, I agree that like I think for video games like with the stuff we've been seeing people make with Astra. Yeah. Absolutely awesome. And so therefore because Dots run on Astra, we've just seen people having Dots just like autonomously building like really fun games and stuff. So that's awesome and I agree with that. But actually I think the audience that I'm the most excited for to like get even more capability is maybe like solopreneurs or like small business owners actually because you know that's where you're doing absolutely everything that your business needs and you're just like swamped with all the stuff incoming. And that's where having a dot that just like has your back and is like managing all this stuff for you. Yeah. Doing the work that maybe you could have done before with Astra, but it was just like a lot of work to figure out that that work even needed doing is helpful. So, I've heard, you know, obviously a lot of great anecdotes from engineers like people waking up in the morning to kudos from their team because their team was collaborating with a dot and it helped them out. It's like good job for having a dot that did this. Which is so fun. But, you know, we also have a lot of great anecdotes from people in completely other areas. Like, you know, running the launch, like dots going around realizing that, oh, this thing about the product changed so we need to go update the blog or like, you know, your travel plans for this conference have changed so let me just like organize things for you. So yeah. Walk me through the specialist dots. Like is that more of like a mental framing for the user to let them know that this capability is there? Is there actual like the harness matters narrative where you do want to go to that particular dot? Or is it because I can imagine as a solo entrepreneur it's just valuable to say, okay, I at least have like one arrow in the quiver for legal, one arrow in the quiver for finance, etc. But also these models are pretty general and the AGI sort of leads you to throw, you know, Codex, Astra, Ultra at every problem. So...
How do you think about coaching people through that dynamic?
Yeah, I mean, so there's two things going on here. So we've got dots and we've got specialist dots. And the reason we're leading with dots first is like just very related to our mission, right? Deliver the benefits of AGI to all humanity. And you know, when I think about product, I want that to be palpable. Like I want every person to have direct access to something that is empowering them, extending them, helping them move faster. And so those are dots, and those dots, you know, each person has one, you'll have more in the future. And it works, you know, extending you. So using your connectors, for instance, right?
So that's great. But when you think about how you scale the amount to which this capability is deployed within enterprise, there's actually a shift that we're starting to see happening that these are a part of. Right, one of those shifts is like moving from local work on someone's computer to moving work in the cloud, right? And so specialist dots have their own computer just like dots. The other shift is actually moving away from working with like an individual's identity to moving with an IT-managed, centrally governed, well-controlled identity. And we actually had an announcement going out today about how we're working with Microsoft on that as well. And so if you are a team or an organization and you have a specific type of functional work that you would like to accelerate the whole team with, that's where you deploy a specialist dot. So it could be something like procurement or accounting. Those are some of the early places we've seen a lot of success, but ultimately broadly we see that it'll be useful in many more functions and departments.
Personally, I feel like I don't know, maybe I'm just like too much of a normie, but I feel like this is the frame of mind that I want to be in. Obviously, there's like the merge where the AI becomes you, and if I want to send you a message, it comes from me, but the AI is like puppeteering me. And I actually prefer the idea of like I have an assistant and they are their own entity and they have their own phone number and email eventually, and I can share information with them and they are on my team, but I am still me. And I think that that is something that I would bet on that being relatable to a lot of people. We'll see. I don't know how.
I actually made some like pretty funny mistakes as we were building this product that we had to like pull back. Yeah. And so like, okay, I totally agree with what you said, and then I'll tell you about the mistakes.
Ask your dot, what is he writing about me from a performance report?
Well, no, that actually there's a lot of layers of defense that that's not going to make that information. But no, it's more like what we realized is that again, there's this kind of bifurcation, right? Like I want my dots to augment me, and so maybe I want to collaborate on an email and then I'm good. I just did that from within the chat app and I hit send, right? And that should just look like me, but you know, I approved that email to be sent, and I made a mistake about that, which I'll tell you about. And then on the other hand, you know, we had for a while all these agents just sort of acting as us, sending messages as us with this like 'sent by ChatGPT' thing internally, and it was like fine, but it kind of started to get noisy. And so that's when we kind of iterated our way towards no, we're going to have like a strongly separate identity, they're going to look different, and we've got these characters, we're going to have like a name that clearly is like owner-bot. Right. And that just clarified everything that was going on in communication. Totally. So it's awesome.
And I love the idea of like, look, sometimes the amount of back and forth and compute that's going to be dedicated to some negotiation, if we're trying to find even just a time to get lunch, our dots might be going back and forth for a long time, but it's not going to ever be like you opened your messages to me and it's a flood of messages because they were going back and forth. It happened below the fold. And then you can abstract that away as much as you need to. Agree. Very cool.
I'm going to set up a dot where every time I get a startup pitch deck that's just clearly AI generated, it'll generate a playable video game that I can just go through the deck in the form of like, you know, possible.
Yeah, but that's totally a thing you could do. Oh, I know I'm joking, but it's actually possible.
So here's an interesting thing about space, right, and sites, which is like a lot of the tools that we use to do work today are optimized for human authoring. Okay, so what I mean by that is like, you know, think documents, right? Like you go into a document and you can type into it, right? And the way that your document can look is like pretty simple. It's like text, right? And like yeah, various document editors have like fancy ways for you to go in and move the images around and stuff, but it all generally kind of sucks because the tool is well designed for humans. Like I think those tools are good. Even though they're a little inflexible, but now that we have Dots and we're collaborating with them, we have AI like HTML authoring is like basically free. Right. And so like I love how space incorporates like HTML visualization. So you could just be like, 'Hey, build me this artifact that I can click around or like literally put a game in the dock or whatever in the page.' And it's super interactive, and I think this is like sort of the paradigm shift that we're starting to see.
How do you think about the road to and the benefits of actual direct integrations? Because at the end of the day, like if I have some service that I want to interact with, like there are APIs, there's ways, there's computer use at this point, like everything is somewhat connected. What's the benefit of actually directly engaging with OpenAI, doing a proper connection? What are the benefits? What are the costs?
I mean, there's a few layers here. I think it starts with, yeah, now that our agents can use the same tools that we use, i.e. computers, then there's kind of this baseline, right? Like your agent, if you want it to and allow it to, it can go to anything. And I do want to emphasize that like dots have many layers of safety protections, and one of them is custom rules, right? So if you're a company and you want to limit what exactly dots can do with various tools, like there's a very easy affordance for you to go in and do that. Yeah. But okay, so that's your baseline, right? The next layer that I think we could talk about is like connectors, and mostly it's just like it makes things faster. It lets your product be a little bit more opinionated. Yeah. Just improves the experience for everyone. Yeah. The other sort of form of integration that I'm not sure if this is what you're getting at, but that I'm really excited about is plugin extensions. Yeah. So, you know, we've long had this idea that like you or there's actually a fun debate in this industry, right, which is like to what extent should UI be generated just in time for the user. And we're doing a lot of interesting work here, and there's like some cool stuff coming out soon around that. But my opinion is that for UI that I'm going to click around in, I generally want it to be somewhat predictable so I know where the button is. I can achieve mastery over this tool. But we don't want to be the only people deciding that. We want other people to be able to build their own UIs for our users. And so something super cool is that now if you're a ChatGPT user, you can install a plugin that has a plugin extension that actually renders like custom UI, custom pages just in the ChatGPT app. So if you're a developer listening to this, you have the ability to go reach your customers directly. And the meetings plugin that we just launched today is an example of this. It's like a fully arms-length plugin integration.
Yeah. Then I feel the same way about on-the-fly UI generation, custom UI. I am excited about that. At the same time, I can sit here and talk to somebody who's extremely bullish on voice interfaces and they think that like everything should be vended over conversational voice interface. And I feel like OpenAI has kept enough irons in the fire across all these different paradigms. Is there a tension there? Is that deliberate? Is there an idea that to reach everyone, to hit a 1.2, 2 billion MAU or WOW, like you don't want to actually go all in and narrow, you need to meet people where they are across multiple modalities.
Yeah. I would say so, like this is something Tibo said earlier today. It's like at some point it's about the active intelligence and the interface stops mattering. Yeah. And you know, my slight way of reframing that would be like it's like the interface should just be whatever you want. Yeah. Right. So like I'm a product manager, so I spend a lot of my day in Slack, and so I mostly talk to my dot in Slack actually. Right. But then when I'm commuting, like I was coming in today, I knew I was going to have a bunch of conversations. I was just talking to my dot like about the day and what I need to be prepared for, was like grilling me on some hard questions, you know. Holly earlier at the demo at the keynote today was talking about how like actually the anecdote was kind of me, like I saw her in this phone booth all the time, and I was going to give her some feedback. I'm like, why are you in so many meetings? Like shouldn't we be like not in meetings? And then she's like, no, I'm talking to my dot. And I realized this, and I'm like, wow, this is crazy. We've like unearthed a completely new paradigm. So going back to your thing about irons in the fire, like I think that some of these things are kind of emergent. Like I really love talking to my dot over voice. But the way that that happened is the voice team that was working on Dots did an amazing job, but it was kind of like we combined these two capabilities, and because both irons were very hot, they combined really well, you know. But if you had started from scratch, I don't think we would have gotten to such a great outcome.
Yeah. Yeah. What I've noticed that I don't think people here have this problem, but as I was thinking about, you remember this viral video that's got Jim from Charleston AI? Did you see this? She might have to refresh my memory. Basically he has a retail store in Charleston that's you can walk in and he'll teach you like how to use codecs and work. Yeah. And I've just noticed this thing where normal people outside of the tech industry don't just know that you can ask the AI, can you do this thing? Yeah. And it's like it feels like as there's 20 new launches today, and it feels like the biggest bottleneck is like effective product marketing to the sort of long tail of users on how to use these different tools. Because when somebody asks me, hey, can I do this thing with ChatGPT, I say, well, have you asked ChatGPT that? Like it'll sort of explain the general capability set and then maybe like even prompt you on how to prompt it, etc. But it feels like outside of San Francisco, that's like the biggest bottleneck to like adoption and diffusion is just people not realizing that the best teacher for the thing is not even necessarily a YouTube video about it. It's like being able to go back and forth.
Well, I mean, two things on that. First, thanks for having me on the show. You know, this is the second point though I think is like at least when I hear that, you know, I think this is actually a product problem as well. Yeah. Like if the product, I mean it's so powerful, right? These open-ended sort of chat interfaces. You can ask whatever you want, but it's hard to realize that you can try that. Like I find that very understandable. And so that's part of why I'm so excited about dots in this era of active intelligence is that well, maybe I don't have to realize that I can ask for help for finding the headshot. You know, it'll just notice that someone was urgently asking me for the headshot, and so it might just go ahead and proactively research like where is the headshot? And they come to me like, 'Hey, someone asked you for the headshot. I happen to know where the headshot is already. Do you want me to send it?' Right. And so I think we're about to achieve this like step function change in how much value people are getting from existing models. Right. Like God is built on GPT-6 Astra, and yet it's going to unlock so much value for people.
Yeah. Yeah. I know it's an exciting time. Thank you so much for coming on the show. Have a great rest of your dev day, and we will move on to our next guest. Andrew Amcino is coming on from OpenAI. We will bring him on, have him throw on this headset and introduce himself because I believe this is your first time on the show. Correct.
Welcome. Good to meet you. Thank you so much for hopping on. Give us a little tour of your day-to-day, your role, introduce yourself for everyone.
So since the original Codex app, I've led that team who runs desktop from the Codex app to now. And I've also picked up leading the space team. From the space product.
You needed some more on your plate.
Yeah. It wasn't enough, man. Yeah. That's top was solved.
Give us an overview of your experience with like the merge, putting everything together. There's been so many hard decisions, so many band-aids to rip off. I feel like what's it actually been like? What's gone well? What do you wish you could redo?
Okay, maybe introduce the idea of the merge because there is that Sam Altman blog post from a decade ago about the merge and that's a different thing.
That's a different thing. Well, there's been a series of merges.
Okay. Should we merge them all together?
We should merge the merges. We split, we merged eventually. No, I mean we had this great app with Codex and we had ChatGPT that served a billion users. And we needed to find a way to join these two worlds and bring the agentic side of Codex in this app into ChatGPT with chat work. Yep. Of course, when we launched it originally, there were still like a lot of things that didn't match up, and so we've been working on merging those. We had two great merges today that are a little bit under the radar actually. Okay. One of them is that we completely rewrote chatgpt.com in the last five weeks. Okay. To be the CEO app.
Wait, your pin tweet? Yes. It says I'm going dark for 6 weeks or something like that, right?
So now we've got the same Codex desktop experience you're used to. The ChatGPT desktop experience is now chatgpt.com.
And then the second merge that we've done is now by default ChatGPT work on the desktop.
Yes. Works on the cloud. Okay.
And still has access to your local machine. So you've got these threads everywhere. Yes. Codex cloud we launched today. But what that means is like by default you will be able to do something that needs a file system, that needs some localish compute even though it's in the cloud. You won't have that tension between like am I just in a chat thread that's sort of like short-lived versus I got to go to my computer. Now by default you have more cloud access. Yes. One of the things that we really like about all of our new products today, whether that's the dots, the Codex cloud, chat work, all of this stuff is running in the cloud and has access to your local machine if it needed it. So if you've got files on your local machine, if you're logged into all your browsers on your local machine, the whole system, we are working on this merge. Yeah. And we are next working on the toggle merges this.
Yeah. So question for you. How do you balance trusting your own instinct around product decisions because obviously you would never make a big product decision that you thought was going to lead to a worse experience with like real-time user feedback around how things were because people, you know, we were just talking with Andrew like the people build up like a habit around different flows and different ways that the product can be oriented and then they sort of build an attachment to it and so even if it needs to change it can be jarring. But you guys are sort of like drinking from the fire hose of user feedback every single day. And whether you're taking like going offline for five or six weeks or whatever you did, I'm sure you're getting like tags and comments and messages constantly being like, 'Do this, do that, do this, do that.'
Lots of tags. Yeah. I think when it comes down to it, we want to put these capabilities into people's hands. And so it becomes obvious that even if for whatever reason we can't make it, you know, we can't bring chat and work together quite yet. What we knew was that what we were getting out of Codex had to be brought to everybody. And now we feel similarly with space, with dots, all of this stuff. We try to put the capabilities into people's hands. We dogfood it extensively internally. So we've got thousands of people who rely on this stuff to do every part of their job. So we get good feedback loop there. But it's all a work in progress. The stuff is moving incredibly quickly as you know. Yeah. Lots of tags on X all the time and we love that.
Can you walk me through how there's this culture of like side project? What was it? Side quest was the buzzword. Sometimes they get killed off.
No more side quest. But sometimes they get brought back like all the computer use stuff that's going on. Huge leap forward out of nowhere. But you go back 2 years ago, I remember operator that was like, was it a side quest or was it really important work to do because it set you up for this moment? And I'm wondering if there's prehistory for any of the new products today like space or anything else that you've touched on where you've like, oh yeah, like actually we revisited some work we did previously in a product. It wasn't quite there. The model got better. We were ready to launch.
I think most of the stuff that we released today is like a take five. Yeah. Even just of things that we did internally. There were so many variations of dots that we saw internally, played around with. Some were local, some were cloud, like there were so many variations of the stuff that we used. Space has gone through a number of iterations. We had part of chat called canvas that we had up for a while. We've had a lot of internal knowledge-based products. So we've gone through the revs on that. Codex cloud is obviously a full 2.0. The original Codex was cloud. Then we got into a local agent era and now we're back putting the full local agent on cloud environments. So I think everything that you see today is one of those operator stories or that you know operator to atlas to Codex, right. Yeah. Sometimes I think that the models are changing so fast that it's hard to get the product timing on these things exactly right to the model when the model is best at it. Yeah. So sometimes we're slightly too early, slightly right, and but I think like the stuff that we have out today feels magical.
Yeah. So tell us more about space. What is the vision especially for where that product goes? How it fits into like a typical person's workflow in the modern era.
Yeah. We find ourselves bouncing around between applications all day, right? We're in this Codex desktop app, chat desktop app, all sorts of documents, we're sharing things across Slack. And so we wanted to make something that felt very agent-native and felt like a way to collaborate with other teammates, their dots as well, right? Something that you can kind of curate this set of teamwork that you're tagging each other in, you're able to use it in a way that feels like an extension of ChatGPT. And so this is really the entry of Codex and ChatGPT into a multiplayer, more collaborative era. Yeah. And we're going to do a lot more here.
How do you think about actually like the balance between how heavy your hand is on the instantiation of UI on the fly? Like I noticed this weird thing just happened randomly and just I believe default ChatGPT. I was asking like is it a particularly hot summer? And it pulled historical weather data from my local weather station and it created this chart and when I would scroll over it, it had haptic feedback where the phone would vibrate and I was like, is that someone at OpenAI like being like we should put haptics in this thing or did the model do that? And then if I love that, am I able to sort of productize that within my organization to be like the way we do charts at this company there's haptic feedback? Like how does this flow from what you get for versus what you put your finger on the scale with versus what an individual user might want as a pattern. That's kind of what this is.
I am very passionate about this subject. And I think, you know, a few months ago with Codex, we brought this visualization. Yeah. That kind of makes UI on the fly. It can either be a full game, right? In space, you can make a chess game and play it with other people if they have access to the page. Sure. That's kind of crazy. That is finally the people are asking for it, right? It's like we're split on this decision. Let's play chess for it. I think the one that you were talking about with the weather, that one is an extra challenge because you can't have your instant chat model writing a full web app every time. Sure. Right. Yeah. And so this is one that we've worked on a lot and it's, you know, the research is finally kind of there to make it feel excellent.
Yeah. Yeah. But then of course like eventually that gets flowed back to the models. The models get faster. I mean we saw the what is it, eight times faster is ultra fast. How do you think that will change workflows for enterprises? Like I mean I think there's like a little bit of a groan at the pricing but Tyler was just like time is money. I want to hit this all day long. How do you think people are responding to like the speed trade-offs? Is that like a new core skill for people in understanding how to size workloads and actually get the resources done, allocate resources, or is this something that it should be abstracted? It is quickly being abstracted.
I like to say time is money. Save both. I would say fantastic off the cuff. It's a great line. No. I think, you know, ultra fast has been one of those things that for us like you just can't go back, right? There's it's, you know, it's not exactly the same as intelligence doing something that quickly, but it sure feels like it when you, you know, especially when we do these dynamic UI things, right? You're working in a page with your dot, you tag it in and you say, 'Hey, can you drop an interactive chart of our Dow usage or Dow growth over time, right?' And it just, it's there, right? It's like you're doing like a whiteboard session, but you can do it with all the live data from your business, right? And it goes off and it does the research in your data sources. It writes the little app for you and then you can just say like, 'Hey, share this with my team, right?' Yeah. That feels so good with ultra fast.
Yeah. Yeah. Yeah, that makes sense. What else are you working on specifically to onboarding? We were just talking about like the boom of solopreneurs. Like who is the core demo that you find the earliest adopter? What industries are on the cusp of getting online? We were just talking about the 3D modeling capabilities feels like what we saw in, you know, vibecoded SaaS products coming to the video game world. But what are you seeing on the horizon in terms of companies and developers here today that might stick out to you as somebody who wasn't here a couple years ago but is now?
Oh, it's hard to say. I mean, I feel like it's just really taken off. I don't know if you've seen the chart from OpenAI internally too, which is just basically, you know, engineering was here, research is here, and every other one is like quickly following in sync. Yeah. I think there's just some of the things that you do together like pulling data. Yeah. You know, researching things in Slack, planning events together. I mean, this entire dev day is planned using coding agents, right? They don't look like coding agents anymore, but under the hood they are. So I think we're seeing really just a lot of everything.
Yeah. How many weeks will you be out of a lock-in until the next six-week lock-ins? I think you're going right back in. That's a great question. Just repin the tweet.
I think tomorrow I'm going to do another lock-in for 6 weeks and we'll see what's at the end of that.
Well, that's fantastic. I can't wait. Anything else, Jordy? No, this is great. Thank you so much for coming out. Crazy. This is fantastic around right now. It really is. Good luck out there. All right, thanks guys. We'll talk to you soon. Thank you so much. Up next, we have Peter Steinberger returning to TBPN, founder of OpenClaw. Welcome back to the show. It's been too long. Last time we talked.
Really much has happened since then either.
Other way. Flip it around. There you go. There you go. We're live. Yeah. How's life?
Have you ever been on a roller coaster that just doesn't stop? Yeah. Like I feel like at some point I see the ups and downs. I'm like, 'Yay!' And it's still exciting.
Yeah. Is the roller coaster more wild online with the chattering class or internally with what you're working on?
I think I built up a little bit of a, how do you call it? A shell. Yeah, that's good. There's always too much online and some of it is amazing. Some of it is just like, oh, you just know so little. Yeah, totally. Totally. But I cannot stay away from it. It's too addictive. Yeah. So it's definitely a new skill set like having a surviving online. Yeah. I'm sure you don't know nothing about that. Yeah, it's the same thing. No, but it's actually insane how angry people get and how emotional people get about like software and just various software products, right? And you've been on the receiving end of like again that full roller coaster, right? Where it's pretty surreal. The last time you were on the show was just like you were calling in from I think were you still calling in from Europe maybe, but then I think you almost immediately flew over here and you probably been here ever since. It took a little bit. I had some London detours. We have a cool office there from Open EI and now I'm here in this crazy AI town.
Yeah. I'd like to go back because I think there's so much of your previous work that leads into the announcements today, the moment in AI broadly, personal agents obviously, but can you just retell us the story of like where did the idea for OpenClaw come from? Like what was the initial spark, like what got you started there?
Yeah. Like I do stuff. You talk about emotions. I also have emotions. I do them on my computer. It doesn't work as fast or as well as I would, I get annoyed. But now in this day and age, it's like you don't have, you can use that annoyance and just like let's fix it. Yeah. And then the fact that I couldn't really just talk to my computer. Yeah. That annoyed me. I like this is so, this all feels so crusty. Isn't there a better way? And then I just explored those better ways. And I keep exploring those better ways. You know this, I talked about it inside like my whole thesis is that whatever workflow we think we have right now it's already outdated. We cannot catch up with the models. You just think about like how much new things you can do because now Luna is like too cheap to ma, now you have ultra fast. It's like wait, so I don't have to do 50 terminal windows because I can actually stay in focus. Oh, it completely changes things again, you know. And it's like we can't keep up building.
Yeah. What do you think as you look back on the boom of OpenClaw, what do you think the key ingredients were? Was it the connectors? Was it the connections? Was it being able to communicate to your computer from anywhere?
I think I just showed people a different way how to use AI and so many were stuck in this world. Oh, it can summarize my email. No, it can like install stuff on your computer or like buy syncs for you or like do whatever you want that you can do in your computer if you just only give it the right tools. Yeah. And I packaged it in a way that was weird and interesting and also felt a little bit different, you know, instead of like giving you a wall of text, it felt a little bit more like lobsterish, humanish, but like it was interesting. Yeah. It showed people a new way and it had some unique ideas like what about if your could actually just be proactive, you know, and yeah, even though like some of it was like token burning, but some of it was genuinely cool and surprised people. So I think this day and age it's all about ideas and imagination.
Yeah. Where you... It was such an amazing moment because it felt like the sort of hacker culture that this entire town was built on. Yeah. Like exploded again in this awesome way. Getting back into the main thing is like you know you kicked off an entirely new paradigm. You set like the whole industry on fire in many ways. Everyone started using it, seeing it, and then immediately thinking, how do we make this like commercially viable to be able to scale from hackers that are buying Mac minis into something that everyday people can use? And I feel like in some ways it's surprising that given how good the models are that it took another 6 months to actually get to this moment right now where there's more excitement than ever around personal agents. Like in some ways like you saw this the roller coaster, right? You saw this curve of excitement and then hey this maybe isn't ready for like a billion people to use and then now we're back to the point of like okay.
Isn't that classical Gartner hype cycle.
And I honestly I wish it would have been less hype. Sure. It was, it's kind of like this roller coaster where like yay and then like it gets even closer like holy you know. So I definitely we had like a massive hype cycle with like inflated things what OpenClaw would do for you that like completely beside of what it could actually do. Yeah. And then you had all these insane stories and then of course people tried it like no it doesn't quite work like this because we still, there's still limitations. Models were not as good. The software was not as good. Of course everything's still hard. Everything is still hard and people were expecting this perfect magical thing that one guy slopped in the basement.
The ultimate slop grenade.
No, it was really a slop grenade for you. Threw it over to San Francisco.
Well, you were one of the first people to come out and just be like, 'Yeah, it's all so vibecoded.' You were unabashed about it and it's like that comes with a lot of pluses and a lot of minuses, but like here is the thing and you know it's free. Like you can go and make your own decisions. You can go rewrite the whole thing yourself one line at a time if you want.
And people did many many many times.
Yeah. Yeah. Exactly. How do you think the ev, like OpenAI and ChatGPT is in an interesting position because you have ChatGPT threads, you have chat work in the cloud, Codex cloud, Codex on your desktop. There's like a few different ways to interact with this. And I'm wondering if you have a vision for like what the next year looks like with like how much people are getting into devices, their own hardware versus things like the cloud just eating more and more things. Like we went through the boom of the open web, then everything went to social networks and Instagram profiles. And I'm wondering if you think there's like a pendulum that's swinging or if you have any sort of mental model for where the early adopter or the early consumer will be in a year. I mean the current situation with like chat and work.
Yeah, of course it's a compromise. Yeah, it is. And I learned this the hard way. It is very hard to like take people from where they are and introduce them into something new. Sure. It's not annoying these people or like confusing these people. Yeah. So sometimes you have to make compromises. Is this going to be the final form? Hell no. Like there's people working on something that's way cleaner. Yeah. What I don't fully know yet is if we will be in a future that is mostly your agent or if it's going to be whole army.
I see, I'm seeing both approaches. I'm more in the first team, but I can also totally see the second team and how all these are like how your personal and your team agents interact. There's still a lot of questions that I don't think the industry has fully figured out.
Yeah, and I always compare it back to how human organizations have worked, right? So like you have a big company with thousands of people. You might task one person, hey, we have this general goal. You figure out how to exactly define it and what it should look like, but then hundreds of people might work on it, but then you don't have hundreds of people report back all back.
Kind of replicate those hierarchies, and then maybe your agent, oh, this is something really complicated. Let me spawn an agent that manages this, and then it manages this little sub-agent. Ah, there's so many ideas, and I am playing with a lot of them. You know what really excites me these days is how do we build software in the future and how do we work better with agents. So like what I just had so much fun with the last few months is moving all of my team to one place.
And everybody was freaking out. Well, what do you mean? We were so embarrassed on how everyone is talking to their agents.
Oh, interesting. And I got pushed back, and then we did it anyhow, and then everybody felt embarrassed for a day, and now everybody's learning so much. It's like because we all have a mental model on how we use agents, and it is very different because that was not something we shared. Are you sharing your code sessions? I'm sure you have some really interesting questions.
And I've actually wanted that back in the deep research era. I would see Tyler Cowen talk about how he was getting a bunch of value for economic research, and I was like, I would love for him just to be able to click share, share the prompt and the result, because I might never think to ask that question. And we've never really developed a pattern around, oh yeah, it's totally normal to just go on Twitter and share a link to a ChatGPT transcript. Occasionally you see raw responses.
They like to take credit for it.
Yeah, they like to take credit for it. It's like it's good content. You did the research, it did the research. You thought of the question.
Where are you going through, you know, the first kind of era with OpenClaw and then now this personal agent boom? Where have you seen a bunch of personal agent activity and you're actually like, I don't think that will be a paradigm? So, for example, like shopping is a good example. Like a lot of shopping with agents right now is like walking up to the front of a store and asking a person outside of the store like, hey, I think I want to get this thing. And then can you give me some information on it? And then maybe they go in and they come back. And it's this sort of strange dynamic where maybe at some points it just makes sense to go into the store and see all the information. And so there's like every single one of these, you know, big debate around restaurant reservations in the last 48 hours as well, right? And so it's like where are these areas where you've seen some activity and you're like it doesn't even make sense to really focus on that from a product standpoint.
There's so many companies that want to own the experience, and I don't want to talk to your agent. I want my agent to be able to use your system in the way it thinks it's best.
But then they push their own agent and they block my agent, and I just go somewhere else.
Well, ultimately we don't see how that will play out. But let me use my agent. Yeah, I like my agent.
Let make your service good so my agent can use it. Now we have like the agent wars. I feel like there's big companies that block their website making agent wars.
But here's an example. So shopping, going back to shopping, there's all these, let's say, companies that make sneakers, right, that are super desirable and there's more demand than supply. Well, they've had bots trying to block for a long time.
Block other bots forever. And now people are just going and they want a generic pair of sneakers and they're getting blocked, right? And even like an agent that is sufficiently aligned to the user is going to, you can just set it, hey, when this thing drops, go and make sure that we're the first person to buy it. And yet that's a thing that the companies have been trying to block for.
Well, if you cling on to your old ways without rethinking how else we could do it, maybe you have to rethink some system. Maybe the who clicks fast enough is no longer the right way. I don't know. So I bought like the Steam machine and there was just like a window of a week where you could register and then date the lottery. You know, that doesn't matter if my agent clicks in 0.1 seconds or as a human click in 1 hour. Everyone gets the same chance. So there's a lot of ways where we just need to rethink how we did things in the past because it no longer makes sense.
Instead you want to block my agent and I buy something else.
Yeah. How do you think about agents going outbound, becoming more proactive? Everything from OpenClaw to ChatGPT has been very user-prompted. I come to ChatGPT, I ask a question, otherwise it doesn't do anything, it just sits there.
But it feels like we're about to enter a more proactive. I'm already seeing it in ChatGPT, it'll recommend some prompts, and you can imagine the future where I'm more in the response category and oh, I want an agent that's always thinking. Okay, it should be thinking about this interview right now.
Yeah. And then maybe I have like a hidden screen I can.
Oh yeah, you will need me proxy.
Yeah, proxy.
It is also really hard to do. Yeah, it's a design problem, right?
Yeah. Not just, it's also a hardware problem. Like, give us more tips.
Yeah. Incredible amount of sprints.
You want looking at everything on your calendar and saying you have an interview. So I went and did a deep research. Here's what John likes to talk about. Here's what Jordy talks about. And they did all of this. They put it together, put it in a video and an audio, and they gave it to you in every single way. It's a game and I want Astro 6 ultra fast thinking about my day and not just looking like if everyone would do that, just not work today.
This is the country of geniuses in a Roomba. So to decide where to go next it down the stairs.
It's Yep. Exactly. It's deploying an incredible amount of super intelligence to solve the most mundane things. But it does feel like that is maybe the natural flow, which is in the shopping analogy, like the credit card permeated shopping and it took a pretty complicated dance of are you writing a check, who are you writing the check, pay with cash, I give you the change, etc., to just tap the card. And there's another world where the inventory and sizing is checked behind the scenes all through my agent talking to your agent. And I think any worldview that assumes that only one side of the interaction will be agent-mediated is mistaken. I think there might be a world where in the future our agents have talked to each other for the equivalent of a million years to prep us, but I don't think there's any world where my agent is spamming you with tons and tons of emails and you're flooded and you don't have any defenses.
The agent war going to fight AI with AI.
Exactly. The agent war will happen below the fold and we will both be blissfully ignorant of the trench warfare that's going on.
I mean, if anything, it's keep being interesting.
Like nothing slowing down.
I kind of dig that we finally arrived in this world where so many people cannot try agents.
Yeah. But also, you know, there's too many people that just we build all these super intelligent things, but we're not really good at explaining people what you can actually do with it.
There's still a lot of work to do.
Yeah. What are you tinkering with these days? I feel like a lot of people are getting the hardware bug, the robotics bug, the video game bug, 3D rendering. I went back into Blender for the first time in years. Are there any new areas that have interested you because of the model advances? You're like, oh, I want to go check that out, build a robotic arm that paints or build a video game or something like that.
Robotics definitely something. I bought some stuff and I'll play more. But yeah, I'm really very interested in as the models get better, how much more can we move up the stack.
So you said before about this like agents building hierarchies.
Like how can I build this? I want to build an agent that creates those hierarchies almost on demand.
Like if I tell the agent, hey, rewrite Postgres in Rust. I don't know, like okay, let's figure out. It's like oh, like is Peter okay? Yeah. And but maybe also okay is serious and how can we plan this? What structure does it need to be created on and also over what timeline because maybe you want Postgres rewritten in Rust and you want to use it, maybe you want it now, but maybe you want it in a couple months. And so it's like okay, well you're going to throw this on the back burner, cook on it for a while. I mean I've noticed this where there's certain things where I'm like oh this is actually something I'd like to play with tomorrow. I'll set a goal and run it overnight and if it's rendering something and it's sitting there, it's fine. And then other times it's fast mode. Let's get it done.
Like I was very unapologetic on trusting the models.
And where people were mocking me, oh, this will never work. It's slop. Well, look where it got me. And yes, it's slop, but also yes, we're going to make it better and the models will keep improving. And then slowly people are waking up and we move from this world where of course you look at every line of code like we scroll through it.
So and it's not stopping. It's not stopping. There's still and there's so much of how we do all those things around this that need rethinking. Yeah. And I just find that super interesting.
How do you think about your personal workflows? Sam earlier was saying historically you would want to think of 30 ideas and maybe whittle it down to one and just work on that for a few months. It feels like the success of OpenClaw came because you shipped 30 things in 60 days. But are you still trying to do that internally where you're just building again silly little kind of projects just to see what's possible to almost generate more robust ideas that you can then invest more time in.
Well, these days I kind of get annoyed if I can't change software as a prompt. Got so used to like having all the tools myself that now I'm kind of annoyed at the operating system, you know, because it feels like it's the last frontier. That's why like I started shifting more towards hey, Linux is not actually cool.
Because yeah, there's like a lot of problems, but I like problems and it's all text-based problems and my agent just goes down and reads the kernel and recompiles the kernel and now my webcam works. Yep.
Yeah. And then that gets me excited. Yeah. Like, oh wow, I can now literally own the stack and there's nothing we cannot fix anymore. Like everything is fixable. Nothing is, everything is still hard, but nothing's impossible. Yeah.
Like yeah, I could write my own OS. Maybe it's still hard, but like let's see with enough tokens and like advances that ridiculous thing will not be that ridiculous anymore in the future.
Well, yeah. It's interesting too because the agent actually becomes the operating system in many ways. Like I grew up using Mac computers and never had like a Windows PC at home ever in my whole life.
Very blessed. Very blessed.
But I got a sim racing rig, right? And so I have a PC for the first time. I'm like I don't know how to do anything on this thing. But you just set up Codex with computer use and you just talk to it and you tell it what you want to do and it just does.
I built a little ESP device and it has display issues. So I glued my webcam to the display, Codex ripped overnight and fixed the display driver.
That's so, you know all you need to do is think about ways how your agent can do the best work. Close the loop.
And then the problem was I also told her to fix audio and then I was waking up hello testing it. Hello ESP the whole night because it took so long for it to figure out that the volume was down on the other thing. Well, eventually it learned.
That's so funny. So funny. Yeah. Yeah. That's amazing. Are you, you said you put the team together, but that was digitally correct. You said you put all the whole team is now sharing what they're prompting their agents. But physically, what's your day-to-day look like? How much time are you in San Francisco?
Oh, I'm here. I'm here. Yeah. And I yeah, we basically have two teams. I do a lot of work on the foundation side. We pushed it forward and then there's a lot of things inside OpenAI where we tried new ideas. There's a lot of hey, we tried this. How can we make this into something that's a one product? Yeah. You know, so there's some cross-pollination. I take some ideas back as well because I want to play with them. Some of you disagree on ideas. Not everything works in that scale to that scale. You know, it's a very different way how you need to build things.
But it's kind of cool. So yeah, I have a lot of jobs.
I love it. Well, thank you so much for coming on the show.
Thanks for having me. Great to see you.
See you next year hopefully.
Hopefully sooner than that. We'd love to chat with you again.
Up next we have Tibo, the CEO of the Reset Company of California.
The master of resets coming on the show. Tibo, here we go. Taking a photo. Can we pan over to them?
Tibo, thank you so much for coming on the show. How you doing?
We're going to have you throw that headset on. The microphone goes on the left. I think you got it right. There we go.
I wish you brought the button. I wish you brought the button.
Oh, he already pushed it today. It was a one-time use.
Yes. But the physical button is a great manifestation of all the hard work. How is Dev Day going? What are your reactions?
The vibes are awesome in person.
Yeah. I feel everyone I talked to here seems super excited. Maybe not about the things that I thought they would be most excited about. Mostly ecosystem and you know us opening up like sign-in with we can talk.
That was the thing that stood out.
You know we had it on a screen and I stood up and just like you know I'm watching my team play you know. But to me that was a standout thing because even having Peter on just now like Peter didn't have the ability to offer massive amounts of compute to the world and so he made a bunch of product decisions because of that right the local approach things like that. And now if you're a developer and you have product experiences that you want to deliver to customers you can do that in a way that will be much much more cost-efficient.
Yeah. So maybe reset on that experience, explain what actually was announced today and then we'll go into the discussion.
Sure. But just briefly on that Peter part. I think like the great thing is you can now focus on like building great product.
You know instead of like focusing on like you know how do I make the token economics make sense?
Yeah a lot of things were announced like maybe even like too many things.
Yeah. No, that was my I mean I always default when I'm talking to you know 70-some portfolio companies when they're like we're going to launch this fund raise and this thing all on the same day I'm like just break it up launch it you know spread it out over two weeks. But in some ways seeing this live I was like okay it actually makes sense it's fun to have this moment where all these things and it creates this like pressure around the team to deliver.
Yeah. So the thing I'm most excited about because it has changed the way I work so much and I can't wait to see what people do with it is like obviously dots.
Where I think this is just a new way of packaging a lot of the things that we were seeing working before like very long running agentic the configuration that we have like the agent never stops it occasionally it sleeps.
Yeah. But it's always online, always doing proactive research, thinking about your goals, thinking about your preferences, and then you just steer it live. Just, you know, you would give feedback to anyone else. You just give your dot a little bit feedback. It learns. And so the more you interact with it and the more you teach it about your preferences, the better a job it does. And at some point, you know, a couple days in, you're like, wow, you know, just can't really quite imagine doing work without this thing. Yeah. So, it becomes like super delightful. But it's something that you need to invest a little bit of time in.
Do you think developers are already, do you think developers are more in a mindset to adopt dots or more excited about sort of diffusing AI throughout their organizations to parts of the organization that may have not onboarded workflows onto Codex necessarily because it's been like intimidating. Is that part of the next leg up here?
Yeah, I think democratizing and they're figuring out like what is the natural way to interact with these things. You know, if you're like, oh yeah, you know, like fire up Codex in the terminal, it's like, you know, it's capable of editing documents and it's capable of creating slides and you're like there were incredible marketers and finance people that were doing that, but there is value to wrapping things up nicely, right?
Yeah. And then also making sure that it's safe and it's secure like right out of the box. That you know configuration is like really minimal. It does something useful you know within the first 5 minutes and you don't have to be super technical but just kind of still feel like the power of it right you know it's like and that's what we're doing with dots. And then in the future you know we'll ship a lot of this the same technology and the capabilities like you know read in Chat as well yeah for you know our 1.2 billion users. But really it's about like getting to the essence of like you know how much can you simplify really?
Yeah. Talk about the current state of the difference between the model and the harness. It feels like this year we went through the whole narrative like the harness really matters and I'm interested in the lessons from both the harness matters and specifically Codex as a harness in Dots. How much translated? How much similar lineage is there? Or is it possible we're even in a different regime and I'm using the wrong metaphor?
Like it's this interesting tension between the model and the harness and you know you always build the harness a little bit ahead of the capabilities of the model in order to like kind of make up for it and then the next generation of models kind of catches up to that and then you have to delete part of your harness because it's actually holding it back.
Yep. And so this like it's very playful and dynamic. It's a dance.
Yeah, it's a dance, right? It's this beautiful dance of the model, the harness.
The Codex harness is used to power dots. So like the same open source harness that we have and then we also use it to power the agents API we have which is like you know managed agents. And a lot of effort is going into this harness to just make it very efficient, safe, secure and so I would say like right now that's most of the benefits that you get from this harness like just really a hardened thing.
Yeah. And then on the complexity of the tools, it's like, you know, we're just really minimizing that because the current generation of models just like so good.
And is that how I should think about the specialist dots as sort of fine-tuned harnesses for certain workflows? What is the mental model to sort of deploy those specialist dots across an organization?
Yes, we have a lot of specialist dots at company eye and the difference is that you use the existing systems within your company in order to provision the identity and then the access and the machine. Okay.
So we have them running on Mac minis.
Okay. So we provision them like exactly like employees like we mint like a new little identity and then you know it's like it's got this IAM and the roles and so we have like you know very good security story there and then we give them like a very premium machine as well because the model is incredibly capable under the hood like you know dots are like beyond PhD level right. And you know if you look at Astra it's just like incredibly capable at like computer use yeah so your dot can just like use a computer just like you would. And so by giving them like premium hardware as well as like you know it just even increases their capabilities further and so we have these set up and inside of OpenAI they do all sorts of things including procurement for example.
Sure. Yeah. On computer use can you help me understand the levers that are being pulled to speed up computer use? I noticed a jump between 55 and Astra. But I wasn't exactly sure whether the speed up was coming from like a fast mode or the actual model or the harness. Like what are the levers? Like what scaling law are we following to when I mean I was demoing it with a video game like Slire just because it's visual and I can watch it work and I know how fast I can click. And I'm interested to know like when we keep joking about it like playing Call of Duty and we're like, well, it's clearly not there yet, but is it a 10x speed increase every year? Like how do you even like what are the levers?
You're going to accidentally create an aimbot for every single video game ever created.
Maybe. But how much of that is happening in the harness and the model both? Like how do you even think about understanding the progress in computer use speed?
I do think we have seen a 10x over one year speed up and it's not that we made the model sample faster. Okay, obviously fast and ultra fast plays a role if you want to go even faster. But that is not what we improved in terms of computer use. It's both the harness and the model. So when you think about the model is like how much does it need to think before it can take the next action reliably. Yeah, that's one element that we really improved. The other one is that we take safety super seriously. So we don't just click around and let the primary agent do things like we have a second agent which we call auto review or internally we call it guardian which watches over this primary agent is like actually oh wait a second you're entering you know this sensitive information on this website like this domain is not quite right you know you should stop that action and so it intervenes there. So we also made this auto review model faster.
Yeah. And then the harness because we had Astra, Tassle talked about this like we were able to kind of use Astra at scale in order to improve the performance of the harness. Sure. Do more things in parallel. You know remove little stalls and then just really innovate there as well. It's like it's all open source. So you know you can go and dig and figure out you know exactly which techniques we used and then combined all of this together you know then we got this 10x speed up.
Yeah. Next year is the 10-year anniversary of the first OpenAI Dota competition. I think it was one-on-one and then they did the Dota 5 a couple years later. And my big question and prediction is like will a stock OpenAI model with computer use be able to beat that original model? And it's going to be close because it's got to be actions per minute. It's not just the reasoning. I think if you were able to pause the game and cook on a really advanced model you'd be able to get there but the computer use is like the final thing and I think that's going to be pretty magical.
Yeah, if you look at ultrafast today on Astra it's 300 tokens per second.
Okay. So and maybe I got to go demo it today.
Not quite that. We know one token is one action. Yeah. Right. It's more going to be like you need 20 tokens to take one action, right? So yeah you know you're going to have you know 30 actions per second which is decent amount and then it's all like how smart is the model like you know can it actually reason about long-term strategy and movement and all the past actions and what we're seeing is that currently we announced as well decisions API we further tuned the constraint so that you can make decisions extremely fast so you can make one decision roughly every 200 milliseconds.
Yeah. So speed is just like we're just progressing way faster than even I had anticipated six months ago.
And I think we'll see like you know definitely like real time actions within the next year and like beating that original model. Yeah.
So yeah I mean I love those announcements. They're always really interesting. So many different models and speeds and cost tradeoffs and paro charts and all sorts of stuff at the same time. How much do you want users to actually understand that internalize that become experts there versus just wait a little bit and it'll be entirely delegated to the harness or to the company even or the product. I'm just wondering how long we stay in this regime of understanding the model picker. Because there's some real value that comes from being able to you know think okay we need this we need this chart but we need it right now. So I'm gonna you know but that could just be in the prompt like go to your dot and say, hey I need a best estimate of revenue for my business right now. Or hey look I need you to go reconcile the books so get it to me by Monday.
Yes. The way that you think about it, it feels exactly right. Humans don't have a knob, right? So you just tell it in natural language. Yeah. You're like I need this right now or I need this tomorrow or you have a $100 in order to achieve this goal. I'm throwing a big party. Yeah. I want to spend 50 bucks. It's like good luck. But the model should just adapt and like the system should adapt and you know configure exactly like how many resources it should use in order to achieve your goal.
And that's what we're working towards. With dots in a way we already simplify things massively. There's no model picker like your dot is just there. It's available 24/7. You can text it. It modulates its effort depending on the ask. And it might not be perfect. This is also why we're announcing it at Dev Day. We want to build it with the community and like we're learning with the pro users. But eventually it will be so good that you will kind of be shocked that we had a model picker. Yeah.
Yeah. How are you spending how are you allocating your time? Because I imagine you're sleeping like four hours a night but the other 20 hours piece of mind. Yeah. The other whatever the other 20 hours a day, how much time are you spending on like recruiting and like who are you actually looking for there? You the team is firing on all cylinders. You want to make sure when you're adding people they're speeding up the team versus slowing you down. But recruiting versus everything else on your plate.
Yeah, I spend a decent amount of recruiting but I try to do more than one thing at once, right? So for example, I might talk to an exciting candidate and have a really fun conversation about an idea that I think you know either the team is already pursuing or you know something new and speculative that we're thinking about. It's like I'm kind of trying to see like you know what this other person is also you know how they would kind of handle it you know if they were on the team and so it's not that I'm you know actually kind of wasting my time in an interview and doing anything boring there I'm like actively thinking about the work and already simulating them being part of the team which then also is interesting to myself like you know as it sort of evolves my thinking sometimes in you know quite novel ways so I would say like interacting with other people and like you know I'm always looking for great new candidates to join the team. I would say it's like you know 20% of my time.
How do you think about the longer term ramifications of memory and going deeper with a particular agent and system because there's this interesting thing where right now OpenAI has business relationships and there are some companies that have gone all-in and at the same time there are individuals who have gone all-in on Codex and have complex workflows and there's more memory there. But there is a world that I could imagine where if somebody shows up on my team and they're doing their work and I notice that they're searching the internet using Bing. I'm probably not going to say, hey look this is a Google organization. But I would say, okay we use Gmail not Outlook. You're on this particular email system. And there's some sort of interaction there that feels like are we going to be more in the regime of you have your agent your tooling and I'm hiring you to work with me and you bring all of whatever tools you use or is there going to be a regime where hey you're joining a particular shop like a Rails shop you're writing Rails you're writing Ruby or this is a Python shop so I need to get up to speed on that. Have you thought about where this goes in the longer term?
Yeah, it also again it feels like the way that you're thinking about it feels very natural. So just like humans, you know, you don't want to just like reexplain the same thing over and over again. And you know, if you're a Rails shop, like, you know, you hired the person and you're like, hey, you know, you understand we're a Rails shop. Here's the documentation. Learn, you know, how we do things over here and then, you know, try your best to do it. If you don't get it quite right, I'll provide you feedback.
Yeah. And so this is very much the same that we're thinking about it for dots and agents in general.
