Satya Nadella LIVE on TBPN | Alexander Embiricos, Kyle Daigle, Jay Parikh, Jared Palmer, Michael Grinich

27:04Yeah. The problem with Excel models is you know when intelligence is created. I mean, it is like a thing of beauty. Sumptions are clear. The formulas are there. Even the formatting and the subpoena. All the stuff. And you can iterate on it. Yeah, it tells a story. It's not like a one shot. And I want to change. I can't go back and say zero shot the entire thing. So in fact, the agent mode in Excel, which I like is, it understands Office.js. It puts the formulas. I can then iterate like I iterate on GitHub Copilot. So those systems. So if you ask me about, you know, first, how do you get rid of this jagged intelligence problem is, you build a great knowledge work system that is multi-agent, multi-model, multi-form factor, get to a great benchmark and an eval where you can trust it at two nines, three nines, four nines.

27:56And until you achieve that, you're not going to be able to move and say, hey, we have anything called general intelligence. Yeah, it feels like there's, there was a lot of uncertainty in the tech community around, you know, is super intelligence going to come out of the lab tomorrow and there's going to be this fast takeoff? Now, it feels like there's more opportunity both for Microsoft to add those nines to products that you have and then also to entrepreneurs who are building products maybe on top of Microsoft. How are you thinking about the entrepreneurial opportunity in the age of AI? I think that's a very good point because at some level, if you sort of buy the argument I made, there is a lot more invention to happen.

28:35By the way, the other thing that we should also talk about, I always say to this, Stuart, right? Today, it's all the conventional wisdom is, oh, intelligence is just simple, straightforward log of compute. So throw more compute in that intelligence. More energy. Who the heck knows, man? One of the researchers comes from here, comes out of here and says, you know what, I got it. It requires compute. Any of you guys have thought about it? Like, that's a game changer. So it pre, and oh, by the way, we're all like excited about reinforcement learning. Guess what? Pre-training is a more efficient form of training because you can advertise it.

29:06So there's a, I think pre-training will have new breakthroughs. mid-training will have new breakthroughs. RL will continue to improve. We will then have to add more innovation to it. And by the way, this is another part of this partnership, which is I'm glad, you know, OpenAI is continuing to do great work. Jakob, Mark, others are great, and we'll partner with them and we'll continue to do so. And Mustafa has built a world-class team, right? You know, Karen, Amar, Lando. I mean, we have now three cool models, whether it's speech or image or text, and we're going to continue to have it. So we'll write our worst as well.

29:40Yeah. How are you thinking about the interplay between OpenAI, what you do internally at Microsoft? So our simple... Are there certain things that you can take your foot off the gas because you're like, actually, OpenAI has got that handled? Or do you want a duopoly? Like, actually, we're going to fight it out on everything. I'm much more like, again, my mindset is all platform, man. Like, hey, on Azure, do you run Windows? Yeah. Do you run Linux? Yeah. Do you run SQL Server? Yeah. Do you love Postgres? I already do. Jotnet, Java. Hey, I'm happy with OpenAI. I would love to have Anthropic, MAI, Turok, anyone.

30:12If Google wants to put Gemini on Azure, please do so. What is that like culturally? Like what does it mean for the next Satya Nadella? Somebody who's working their way up in Microsoft, do they need to be, okay, I'm building something internally, but my company isn't going to favor me? I need to fight it out with all my competitors across? We all grew up in that culture where it doesn't mean because it's always, we're going to bring our pieces together. Sure. We are going to innovate across these seeds. Yeah. But as a platform company, you kind of want to support everything. Sure. Most people don't, Office was born on the map before Windows was even.

30:51Yeah, if you don't give people choice, developers here will churn, right? They'll find other platforms, right? The concepts Bill had when he started Microsoft was, hey, we're a software factory. We love all software categories and we're just going to go create software. And so to me, we definitely want to sort of have that same attitude to innovation. We definitely need to stitch our stuff together so that they come together to solve bigger and greater problems. But it doesn't mean we can't create opportunities. And the other thing that I grew up, like, for example, you know, building SQL Server with SAP.

31:26And so we've always partnered. And Intel, Microsoft, right? I mean, we don't have the PC industry, but, you know, it is called the Grove Grades model. That's a good model to create value. Do you think that there's increasing returns? This is going to sound like a loaded question, but I promise you it's not. There's increasing returns right now to being a deals guy or innovating on the deal structuring side. And what I mean by that is there's all these difficult problems to solve with energy and data centers. And it feels like there's innovation in tech that we normally think of as like the code or the algorithm or the design of the system.

32:04But then there's also this difficulty sometimes to just marshal the resources. And is that like a new phenomenon? Has that always been true? Is there, if somebody is pursuing a career in tech, is becoming a great deals guy or deal maker, like an important path now? Yeah, I was just thinking about it, man. Like, which is, yeah, you have this great investment and it has great return and no carry all the value to my shareholders. That's awesome. Maybe we should start a venture for. You might be well there. I think the thing you're touching on is something that actually platform companies should think about, which is what's the ecosystem upstream and downstream?

32:47To your point right now, we have to as an industry. The reality is let's take power, which is if they sort of say intelligence is about tokens per dollar per watt, we've got to get efficient on all of it. In order to get more efficient on it, you got to really think about, even in our own industry, the token factory itself really getting better order of magnitude. This is like, again, a renaissance time for systems architecture. And so, you know, obviously, NVIDIA is doing great work. AMD is doing stuff. Broadcom is doing stuff. All of us are doing great work to just push that. Then the next barrier is going to be, man, can we generate energy faster?

33:27Can we build faster? Can we build a cooling? I mean, I now know more about campus cooling systems than I ever thought I'd know. I mean, and these are all choke points. How much do you want that to live within Microsoft versus you want to just be a buyer and all the different power players are out there building nuclear, wind, solar, and you're just dealing with it at a higher level of infrastructure? For sure, the vast majority of this infrastructure now. Now that, you know, if you think back at it, right, our data center builds, mostly we built and we lease some because no one was in the business of building at the scale at which we were built.

34:04But now I think there's going to be opportunities for us to lease. And there's going to be significant competition amongst builders. So therefore the lease prices also. Yeah, do you think you're more ROI focused than others that are throwing around big numbers? I mean, I'm always focused on long-term return. Well, and we're at a time right now where there's people that have come out and effectively said, I don't actually care about ROI. I just care about winning, right? And it seems from your... Yeah, a couple years ago, there was the mood of like, if you... This might be the laugh invention. You always have someone else willing to give you the billion dollars when...

34:40Or the$10 billion, you can always be about, I'm out winning. Sure, yeah. There's a ton. But at some point, that party ends and everybody needs to sort of have a plan. In that context, in these platform shifts, to be short-term oriented doesn't help at all. Because you got to have, you know, I always say long before it's conventional wisdom. I mean, if you look back, you asked how we put the billion. And the reality is we put the 10 billion. That's right. Or the 13 and a half was fully committed before it became a thing. Right. I remember that was all done before chat GPT became a thing. So yeah, how do you, I think there's a general consensus now that it's, it feels very possible to predict like a year out, two years out, and then 10 years out is extremely fuzzy.

35:28What's your view on that, given that you look like going back to the original open AI investment and the original partnership, it seems like you've had at least really good, like six-year kind of like foresight abilities to sort of invest against like a six-year time horizon. But how are you thinking about managing over, you know, the next decade? Yeah. I mean, I think, you know, to me, you know, one of the things about tech is as a percentage of GDP, get it right around four or 5%. Yeah. And if you ask me five years from now, 10 years from now, is that percentage going to be higher or lower? I think the answer is pretty straightforward.

36:07it's going to be higher. It's just a question, is it going to be 10 or 15? So why is that? Because the rest of the pie, the rest of the GDP would have grown faster. So that's why I always go back to at the end of the day, the only rate limiter here is the overall economic growth and the factors of sort of input to it. So tech as an input, I think AI and everything that it entails is going to be a core driver. And some of it will come from just this intelligence and its sort of continual march of capability. But it'll also come from, I'll call it, great engineering and product making around it.

36:47Like when I look at GitHub Copilot today with Agent HQ and what have you, that's great product. Because right now, I'm inundated with multiple models. And everything is slightly different. Except I have one repo, and I want all of these agents to come work on all of my repo in different branches. So you need great product making to bring more coherence to the chaos. And that, I think, is going to be the big difference maker. I was talking to Eric Lyman at Ramp, who makes the show possible, of course. And he had a question about what advice you would give to someone running a Decacorn, thousand plus employees in this age of spiky intelligence where there is the possibility that tools are going to get better very rapidly and maybe you don't want to scale up too fast and then up to do layoffs or retraining like you run a huge organization how do you think about managing human capital in what feels like an uncertain time does it feel more uncertain to you now than it did 10 years ago for it's a great point i mean actually eric's a great founder i I know him well, and they're doing some unbelievable work.

37:53And so, in fact, whenever I've talked to him, in fact, I learned from him even how he's rapidly changing. Yeah. The agents they have built. So to some degree, I think, as I said, Microsoft or its RAM, I think the key is learning the new production function. So when I look back at Microsoft, I feel like, hey, look, platform ships, we've navigated. I joined Microsoft when our existential competitor was Novell, right? And so in 92, and here we are. And so the bottom line is we've, over the years, navigated many platform shifts. We've also navigated tough business model shifts. When you suddenly have a 98%, 99 % gross margin server business, and you have to move to the cloud, and you don't even know, man, is there a margin here?

38:38And yet you have to make the shift and figure it out. This one is, interestingly enough, both a tech shift, a business model shift, because this is the first time you have marginal cost software. That's not like COGS of the SaaS world, but true marginal cost. And three, the way you produce your artifact, your software, is changing. So the product development process is completely getting ripped and replaced. And that is a matter, whether it's for RAM. Yeah. Even the competitive dynamic, too, because you have people that can say, hey, we can build this product in two months. Previously would have taken us 12 months.

39:15Why don't we enter that category? And it's kind of like rewiring yourself, right? Unlearning is the hardest part. Learning is easy. Yeah. Sometimes. If you have to unlearn and learn, it's much harder. Yeah. And so to me, that I think is what all of us have to do. I mean, it's funny. I met a bunch of student developers right here. I'm not sure. It is the first cohort of developers who grew up with GitHub Copilot, a standard issue. It's crazy. It's a completely different environment. They say, oh, there was a word before GitHub Copilot. Crazy. I don't want to live in that. It's not a completely different abstraction.

39:49On the topic of changing business models, shifting your business model, it seems like the console wars are over. Take me through the journey. You're a peacetime CEO now. We're a peacetime CEO. The war is over. But take me through the evolution of the business model shift on the gaming side of the business. It's one of the most interesting pieces of Microsoft. I think you've got to remember, in fact, Flight Simulator, I think, was the first product Microsoft built even before, I think, our dev tools were first. Flight Simulator was second. That says so much about the culture. As soon as you gave the developers the ability to write code, they were like, let's make a game.

40:28It's amazing. And so to me, remember, the biggest gaming business is the Windows business. To us, gaming on Windows. And of course, Steam has built a massive marketplace on top of it and done a very successful job of it. So to us, the way we are thinking about gaming is, first of all, now we're the largest publisher after the Activision. So therefore, we want to be a fantastic publisher. Similar approach to what we did with Office. We're going to be everywhere in every platform. So we want to make sure whether it's consoles, whether it's the PC, whether it's mobile, whether it's cloud gaming or the TV.

41:05So we just want to make sure the games are being enjoyed by gamers everywhere. Yeah. Second, we also want to do innovative work in the system side on the console and on the PC. Yeah. And bring, you know, it's kind of funny that, you know, people think about the console PC as two different things. We built the console because we wanted to build a better PC, which could then perform for gaming. And so I kind of want to revisit some of that conventional wisdom. But at the end of the day, console has an experience that is unparalleled. It delivers performance that's unparalleled. That pushes, I think, the system forward.

41:42So I'm really looking forward to the next console, the next PC gaming. but most importantly the game business model has to be where we have to invent maybe some new interactive media as well because after all the gaming's competition is not other gaming gaming's competition is short form video yeah uh and so if we as an industry don't continue to innovate both how we produce what we produce how we think about distribution the economic model right best way to innovate is to have good margins yeah because that's the way you can fund So interesting saying gaming's competition is short form video.

42:19It feels like the entire world's competition is short form video. Yeah. I mean, we've heard this thing a while ago. It just comes up again and again with public SaaS companies that are maybe a little bit more of a point solution, and they have to go through a business model transition. And that can be harder than a tech transition. And we hear about, oh, well, if you want to change your business model, maybe you want to be private. But it feels like, is there some sort of advantage of being a hyperscaler, a$4 trillion company that you can go and retool a piece of the business over here, or change the business model and have almost the privilege of, you know, not having shareholders come to you and beat you down about a slight shift in a business model in a subdivision that you don't get that.

43:02I can't deny that, you know, diversity of business models, diversity of the portfolio that Microsoft has, has been helpful. I mean, it's kind of, but that said, I don't think you can take that and say somehow you can make it if you don't reinvent yourself. See, I think what happens in tech, unfortunately, is that when these shifts happen, whether you like it or not, you have to first be relevant after having, it doesn't matter what the business model is. The business model may be like, hey, I had whatever, 90 % margins. You are going to at best have 10%, but you have to jump all in because even that 90 is going to zero.

43:41And so given the binary nature, you got to make it to the other side. But then the category economics matters. Because if you can't sustain long-term innovation, if there is no category economics, I mean, hyperscale is a great one. In fact, the best day in hyperscale business was the day Amazon announced their operating margins. Oh, AWS IPR. Yeah. You know, because that's when everybody knew hyperscale business is an unbelievable business. It's a commodity. But at scale, nothing is a commodity. And so to me, that is kind of going to be the key here. Even SaaS applications. And quickly unpack why that was so good for you again, just because the market recognized that you were in the same business and it was fantastic?

44:20That is one. And more importantly, it was much more expansive, right? I mean, think about our server business. It's super profitable. Yeah. Except it was one-tenth the size when I look at it compared to Azure. So we sold a few servers. Yeah. But man, we sell a lot of cloud VMs. Yeah. Or containers. Who would have thought how expansive the cloud consumption model is going to be in terms of people being able to sort of, it's kind of the Jevons Paradox that sort of really played out in a massive way. Just with Doug Rodley. Yeah. So I think on the business model. Well-timed Jevons Paradox post, by the way, back during the DeepSeek moment.

44:59Oh, yeah. That was an important. Spot on with that analysis. I wish we could keep going. Okay. I think we have to wrap up. We would love it. We would love it. And if you have a hit for 27%. Wow, that's a good hit. That's a very strong hit. It's not as big as possible. Oh, there we go. That is a fantastic signature. Thank you so much. Thank you for coming on. Thank you for having us. This is a really great time. I've always, whenever these big news, these big tech news things happen, I always wish I could talk to the person who's making the news, and now I get to. And so what a wonderful conversation.

45:37What a moment. What a CEO. What a moment in the tech world. Well, thank you. If you're new here and you tuned in just because of Satya Nadella, the CEO of Microsoft, live on TBPN, please follow us. Leave us a comment. Add us to your RSS feed. We have a 15-minute version of the show called Diet TBPN. You can hear all the flavors. All the flavors. 15 calories. 15 calories. You'll also hear ads from our sponsors like Banta. Automate compliance, manage risk, prove trust continuously. Banta's trust earns your platform and takes the manual work out of your security and compliance process and replaces it with continuous automation, whether you're pursuing your first framework or managing a complex program.

46:20You could have kept going forever. Oh, absolutely. We both could have. We should do a giga stream with Satya sometime, just 12 hours. It would be super easy to get that on this calendar. I'm sure there's a 12-hour block. For sure. Somewhere out in like the 2030s that we could lock in. But, I mean, there is really so much. I mean, that's the problem with these conglomerate CEOs. They just got too many business lines. You know, you could do a whole hour just on Xbox and Activision. Yeah, they didn't even do anything that stands out to me. The thing that stands out to me is when you see some of these other hyperscalers or players, their reaction time just, Satya makes them look incredibly slow.

47:00right oh yeah he's been making these like sizable bets like he's been seeing the future and yet only this year you've had other players who i won't uh i won't directly name deciding like okay i want to get in the game now it's like what what were you doing when uh when satya was in the kitchen cooking um it'll be fun anyway what a wild day all right what else is on the timeline what other news should we bring the folks while we are here? I'm seeing mostly people talking about 996 Porsches versus 996ing working hard. What else is driving the news cycle today? Of course, OpenAI just did a live stream with Sam Altman and the head of research over there talking about their side of the deal.

47:56All parties kind of align to, hey, we got a clean cap table. Let's move forward. PBC. Yeah, which is what Anthropic is as well, right? Yep. Microsoft now owns 27 % or$135 billion stake in OpenAI. And OpenAI is contracted to buy $250 billion of Azure services. That's a lot of Azure. should make it easy to underwrite future CapEx on the Azure side. Earnings is tomorrow. We'll be back in Los Angeles at the TBP and Ultradome. And we probably won't be live when earnings drops post-close, but we will be bringing you the news on Thursday, of course. Meta also reporting earnings tomorrow, which will be notable.

48:47Yes. We also have Alex, the product lead on Codex. but we're wiring him up so that we... I can go here through a post from Semi Analysis. They have a green text here. They say, Be me, Qualcomm. Time to enter AI chip market. NVIDIA is making money. How hard can it be? Spend years developing AI 200 chip. Finally ready for big announcement. Make fancy slide deck. Put 768 gigabytes of memory on there. Sounds big. 160 kilowatt power consumption. Sounds powerful. Add liquid cool. sounds cool uh oh shi t jpeg what about flops hmm decides just to not mention it also don't mention price or how many chips per rack or actual benchmark numbers just vibes launch presentation qualcomm ai200 it exists and uses electricity refuse to elaborate stock goes up 15 percent uh that feeling when investors don't know what flops are either uh my uh face went greater than 10x with no baseline ships in 2026 ai 250 ships in 2027 still won't tell you the specs by then probably uh low tco trust me bro confidential computing the performance is confidential unreal uh i have more there but uh let me first tell you about julius what analysis do you want to run chat with your data and get expert level insights in seconds um no you know what's funny is that in the 2019 blog post from OpenAI announcing the deal with Microsoft, they say Microsoft is investing$1 billion in OpenAI to help us support building AGI.

50:20But specifically, we're partnering with Microsoft to develop a hardware and software platform within Microsoft Azure. And so it feels like, I imagine what they mean by hardware platform within Azure is just like a bunch of NVIDIA GPUs at that moment in time. But it does lead to the natural questions of how deep do you go in the stack? And if you're SAM and you're OpenAI and you're seeing that you're ultimately limited on cloud capacity and then chips and then at first dollars, because there was plenty of Azure capacity. It's not like in 2019 they used all of it. But they needed the money. Then they needed the data centers.

51:04Then they needed the chips. And now they're needed electricity. and it's just going deeper and deeper in the stock. Yeah, the thing that's notable is just how married OpenAI and Microsoft are. When you look at OpenAI's relationships with NVIDIA, AMD, Broadcom, and these other players, everybody in the chip space is sort of like, you know, in these sort of like complex dynamics, right? We got some backstory on the dynamic between just like the whole series of events between OpenAI and NVIDIA and AMD and how that all came together. And it seems like everybody in the chip side is sort of, I wouldn't say desperately, but desperately sort of like competing for OpenAI's attention and resources.

51:52Meanwhile, Satya is able to just kind of sit back and ride this partnership out. So I wonder where Qualcomm is retraced, by the way, it's now only up eight and a half percent over the past five days. So dropped a little bit after that on public dot com investing for those to take it seriously. They got multi asset investing in food is real. They're trusted by Maryland. We got to throw in a post here from Spooks, early early friend of the show. He's quoting a post from Samuel Hammond, who said melatonin in the U.S. is sold in five milligram doses with the effective dosage range is 0.3 to three milligrams.

52:28Americans essentially overdose on melatonin by default for no good reason. Spook says, okay, well, if there's an easier way to soul speak with my ancestors in the dream realm, please let me know. I don't know if this is going to read it all on this particular string. We don't even have your tweets. Glow time banger. I mean, Rockopedia is now live. Do you think, do you think opening I will launch a, a, a Rockopedia competitor? Is I, I clicked on, I clicked on a Rockopedia. Yeah, yes. But I clicked on a Grakipedia like entry and I was like, oh, wow, it's like a pre-baked deep research report on something that I already would have wanted to search for.

53:01Yeah, actually, effectively. It's a great product. I mean, deep research has, if deep research has disrupted, in the same way that ChatGPT has disrupted search. Yeah. Deep research has disrupted Wikipedia. Yeah. And there's so many prompts that I run that I'm like, I shouldn't be, like, burning up the GPUs for this. Yes, yes, yes. This should be stored in a database somewhere. I should be able to access it. This is the frettiest thing is that, like, over time opening accounts. We have a story changes. Alex, welcome to the stream. Let me tell you about Fall while he hops on. General Need to Fire Phone for developers.

53:38Nice to meet you. Welcome back to the show. Thank you. Congratulations. Incredible event. How many have you been to? Give me a little read on the ground of like, what's the scale? Is this the biggest ever? Tell me a little bit about what's going on today. So we just had OpenAI Dev Day two weeks ago. That was awesome. It was massive. Actually, I got COVID like the day before, so I was not there. So yeah, this is actually the biggest event like this I've been to this year. It's tough with the COVID now. You don't get the same level of like, oh, I'm sad. It's like, okay, so you're sad. I didn't think it was a thing anymore.

54:11But anyways, here today we announced a couple of things. We announced that Codex is coming natively to GitHub. Okay. Very nice. And then we announced that today, actually, we're bringing Codex to co-pilot VS Code. Sorry, I'm going to mumble this. Sure. Co-pilot ProPlus subscribers and VS Code can use... Throwing mad part of this. Okay. Yeah. Amazing. Walk me through the different flows and the different actual user journeys, because there's something very interesting about GitHub has the ability to even host pages. Yeah. And then Codex allows me to, from my phone, potentially like write a web app that then can be deployed.

54:51Is that helpful to actually like instantiate a web app on the fly that actually lives on the internet that I can send to a friend? Is this like, is there, is there the beginnings? Are you, are you starting to see like what the next era of vibe coding might look like? So totally, I mean, so mostly Codex is used by professional software engineers, although we have a good amount of people who aren't as familiar with coding using this as well. But like, I think the best analogy is to think of Codex kind of like a human teammate, right? Sure. If we're working together, I could talk to you in Slack.

55:24Yeah. I could talk to you in GitHub. I could text you. Yep. I could come by your desk and we could like jam on something on your computer. And so it's kind of the same you, but you're present in all those tools. So that's what we're trying to build Codex into. It's just like an AI software engineering team that works with you wherever you like building. Sure, sure, sure. So, yeah, use it from your phone and like make some updates there. Maybe that's PR. You push into GitHub. Yep. You know, maybe you use Codex to review your PR in GitHub and then you land the PR. Like all those things, no matter where it is, it's just the same Codex agent.

55:53Yeah. Yeah. What is the, is there some sort of like business model flow through them? Like you have to be subscribed to OpenAI, but then you also subscribe over to GitHub. And that's just like the default stack for a lot of people. So this is like actually a kind of an interesting part of the deal. So to use Codex today, the main way that most of our users use us is to have a ChachiPT account. Of course. You know, they're on ProPlus, Azure Enterprise, and then they can use Codex. Now, as part of this deal, what we figured out with GitHub is how to partner. That if you have a Copilot ProPlus account and you don't need to have a ChachiPT account, you get the full power of Codex anyways.

56:31And when I say the full power, I mean, you get to use our model and you get to use our model harness, which is kind of like the code that provides the prompt. There are rules to run loop. And so, you know, our goal in this is just to, like, make Codex as ubiquitously available as possible. And so, yeah, you don't need there's no like flow through there. It's just like, you know, you just need your copilot account.

56:52there's uh somebody that just started cs in college and they're only gonna live a life that like just running codecs uh in github just naturally yeah and well i mean look so actually it's interesting right like if you think of github it github does a lot of things yeah right but at least like where i personally spend the most time in github is like actually collaborating with the other people contributing to the code base right yeah and so like i actually think it's quite unlikely that you would only spend your time doing that type of activity you're also going to spend a ton of time in tools like vs code or like the codex cli or id extension because that's where you're doing your work yourself, right?

57:28Like, again, like I think the human teammate analogy kind of works pretty well here. It's like most of us, 90 % of the work we're doing is kind of at our desk. We're not like having a, well, hopefully not having meetings like 90 % of the time as a software engineer, right? So probably that, you know, that person who's going to become an engineer, but is currently in college, they'll spend a lot of their time with superpowers, but working at their computer, doing stuff individually, like commanding fleets of agents, right? And then they'll spend some of their time collaborating with their team, like obviously quite a lot of their time.

57:57but not all of it. I don't think the sort of individual productivity is going away. How much, one kind of follow-up question, like how much, how important is the metric of like how long Codex is spending working? Is that something that you guys are like explicitly like tracking and trying to scale? Because it just means that it goes from like 10 more value to two hours. We're tracking like the meter thing. It's about like 200 years. I click it and I come back. My ancestors come and they watch. What's it into? I want to set it off. 200 year AGI. So like this, it's interesting. I actually shared at the keynote today that we lost the leak and then here on the Codex team ran Codex for over 60 hours on a single incredibly hard task.

58:40And that's crazy. And we're like, we're excited about the capability of that from the perspective of it means the model is actually able to do like very productive work for a long time. And it means the model and the harness are working together. Sure. Manage the context window. Yeah. It's more than one length. However, it's not like we have an eval that's like, how long did the model work? And let's maximize that. That's much more sort of a lagging indicator of the intelligence capability of the model. What we're really trying to drive is how smart is the model? Yeah. And how easy is it to work with?

59:09Yeah. And then it just turns out that as you make the model smarter and smarter, it can take on longer and longer tasks. Oh, what do you think about the user experience if you're setting Codex off to go work for 60 hours? There's some risk that it's doing things maybe incorrectly and you come back after 60 hours and you're like i just i kind of just blew 60 hours that i could have been doing this myself i'm maybe in that 60 hours all of game of thrones that's possible or you're watching subway surfers here we got a no but my question is like the workflow that uh that feels like the most natural give it using the teammate analogy is you just get a ping and it's like hey can you double check this before i continue?

59:52Is that a workflow that you're thinking about? You're a developer, you get a push notification, you're at the gym, and you're like, yeah, it looks good. I think there's two things you said that make a lot of sense to me. One is just steerability. That's something we're working on. We want us to be able to steer the model. Short tasks, long tasks, whatever. The other thing is proactivity. Again, when you hire someone onto your team, maybe at the beginning you're hanging out, you're collaborating directly, then you start delegating small tasks. At some point, you will give them a 60-hour task without like specifically prompting every single detail right yeah and what you expect of like a good employee is that they know when to ask you questions right yeah and so it flips from like the bottleneck of my productivity is how frequently i'm able to prompt an llm sure to the bottleneck of my of my productivity is actually kind of like how i can structure the work yep so that like independent agents yeah or humans or whatever can like go do the work and then ask me questions when they have to feel like...

1:00:47It's so interesting because we're so used to now like trusting a CRUD app, right? Like you can trust that you can put data in and you're going to come back and it's going to be there, right? We have that faith. And it feels like with agents as a product category broadly, the agent, if you're building agents, you need to be focusing on like, how do I develop trust with the user? And it's come, you know, again, maybe it's like focusing on short-term tasks initially and having that steerability. So if you think about it, Like right now, the place where most people are using coding agents is to write code, right?

1:01:20Like CodeGen. And again, I keep coming back to the human analogy, but like imagine you had a human teammate and the only thing they can do is write code. They can't read user feedback. They're not in Slack. If there's an outage, they're not going to see it. Sure. Right? So like, sorry, human teammate, even if this is the smartest human teammate in the world, like no, right? So like what we need to do to like build this trust is we need to like extend what agents can like look at across the software development lifecycle, right? So they're like present in more of the team conversations and ideation and prioritization and planning.

1:01:49They're also able, like more and more capable at the code review stage. And actually that's one of the recent product releases we ship is like Codex Code Review and people loving that. But more present at code review, more present at like the deployment stage and like code maintenance stage, aware of like what's going on and like your telemetry tools. And like, I think that's actually how you get to the trust. So some of this is like increasing model capability, but a lot of this is actually changing the form factor of like how these models are harnessed so that they can like interact with more of what you do.

1:02:15you need. So it feels like a couple of years ago, we were just trying to predict the next token. And it was like, oh, wow, I can do poetry. Oh, wow, I can write code. Like, this is amazing. And very like undirected fundamental research. Now we're in the age of spiky intelligence. Is there a feedback loop where you're actually trying to take feedback from, okay, Karpathy says that the agents can't write, you know, nano GPT. So let's go work on that. Or, oh, we've seen in the data that it's easy to write a website in Django and Python. But if you're trying to use Fortran to refactor some obscure to high-francy trading system or something like we're falling down on that, let's actually put a team on this.

1:02:55Like how does feedback work now? And like, what are the humans on the team doing? Or is it all just zoom out and hope that the emergent properties like solve? Like is it deus ex machina? You know, this is OpenAI. So the way that we build is like constantly evolving. Like the Codex team was just like five engineers like a few months ago. And now we're like, I actually don't know, but I think we're like 25. Okay. So it is constantly evolving. But what I can say is that our team is like possibly slightly unhealthily on social media, just like reading all the feedback. Sure. We love it when people send feedback to us.

1:03:28We're also starting to like try to get like a better understanding of like, okay, like how do different like model snapshots and stuff compare? Sure. And so, yeah, we're starting to build up and more in our systematic way of doing it. but I still think it's like quite early days. Yeah. And it's still a lot of taste involved. Yeah. It does feel like we're entering the era where if you see in the data that, you know, maybe there's developers here that are like, I want to use codecs for this specific thing. You're, you're, you're falling down on this. You're great at everything else, but you're not with this.

1:03:55There is the world where you can actually go in and, and run. How do you make sure you don't get the wrong signal from social media? Because like, as a certain type of person who posts about post product feedback, publicly and then there's for every one person that does that there there could be one you know a number of people that just turn yeah and and never say anything there could be a number of people that uh are just like super hungry power users and yet you know people think like getting a push notification and somebody is saying like you know they're dming you something that somebody said about codex it's like you need to make sure that it doesn't you know consume like 100 of your like worldview on how the product is actually resonating with users totally so um yeah let me answer that and i actually i want to go quickly back to the fourth topic as well but on that note i think like yeah a lot of the feedback you're going to get on social media is like your power users yep right and so power users i think are really good like i kind of put that in the like how do we advance the capabilities and like we are trying to advance capabilities so there's good feedback like what are they doing like what should the product do make easier for them yeah and then at the same time i think i like to balance that kind of with like just like what what is the first mile of the product?

1:05:01Like literally like the first 20 keystrokes, like what are those? Right. And I think for me, I just like kind of focus on mostly these two extremes. So we're constantly looking like, okay, what is the new experience to get into the product? And frankly, I think there's a ways to go. It's still a very power user reproduct and lots improved there. On the sort of like the foretrain question, like one of the interesting things about building codecs in open source, which we're doing, is that we're seeing like larger enterprises with very bespoke needs who are excited about the capability. You know, maybe an engineer was using codecs on the side, wants to bring it to work, notices like oh in this code base it's not doing as well as it's doing like the code bases that openai like sure has more you're going to see more so like what we're actually seeing is like certain customers are starting to like fork the cli or work with us to deploy it in a very specific way where you can inject like you know more instructions for you the like company specific language like into the context so if i'm a business if i'm a big enterprise and i have a million lines of fortran for whatever reason yeah uh and i you know authenticate with codex You're not training on my code, which is good for privacy, but maybe bad for performance.

1:06:05So what you're saying is that there's a world where we could work together to figure out how to actually fine tune the model or train the model or work together to have the actual product work better on my code base. Yeah. And I think like fine tuning and training are like definitely levers that exist. But I think even before that, there's like a ton of work you can do. In the harness. Yeah. In the harness. Got it. Even in like, in terms of like, you know, agents.md and just like how you tell the model what it needs to know. So in every layer of abstraction, there's opportunity to squeeze out extra performance before we go back to like, hey, let's pre-train on your data, which is not necessarily important.

1:06:41Yeah. I think there's a giant capability overhang from models today. And so, yeah, it's kind of, it's exciting. We're seeing like a lot of pull from enterprise now. Yeah. And it's exciting because we get to kind of like go deep, right? And invest a lot of time on our side to figure out how to make it work, even with like the current model. The pull from Enterprise. Do enterprises care about benchmarks or do they feel like they've been hacked? Going back to the social media thing, I think people were really into benchmarks and then pretty quickly everyone kind of assumed that they were saturated, they were gameable.

1:07:09But what are you hearing on the Enterprise side? I think, yeah, I don't hear a ton about benchmarks to be completely honest. I mean, maybe folks read it, but I think it very quickly comes down to... Yeah, think of the sass there. It's like our CRM is a split second, actually fast to the competitor like you should use us instead actually yeah so mostly i think what it comes down to uh at least on a lot of things we're seeing it well it's actually there's kind of two motions sure one is like it's just like they give the tooling to developers and it's like do you like it yeah right and it's just like what do developers like more luckily developers love codex that's right the other side is actually it's like hey like we have this like really big project that we want to do it's like a replatforming like a migrate migration from one cloud provider to another or something like that.

1:07:52And that's where we're actually like working more closely with enterprises to figure out like, okay, let's actually set up like a meta harness almost for like Codex to do this work. This started with like some customers like Instacart runs Codex like in one of these, I don't know if they don't probably don't call it a meta harness. Sure. But like in basically a system that runs Codex automatically to do stuff that, you know, that they want to do for code maintenance. Yep. And then now we've been like, okay, this is actually a pretty good idea. Like we can go help customers who have these like larger things that they want to do to set up this kind of like workflow automation.

1:08:21so there's kind of the two sides last question are you feeling gpu rich or gpu poor right now any any requests for sam and uh sarah we we so the codex team is getting a ton of support so feeling gpu rich i think i think the codex team feels supported but i think open ai like we could definitely like things are growing and more gpus would be more good so yeah definitely okay so so internally gpu rich yeah every time i wouldn't say that that's probably overdo you No, every time I'm in chat GPT now and I prompt it on something that it should really think, you know, you should think a little bit. And it's like, you know, I had a request yesterday for like, give me a list of like 50 companies that meet these criteria.

1:09:01And it was like, I can't do that. And I was like, and I was like, yes, you can. But then in that time, it was like, yeah, the GPUs were like, this is you guys, I'm sure. And somewhere out there, there's millions of codex agents, you know, running. and yeah this is the this is the uh the the endless product feedback but thank you so much thanks for coming guys uh we have a few more people joining this show if you're tuning in for the first time please subscribe follow us on youtube or add us to your spotify and before we're on our next guest we are on linkedin don't forget about linkedin yes honestly honestly i know a lot of you who are listening on another platform do not follow us on LinkedIn.

1:09:42Head over there. We are actively hiring someone. We just hired someone. We're working on ramping up our LinkedIn presence. Very excited for that. We're also excited to tell you about Turbo Puffer. Search every byte. Serverless vector and full text search. Go from first principles on object storage. Fast. 10x cheaper. Thank you for the pop. Scale bin. Thank you. Next up we have Kyle COO of GitHub. Very excited. Very excited. This is where we are wiring him up. We'll be coming on in a second. Next, we'll have Jay, EVP of CoreAI. I talked to Jay a couple months ago at this point. I was in the back of a car.

1:10:19It was kind of hard to get, to bring like the level of enthusiasm that I had, but I'm very excited to talk to him because a lot of the questions that I got when I asked people, hey, we're going to Microsoft. We're talking about their relationship with OpenAI was we want to hear what Microsoft is doing inside. What is the CoreAI team? where's that driving the business? So we're excited for that. And then next we'll have Jared Palmer, VP of Product Core AI, and the SVP of GitHub. And then we'll finish it off with Michael, founder and CEO of WorkOS. Very exciting. Excited about as well. Ken heads back to the timeline.

1:10:55Tomorrow for the show, we'll have to dive more into Grokpedia. We got to start using it, checking it out. While we have time, there's an app launched by a product designer at Meta. Did you see this? It's an AI app that if you can't afford a vacation, and The Verge is saying an AI app will sell you pictures of one. So you upload images of yourself. I saw it. And then it creates a vacation photos for you. So I guess short the tourism industry. This was announced by a founder who just said, like, I wanted to feel the feeling of, like, the warm and fuzzy vacation photos. And so I used AI to generate those.

1:11:32it was pretty well received originally, but I think that they're spinning it, and so the narrative might be getting away from them. But we have Kyle from GitHub coming into the studio. Hey, guys. Hey. Great to meet you. Great to meet you. Thank you so much. Top of Michelle. Yeah, Mass Day. Beautiful event. Good weather, too. I know. It's a little late here when the weather's not great, but when it is, it's awesome. Give me a little background on you. How'd you wind up here? how'd you wind up at Microsoft? Yeah, so I joined GitHub 12 years ago. Before we had managers. We call it open allocation now, but it was anarchy.

1:12:13Freeman, yeah. Nat joined in 2018 as part of the acquisition. So yeah, we were 140 employees when I joined back then. Wow. And do you even think about employee count in GitHub now? Because it's so merged into Microsoft, but it's still its own brand. Yeah, I mean, we have over 3 ,000 employees that have worked full-time on GitHub. Then obviously we partner with Microsoft Teams, do a lot of, you know, the AI model hosting, training, and so on and so forth. On sort of a meta question, I mean, AI and, you know, AI can do so much. How are you thinking about scaling that team over the next 10 years?

1:12:49Is it harder to forecast, like, human capital allocation in the age of AI? Yeah, I mean, part of the problem is that there's places where AI is, like, incredibly helpful. like in software, I think for sure. There's a ton of places that AI hasn't hit. You know, I mean, there's what we talk about, like IT, so many of the sort of business operations side that AI hasn't proven to be as valuable to us yet. But I think over time it'll get there. Events like this take people, you know, full in, full out. And so there's just an imbalance, a little bit of where software has been so great and then the rest of what makes GitHub GitHub, these people still.

1:13:28yeah uh earlier with satya before we jumped on we were catching up with him and i said uh my words uh github uh copilot is criminally under hyped and i think the reason for that is like you guys don't need to go out and raise a venture round every couple months and uh you know i mean you're obviously well well capitalized but uh can you give us a sense of the scale and kind of the growth of of Copilot over the last couple of years. Yeah. I mean, you know, GitHub is used by like 80 % of everyone that joins GitHub right now. It's 36 million, I believe joined in the last year, the first week, like one of the first things they do when they join.

1:14:08So we're definitely still hitting the world's developers with Copilot. Yeah. All the time, like every day. Now these days, just like, I don't know, 10 years ago, devs are using whatever tool they want and they're changing so quick. You got to keep up. You got to try all these tools, but we keep seeing folks using you know co-pilot over here and then trying out a new tool or using co-pilot over here and finding this new flow that's a big part of this like reopening up like github is done over and over yeah let's bring us all together so you can collaborate uh in that single place while you're going to pick whatever tool you're going to use and that's cool i can't tell you that's part of being that's part of being a platform i was saying what was i did when he was on the air it's like you can't uh if you want to be a platform is the more closed off you get the the more you're encouraging other people to go elsewhere because the tools are changing so quickly.

1:14:54You just, you need to be able to give people that flexibility, right? And we've had Alex on from Codex and that's a good example of it. Yeah, I mean, this AI moment, it feels a little bit like before every app had an API back in the day. Because that wasn't the norm. And now we're in a kind of quasi-walled garden moment where everyone's making their thing really, really great. You know, their model, their app, their service, whatever. but in order for all those tools and agents to actually be valuable we have to interconnect them so my hope is that just in general not just for software devs we can deal back to that platform first approach just like as an industry uh because then each of our products will be more valuable for our customers because we're not going to have to deal with well how do i actually place the grocery order yeah because there's not an api for that you know what what were the kind of key moments for you and understanding uh that ai would completely change software engineering because when you you know we're just going back through like the history even of microsoft's investments in uh open ai obviously because of the announcement today and it just feels like satya had this incredible foresight you know in in 2019 in the early 2020s that only now a lot of other CEOs are kind of reacting to, but I want to know for you, you've been here for 12 years, like what were kind of the key moments that were eye opening to you where you thought, I've seen the future, we need to just invest heavily, heavily in this?

1:16:23Yeah, I mean, the first moment was kind of a pure open source moment, right? Like when it first started happening, and we were talking about transformers and everything, like you see the groundswell on GitHub from the open source side. So we started talking about that. And then when When we got access to the first, you know, GPT-3, I think, you know, model to ultimately build Copilot, the thing that was so interesting was we were building it to write docs. Like, that was what Copilot was. Copilot was taking, oh, yeah, Copilot was taking that model. We were going, oh, you're working on it. We're going to do it.

1:16:52We're going to do it. We're going to do it. We're going to do it. We're going to do it. We're going to do it. We're going to do it. We're going to do it. So we're going to generate the docs. And then what happened? Wait, code is the same characters, actually. Exactly. And so then we flipped it, and then we got to this new setup. it seems very simple now like the idea of ghost text like more than just like an auto completion or whatever the first time we used it that was truly the moment where we've said oh crap because no one had to do something differently and i feel like that's the big problem with some of the ai tools you got to go interact with them in a way that's not normal you got to go okay i want to go write an email write an email for me that that's not how our brains work we just start typing and when we were able to do that with the IDE, that very, very quickly kind of shook all of us because it meant, oh, it won't be the same anymore.

1:17:43Now with agents and whatnot, like because we can verify the code, we have an advantage in software versus some other agents where it's harder to verify. But that all started that first time you like wrote something and then it just appeared and I didn't have to learn anything. It just happened. And now we just, you know, we all take that for granted because it's de facto. So, yeah. Do you have a philosophy of how, where inference happens will change over the next few years? Like, in a lot of worlds, there's, it used to be there's a decision between, like, fire off an agent, wait 20 minutes, wait an hour or something, or do it quicker, you know, a couple of seconds.

1:18:22But then there is the world where a lot of the work that's going on in software development is so high value. Like, why not just do all three? Inference it locally, immediately. and then also in the fast model and then also kick off an agent for every single task. Is that where we go? Or is there sort of like some sort of shift in where inference happens over time? Yeah, I think, you know, we clearly are going to have more and more inference tests that should happen locally. Yeah. That seems pretty obvious at this point, I think. And then I think when we're talking about, you know, how much we're going to kick off into what cloud ANs and models, etc.

1:18:57I think the thing that's really interesting is that we're pretty close to having that now. Like we're, you know, we talked about it a lot today. Other folks have that. The real problem is like the age old, like garbage in, garbage out problem, which is like we talk about abundance and you just fish off five tasks and we pick the winner. Well, if your input was crappy, then probably all five of those are also kind of crappy. They just did that variance of crap, I guess, you know? And so I really think it's about when you're having a discussion with a colleague or you're on a Zoom call or you're in an issue or in linear using a ticket.

1:19:32That is the moment when we can actually get as much context as is possible and ask questions then. Like, why isn't Copilot in that moment going, I think I know what you're trying to build, but are you sure about that? Why do I have to carry that even via a click and then go, let's plan to build this? Again, humans don't do that. You just read it and you get started. So I think it is a bit about where inference happens, but I think it's how early in I have a problem that needs a solution. I should be doing inference in the background immediately before I ever invoke something, you know, to go say, now it's time to work.

1:20:06It started while we were in the shower thinking about the idea. That's, I think, what we need to get the AI to do more of. Yeah. I'm thinking about, like, AI at GitHub is the funniest thing because it's, like, seven different initiatives. Yeah, yeah, yeah. walk me through your thinking on with a lot of companies, I see people like AI above the fold or below the fold. So above the fold is kind of like I put a search box and I let you interact with my app, my SaaS, my CRUD app, be a natural language. But then behind the fold or below the fold is like behind the scenes, I'm running inference over the data to improve the user experience, but I'm instantiating it with HTML basically at the end.

1:20:50But with GitHub, you also have inference that you're selling directly you have a whole bunch of stuff are you seeing any uh exciting developments on like that behind the scenes like using ai to improve github as a product that isn't actually bubbling up to the user experience directly in the form of just like uh you know a text box have you have you seen developments there yeah so i mean a big part of what we've been figuring out is like we have so much information about your interactions yeah you know pull requests the ones you got closed like i told a story the first pull request i ever shared didn't make it and like but now you know that about me and so i think the thing that we're figuring out is like we have like new uh models that uh allow us to really deeply understand your code more sure beyond just like the tabbing and asking a question because then if we have that then we have to understand what does kyle how does kyle work in a pull request yeah what mistakes does he make every single time i'm super fascinated by this that's the thing and we were just talking with like uh grokipedia today where basically it seems like the xai team went and ran a bunch of deep research reports for all the topics that you'd want to know and then you just have that pre-cached output and i'm so fascinated by this idea of you have a ton of user data you have a ton of inference it's going to be really inference expensive but what what is some sort of you know cron job that you can run over your entire user base all the data and then just have surfaced results or surfaced action items uh that seems like an interesting like under explored territory yeah if you think about today we're saying we're going to bring all these coding agents sure sure why why does each coding agent have its own memory of how i've interacted with it yeah i've been a developer for 20 something years we can just go here you go take this with you yeah you know you can understand how i work so i'm going to get a result that matches what i'm looking for okay walk me through the game theory around enterprise pre-training for coding agents so if i'm coke and he's pepsi yeah and we both have written a bunch of corporate code and you have a massive github installation yeah and if we both say yeah we're gonna train on us maybe we'll get better models yeah uh but at the same time we don't want to leak information so like what's the current thinking among like big enterprise customers around like will they jump over and say yeah you know what it's worth it or or from a from an actual like when an amp scientist just be like yeah, I don't need that code anyway.

1:23:14What's the current thesis? So we spent a long time trying this. And the problem is that everyone goes, hey, our code's very different. It's super unique. You work a certain way. You don't. Like, so many companies don't. Now, there are examples of where that's not true. Sure. Particularly really companies with a really long legacy of like COBOL, mainframe code, et cetera. We've been kind of discussing with them, what would it take to get another 100 million lines of COBOL code? Sure. Because then that does matter. That actually moves the needle under quality of the coding. A hundred percent. Because the problem is that the practices and principles don't change that much.

1:23:50And then most of these companies are also trying to modernize. So they don't want the code to look like their old code. They want to use their unique IP and look like the thing they want it to look like in the future. But if I have a hundred thousand line Django project and he has a hundred thousand line Django project, like you're not like, oh, if only we had that. No, no, because we've done it in margin of error improvements, but nothing major. When we look at the chain of commits, then you can get to some interesting information. Okay, how was the enterprise built? Exactly. That history. Exactly.

1:24:25Why was that choice made? There's a little bit there. You can obviously still instantiate that on the fly with an enterprise partnership. There's just always the question about is there something beneficial to all the companies working together? But that's very helpful. Thank you so much for hopping on the show. This was a lot of fun. You have incredible voice for podcasting. Come back on. Anytime. Thanks, guys. If you ever wrap up, we have Jay Parikh next. Oh, no. We are moving on. Okay. We are going to take you back to the news. Thank you for tuning in. We also have to do an ad read for Google AI Studio.

1:24:58We are behind enemy lines here at a Microsoft event, but we are presented by Google. Google AI Studio. It's the fastest way from track to production with Gemini. Chat with models, vibe code, monitor usage. We are obviously very happy to be supported by all of our sponsors who make crazy events like this possible. We are obviously able to come up here on short notice due to our sponsors. And what else? Jay looks like he's getting mic'd up here. Who would run Jay Parikh in? Me, I brought this up yesterday. um uh john yeah remember i brought up uh i did not know that interstellar that with interstellar christopher nolan spent a hundred thousand dollars to plant 500 acres of real corn in alberta yes uh then sold the corn for a profit after filming yes uh and it remains the most profitable commodities trade in hollywood history yes you acted like this was uh this is old news i I knew about this years ago.

1:25:56This is what I think is viral. I get inspired in the AI area of like, well, I could just generate the scene with AI, but I want to - You're going to miss out on the commodities trade for sure. No, I think Christopher Nolan just got lucky here, honestly. It is pretty hilarious. I also wonder, you know, how apocryphal is this story? Because it doesn't account for everything else that went - Like if the corn was planted by production assistance, right? Like you have to burden that cost into the actual ROI. Yeah, it's to plant corn. Who planted the corn? I'm just saying like, is this gross profit or net profit?

1:26:40That's what I want to know, Christopher Nolan. Everyone's talking a big game about the interseller trade. And it might not have been as good as you think. Also, who owns the rights? Does the value of the corn accrue to everyone who has points on the back end? Like, does Matthew McConaughey make a couple dollars off of that corn trade? I don't know. We'll have to get to the bottom of it. Or just go ahead of the budget. We have our next guest. Welcome for the stream day. How are you, Derek? Thank you so much. Welcome to this week to the UK. Introduce yourself for anyone who's been living under our data set.

1:27:17And explain a little bit about what you're working on today. I'm Jay Pree, and I am the EVP of Core AI here at Microsoft. Okay. Important job. Do you have anything to share that updates your job today based on the news with OpenAI? Or is it just exactly the same? It's exactly the same. It's exactly the same, yeah. Really? Okay, so are you marching towards AGI? Is it a race? Star and developer. Is it a race? If Microsoft becomes a platform for AGI and OpenAI can compete there and Microsoft's internal AI team can compete, is there a world where you're racing to AGI against them? I think we have a process for figuring out what AGI is.

1:28:01Yeah. And that is something that both companies will continue to collaborate, work on, research. Yeah, yeah. And in the meantime, we have this mission, which is to focus on developers. Yeah. and how we unlock way more creativity and to build a ton more things. So I have this idea, which is, or this concept, where you think about all of the potential, you think about the Hoover Dam, for example, right? You guys are familiar with the Hoover Dam? There's 9.3 trillion gallons of water behind that. It's about one gigawatt, right? And it's like massive, right, in terms of the amount of energy that it can generate.

1:28:38Right. So think about all of these large language models, whether they be small ones, big ones, closed, open ones, multimodal, video, audio, text, etc. And you think about how we're going to unlock that intelligence. And in order to unlock that intelligence, we have to write a lot of software. Right. And so if you think about the history of Microsoft, Satya commented on this earlier, you know, it's like Microsoft has been around 50 years. right and you think about all the software that's been written by microsoft and everybody in the last 50 years and i've i would posit that only one percent or less than one percent of the software that has been written in history and that what we're going to see in the next 10 years is just this like prolific expansion of the other zone where that's going to create a league you might It's going to be the enderponies, right?

1:29:30It's going to be crazy. Right. So that is why we're all here today, right? Which is like, how do we really drive and use agents, use this technology with the right guardrails, with the observability, being able to customize this, personalize it, be able to tap in and bring in open source, being able to bring in your enterprise-specific knowledge and controls and all of that. And to really just change that trajectory of creation of imagination in a building, right? And I think actually the notion of even what we think of as a software developer is going to change right now to make this way more approachable by anybody who has an idea, being able to translate that into, you know, showing something, building an app, getting out there, getting feedback, iterating on it way faster than we've historically been able to.

1:30:16In CodeGen, I want to get your read on how you're thinking about today developers, maybe they have some favorite tools, but they're willing to constantly be experimenting, trying new things. You guys are in a great position to be able to support that through partnerships and sit at a foundational layer with GitHub. But how are you thinking about what's your view on switching costs today and how that might evolve? As in five years from now, do you believe developers will continue to just want to always be trying the latest thing? Or do you think they'll like eventually switching costs will get to the point where it doesn't make sense to just constantly be looking over in other places?

1:31:00and it really makes more sense to just focus on what you have. Yeah, I think there's like an element to, you know, developers, builders around craft. And I think you're always going to want to find like the best tool or the tools that suit your sense of craft, right? Whether you're a wood, like a wood maker, you're a painter, you know, and there's like a big element of craft. So I think that there's going to be use cases where, hey, this is the way to do it. These are the best tools to, say, modernize or upgrade some version of like old Java code that you may have. And there may be just like this is the one or two just true tried ways of doing it and proper tools that you use.

1:31:40Then there's going to be new use cases that we haven't even discovered, seen yet today. Think about some of these rapid prototyping apps. And I think, you know, it's great for the ecosystem that we're seeing different startups. We have different, you know, we have GitHub Spark. We have different ideas that are all kind of competing and trying different versions of this. Those things do mature and there may be a smaller, narrow field. But I think right now with this inflection that we're seeing in terms of building and velocity of change, that there will always be lots and lots of things to go try out.

1:32:12And I think that's good for developers, right? And I think our platform is such that we care a lot about that ecosystem of startups, other companies that can bring that choice, bring those tools into it. But then we can help like hook those things together from an observability controls, like just a sensibility perspective. So if you want to scale this adoption inside of your enterprise, you need those rails, so to speak. there's there's so much there's so much like practical on the ground just make the piece of software five percent better with ai today uh there's so much low-hanging fruit it's a very exciting time uh at the same time we're in this like i feel like we're taking a breather from all the ai fast takeoff and it's exciting because they can go build so much enterprise software so much value so many new companies so many things built on top of azure and microsoft uh But at the same time, it feels like there is a new need for going back to the roots of academia or these like academic labs or these scientific labs.

1:33:11Do you have a pitch for if there's someone out there who thinks that they're going to be the they're going to write the next attention is all you need. They're going to write the next transformer paper. And you know what? In the short term, they're not actually going to help optimize, you know, knowledge retrieval or cogen for this next couple of years. But they believe they want to do it. Do you have a pitch to them where they can come and work at Microsoft and do that level of research? Yeah, absolutely. So I think there's lots of different adventures you can pick inside of Microsoft for and focused on builders, developers.

1:33:46Because one of the other fascinating and fun things about the Core AI team is we have this super tight collaboration with Microsoft Research. Yes. Right. So Microsoft Research has all of, you know, 30 plus years of history and science research and programming language research, compilers, security and you name it. Right. So we actually have a lot of collaboration and joint problem solving. Right. Where they can focus more on that open ended research, whether it be, hey, here's I'm going to go optimize this model. Here's how I'm going to do formal verification of the code that comes out. Here's what I'm going to do in terms of how to secure this code better.

1:34:22And so those things are out there. They're like big unsolved problems. There are longer time horizons. Then as those innovations, those inventions happen in research, we can do the tech transfer. We can do the combined like product making together. And then that accrues into GitHub or in VS Code or into Foundry, whatever, you know, whatever is the right avenue to bring that stuff to our customers, to developers. Yeah. Where do you stand on the should you learn to code debate? Oh, that's a good one. I think, yes, I think you should learn everything you can learn about these systems because the fundamentals, you know, ultimately, if you can understand like how this stuff shows up and it's instructing a computer, a GPU, a mobile phone, then I think that it's less about maybe even knowing kind of the code, but it's that systems thinking mindset.

1:35:17right it's the cultural aspect of it it's like hey i'm creating i'm prompting and guiding this thing but here's how the code is going to generate i understand what these models can and can't do how to guide them more with a higher efficacy right so absolutely but i think of it more as like less of a narrow question of like hey should i learn the code or not yeah it's like how do i understand the system the new system for how we're going to build software build innovation there's understanding the hardware understanding the software understanding for example evals yeah super super uh like important concept totally underreported like yeah right in terms of what's going to happen you have these offline evals we have these smarted what we're saying is to say in the media yeah in terms of like how important that is to get higher like quality outputs of these things because there's the offline evals that we can sit there and we can score and say we got these evals.

1:36:15Then there's the online or the lived experience, right? When you put this AI into this product, you're like, wait, that doesn't quite work the way it is eval said it was not a work, right? And what Mark is saying is you're in ways, right? In terms of... Sorry, I have one more on that. We got your answer on should you learn to code? I want to know, should you learn to deal? Should you learn to do deals? Is deal making underrated in 2025, in the age of AI, being a deals guy, understanding incentives, bring people together around a table, iron out a deal. This is something that's, it feels like it's growing.

1:36:51We'd saw it with the Microsoft OpenAI deal. That was a very unique deal. That was something that a lot of people, if they were just saying, oh, well. It was much more than a traditional. There were a ton of reasons not to do it. And it got done. And it's probably one of the greatest deals in tech history. So is there value in learning how to do deals and becoming a deals guy? I don't know that that's a 2025 question. I think that is a life skill to know how to collaborate and how to negotiate and how to compromise and how to see, you know, and sometimes like there isn't a deal to be made. And other times there's a greater output or there's sort of a greater like a global maxima that you can attain, right?

1:37:31And that's where even if you look at the news today with our announcements of partnering with OpenAI and with Anthropic, bringing that all into this platform together, I think is what we can go build and what we're going to discover and how we're going to accelerate our joint learning, I think is important, right? And that can turn into a deal. But I think that comes up with this like, hey, there's a greater good, there's a greater opportunity, there's sort of a greater market, there's a greater problem, a bigger problem to go solve. Then, yes, figuring out how it's going to work, nuts and bolts.

1:38:04How do you think about Jevin's Paradox in the context of code? During the DeepSeek moment, Satya quickly came out, and I think he posted the Wikipedia link to Jevin's Paradox. And it sort of like steadied the market broadly. There was people that just weren't familiar, but I think it was well-timed from his side. But when it comes to, you know, on our side, you know, we're a media company and we have a developer on our team. And I think that like five years ago, we wouldn't have had a developer. And as it's become basically cheaper and faster to create software, we now want to make software. And we're a company that historically just wouldn't have.

1:38:44So I'm curious how you think of that in the context, you know, going back to your earlier point of like we might have 100 ,000, 100 ,000 times more code. And so what's your view there? Yeah, I think that's what we want to see, the acceleration, right? I think we talked about today, there's 180 million developers in GitHub today, right? And new developers joining GitHub every sector. Somebody said it's a country. It's a country. And I was like, that sounds like miles per hour. But this is just such an abstract concept. That's how the countries talk. They're like, every second. Yeah, there's a baby born every second.

1:39:21Yeah, but you think about the types of personalities and backgrounds, right? You can be a product person. You can be a designer. You can be a marketer. You can be a dealmaker. Like all of this stuff. You can join GitHub. You can start building. You can start checking in code. You could start mashing up different things. So I actually think it's a super exciting time to see what the industry is doing, right? And I think it's hard to predict the future, but I do actually really, really fundamentally believe, like from a mission perspective, in Core AI, our job really is to unlock that creativity, both in the AI-powered tools that you heard about today, plus the platform, making these things secure.

1:40:02And really, anybody who's got an idea, wherever you are in whatever department you are in an organization or an individual, you should be able to actualize that. We have this saying in our team, which is more demos, less memos. It's all about building and showing and iterating lots of stuff gets like we don't like it you know yeah but the fact that i can in 15 minutes go through 15 iterations versus in the past i might get a quarter of an iteration done that i think is gonna matter how good a memo is like seeing seeing the product tells you you're 10 times more it sort of gets more creativity from the team a small group of people now we have to make sure we We also spend time dealing with the fact that there are gaps in the technologies, right?

1:40:51They don't work perfectly, right? So we've got to keep building those guardrails. We've got to keep building that training. The models will get better. The tools have got to get better as well. And that's where I think the GitHub community, working together with these different partners that we have, the platform, we just have to keep learning faster and faster and faster. That's what we're focused on. So more demos, less memos. Let's role play for a second. More deals. More deals, potentially. Let's role play. We're trying to do a deal. If I'm a Fortune 500 CEO and I'm coming to you and I'm saying, I want to transform my business with AI.

1:41:28I don't want to make mistakes. What pattern should I avoid? What mistakes have you seen broadly, trends that I want to stay away from so that I can move forward with something that actually drives shareholder value and isn't just rah, rah, I'm doing AI now? Yeah, so the first thing I would say is like really understand what the top one, two, three outcomes are more like specifically, like, hey, I want to transform my business. Okay, well, what does that mean? Yeah. Okay, do you know what that means? Are you saying, hey, I need to, I'm in an understand phase where I even just need to even create some bright lines around what is the ideal or kind of my dreams around the outcomes of what transformation means.

1:42:15So get into the specifics of the what that actually means. Is it some revenue thing? Is it some product thing? Is it some... You can just look at the difference between how you're trying to cut costs or are you trying to grow top line? Right. That's number one is just understanding like, one, how are you in the customer? I keep asking why and to try to get more grounded in what those specific things are. Number two is one of the things that I will always encourage them or talk to them about is to then don't just talk about these things. Right. It's like, what are you doing to start learning? Because if you're early in that journey of understanding AI, there is only so much that you can sort of like read and talk about and conduct meetings.

1:43:01you do need to have like this internal adoption, right? Where people are, and you're encouraging, you're incentivizing, you're really driving that experimentation, that curiosity of your organization, right? So how do you understand what your base level of curiosity and risk-taking is? If you are a more risk-averse company and a slower-moving company, then how do you change that culture, right? So cultural transformation, it comes up in 90 % of my customer conversations. We'll talk some tech stuff, and then they're like, okay, Jay, how do we do this people-wise? And then the third thing... Headcount planning.

1:43:40They're like, what's your plan? Maybe I'll adopt that. Right. And then the third thing that I always will encourage folks to do when we'll have a conversation is raise your level of ambition. Wherever you think you are in terms of ambition and that outcome, I promise you it's not enough. Because with the technology, the models are growing way faster. They're getting way smarter than we humanly understand. So whatever ambition you have for this fiscal year or this half or this quarter, take it up a notch or two and then strive and push and lead to that point. Yeah. Do you have a right line internally with, I feel like there's some organizations where core AI means not generative AI, but I don't think you use that exact dividing line.

1:44:22But should there be a dividing line between like machine learning recommendation systems, how Netflix recommends me the next thing to watch, for example, like that is an AI system. What pops up on my newsfeed is AI, but it's not generative AI. It's not what we think of when we think of generative image models. Is it worthwhile in 2025 to have a bright line between those teams or those skill sets or is everything bleeding together? I think things are definitely blurring together and there's stuff that's informing, you know, from one set of techniques to the other and vice versa. I do think that those systems are very, very sophisticated.

1:45:01They're very, I would say, powerful in terms of like user experience today. There are definitely places where people are using Gen.AI when they shouldn't be and they should be using, you know, machine learning techniques that just really work and are faster, better, cheaper. Sure, cheaper. Yeah. You can imagine a bunch of things. Those are the things that, you know, we have to watch for in organizations where, you know, Gen.A.I. is the hammer and everything looks like a nail when we actually have these mature, optimized and like really exceptionally bright people technology to use those and not forget about those.

1:45:37But I do think that at scale, the stuff that we've learned in these maybe more mature, more scale-out machine learning systems will feed back into how we make products using GenAir. Well, thank you so much for coming on the show. Thank you, fellas. This is always an LF. Take care. We have Jared Palmer, the Vice President of Product Portrari, and the SVP of GitHub. The SVP. The V. We're going to wear it. Let me. You're both a vice president and a senior vice president. Yes. Is this like a two phase situation? Yeah. Title max. I might just.

1:46:21So to regional branch manager. Yes. Right. Yeah. Yeah. It's like assistant to the CEO, assistant CEO. Technically, it is VP of product core AI. Okay. And SVP of GitHub. Okay. does this incredible welcome welcome to the gig thanks monday 30 13 oh 13 yeah okay what do we do more for the month i was uh vp at ai ever sell that was right we had darymo on the show yesterday we rated a little thing called v zero zero yeah resolution yes uh so and been in a game for i don't know a little bit doing that stuff so yeah it's been fun at for something great yeah so i mean have you had time to actually develop like a vision for what you're building here is it too early to ask or or are you still just in kind of like let me assess the tools in the tool chest over here uh yeah it's day 13 so definitely uh but i've been a long time get up user for for very long time like i don't think over 10 years yeah i made my account so um and i imagine you've been thinking about like a broader developer experience and what this means in the age of ai all through the last i mean the last five years have been like a deafening ring of like AGI and takeoff and timelines and stuff.

1:47:33You must have engaged with that, of course. Yes, yes. And obviously at Vercel, we've thought deeply about developer experience. I think that's really the vision is how do we bring apart with Core AI and the formation of it? We just had Jay on. I think by combining Microsoft's assets across the stack, right? I've got DS Code, Visual Studio, GitHub, and putting these actually all in one work make for the ultimate developer experience. And that's what our goal has to be. And focusing just on that is I think my first and foremost. Yeah, do you think that developer label just melts away eventually?

1:48:02It feels like you think there will be a dividing line in five years, ten years? I don't know, five or ten, I guess. I mean, it just feels like there's a world. I don't know if that's about that. But it's just feeling great. You know, like there was a time when to take a photo, you needed to be a professional photographer because you needed to understand how to change film in a darkroom. And now everyone has a smartphone camera and everyone's a photographer. That feels like it's coming. I don't know. I just see like, I can open up an app on my phone, type a prompt, get code. Sure. It's like kind of hard for me to, I need to link my GitHub account instead of pages to like actually deploy it.

1:48:38But like, we're only a couple months away from that, I feel like. And then eventually it becomes like more prompt driven, but then there's still value. I don't know. How does all this play out? I think there's always going to be a market for people who get stuff done. Yeah. Right. Yeah, just high agency people. Lighter people. So builder who, and whether it shifts into more product focus. Yeah. Knowing how to build systems that are big and large. Yeah, yeah. That may be outside the training set. Sure, sure. I think it's always going to be important. Out of district. I also think that some of the, the way I think about it at least, is some of the pipes, the tooling, probably aren't changing as fast as the AI is.

1:49:13Yeah. What I mean by that is like the way that packages and code is distributed, tested, built. I don't think that's going to change as fast as maybe the models will, if that makes sense. So with that infrastructure in place, I think you're still going to have human involvement for quite some time. I think the things that people will build may be more ambitious. I think that's really exciting. And our job is to facilitate that and empower developers and think about what they need. But in five years or so, I still think people are going to be building stuff. They're still going to be coding in some respects.

1:49:43it just may look very different. Where do you want to see model progress? People talk about the models are going to get better. Like they're just going to get better. Right. And plan around that. But like when you're talking to labs, like when you're at Vercel or when you're now at Microsoft, like where specifically are you even thinking and kind of pushing them to say like, hey, like it needs to be better here. Ah, yeah. I mean, that's a great question. At Vercel, we worked deeply with the model labs. We obviously were very focused with a product like B0 on a specific subset of what models can do.

1:50:18Even in the coding realm, Vercel was always focused on front end, right? And specifically Next.js. So not just one language, but one specific tech stack. And so we're always engaged with how can we make it better for Next.js. Switching gears for a second to GitHub, obviously we're now multi-languages. We care about everything, but we do care about coding. That's the primary focus point. But coding involves so much more than just generating like more than auto-horse code, right? It's more than autocomplete. We need models to be great at research, to be great at reasoning. And I think, and then also delivering mergeable code, right?

1:50:55That's, I think, something different than just complete my comment. So we've been focusing a lot there and focusing on quality and something that we look to continue to hill climb on as time goes on. How much have you studied the open source company, like scalable business model? Like what Vercel did with Next.js? Are you familiar? Can you give me like the crash course? If I'm like, I'm a developer, I want to build a business, I'm going to open source a package that does something and then I want to build a business about it. Like, what are the pitfalls that I need to avoid? How do I actually balance?

1:51:32Like, what are the trade-offs that I'm making to actually build a great, like, open source, for-profit company? Because there does seem to be some tension there, but it's held, that model's held for going back to Red Hat Linux all the way to Vercel today. Sure. I think I'll, I'll, I have a controversial take. Please. Um, there aren't as many pure open source companies where the core product itself is open source. Sure. I think the more successful strategy is actually, if you really dig into Vercel, because Vercel is not open source, but Next.js is open source. And Next.js is a complementary satellite product that drives attention, that is used by Vercel to make a better product.

1:52:13They've got this amazing feedback loop of internal dogfooding. but there is a community around the project which then I think some, you know, Vercel has a material amount of Next.js overall builds and developers use Vercel but it's not like Vercel is an open source business, right? It just has Next.js as one of its largest pieces of the open source portfolio but it also has now AISDK and with Vercel, the idea was to do something what we used to call framework-defined infrastructure so framework-defined infrastructure and the idea was you can build this framework and with no configuration you can deploy it and you don't have to think about scaling it.

1:52:51And so the analogy I would make is like, imagine you were asked to, I don't know, cook food for everybody here at Universe. With Vercel, the idea was like, oh, what if we gave you the pots and pans and all you had to focus on was like cooking for your family of four. And then Vercel would worry about like scaling it to everybody here. And so I think to your point about like open source, my suggestions for the crash course is, a common pitfall that you should not run into is just assuming that your free open source users are going to directly translate into paying customers. I think that's actually really hard because you've set up expectations that you're giving away a free service.

1:53:30We have this free, this code, right? Yeah. And then all of a sudden they're going to convert and pay you X dollars a month. You're going to have an enterprise business which you haven't been really honing in on and grinding on. And that's just going to happen overnight. I think that's, I think that's, you need to start from the beginning with both and also set expectations with your user base that this is paid, this is open source. So if you can find a beautiful symbiosis between those two, that's where I see like it really being successful. Is there some sort of like barbell strategy where you should actually go really broad with your open source package?

1:54:01Anyone's using it, but probably like, you know, small developer startups, solo indie devs are using it. And then if you jump all the way to like, oh, you notice some big corporations are using it. So you go with an enterprise plan on day one. And it's like, they're not going to be, they have no ground to stand out if they complain. They're like, so, so it's a lot easier than being like, okay, actually I'm nerfing the open source thing. And now all the indie devs need to pay me 25 bucks a month. You know, that's way different than going like, hey, look, fortune 100 company was using this. Now we got a million dollar contract with them.

1:54:32Is that best practice? It's hard. Sometimes this big contracts early on can really be devastating. Oh, sure. Because they can remove your focus on growing that inertia. Oh, yeah, yeah. And so you have to be careful. Okay. Obviously, they're great, but focusing on your core value proposition, your core customers, and it's really great to get feedback by those enterprises early on. And many, many projects I've been involved with, whether it was Turbo Repo, whether it was Next.js, even VZero, like we didn't launch Enterprise for almost a year or so. And we even, I even, it was a big debate between me and Guillermo.

1:55:07I think we were actually early. You should actually delay it even further. Really getting that groundswell is so important. and you can always do enterprise. Okay. You'd be careful. I say always do enterprise. You'd be careful. Yes, somebody could come in, but just driving up, even like ChatGPT, by the way, didn't have enterprise for like a lot. People were like, I came for it. Yeah, yeah. And then when I find out, a lot of companies report like, yeah, we don't pay for ChatGPT, but our employees all use it. And then you go to the CSO and you're like, hey, by the way, we have a lot of your data.

1:55:32I know exactly. I can't wait to see that. It is a wild choice. How do you think you know a lot of there's so much excitement around the potential AI and science and law these other categories and obviously adoption is happening but how do you think adoption will will kind of how would you imagine adoption will look in those categories because I think AI adoption in software engineering is very natural because the people that are building and and doing the research are adopting the product and like it's a super tight feedback loop and you're not really going to see in the same way and some of these other categories so uh yes and i don't know i before i got into software development i uh i was actually um a banker and so let's go yeah go go the sax speak thank you very much let's go uh so uh i did i did my sank or yes or banker right so i i think i gotta tell you i'll be honest like I think Anthropic, I was talking to Mikey, they announced Claude for Excel.

1:56:41I think that's going to do wonders. I think if you talk to any Goldman Sachs analysts, they'll be very excited to have that deeply integrated. And if you're building Excel Mara's all day. There's whole businesses that are built just on templates. Totally. And it's like obviously. What's interesting though is if you look at Claude for Excel, I think their core foundation is still the coding agent. And that there's something about the coding, the coding runtime that can be then augmented to other verticals. I think that's what you're going to see in next year or so is, is, is these model labs build out these harnesses and go vertical by vertical, whether it's banking, healthcare, or consulting, right?

1:57:17They're going to go through that through knowledge work. And they're going to iterate on that just like in the hill climb. Yeah. Yeah. What do you think on how do you how do you think about switching costs now and over time if you're a products company and you're leveraging intelligence from a lab? Like, do you think the labs will over time make it harder and harder to kind of like rip out one model provider and use another? Because right now it feels like there's this land grab happening in enterprise and this race between Anthropic and Gemini and OpenAI. But like, how do you think that evolves?

1:57:55I think most of the product builders that I've talked to, like the companies that are on here all the time, most of their teams are working with multiple models. and they're constantly evaluating whatever sort of product analytics or test harnesses. They're looking for any edge they can. They're so competitive. Yeah. At least in like the startup space that like... and but switching models is not easy that takes time and especially when there's big rearchitectures like when reasoning came out, for example. That may require a rewrite of all the prompts and all the edge cases that you've been massaging.

1:58:29And these models have different characteristics. But I think most of the high-performance teams are dialing in harnesses for each and every lab, and they're just so hungry that... So switching costs are high, so the answer is, like, you get all of them in the beginning. Correct. And there may even be certain subsystems or certain tool calls where you're going to switch models and mix them together. And that just is, you know, part of the part of this. If you look at like what Windsurf did, where they released a SWIG rep, that specialized model for research. You know, they're combining, they're mixing and matching.

1:59:03I think that's the next, you know, we'll see that throughout the next year. I don't think it's like, oh, we're just going to use anthropic models or we're just going to use opening eye models or we're just going to use, you know, whatever model. I think you'll see a lot of combination. Thank you so much for coming on the show. Thanks, guys. Big fan. Congrats on having a great come back. Come back every time. Thank you. before we bring in Michael Grinich from WorkOS, we got some breaking news. Did you see the blimp? Did you see the blimp? Do you know whose blimp that is? It's Sergey Brin's blimp.

1:59:32Let's go. I mean, this is bag seven on bag seven. I'm actually going to take a little bit of credit for that. I told the Gemini team. Oh, you told them? I care. I haven't texted Evan Logan. What's up, guys? Hey, good to see you. Great to finally meet you. We had to see you as your old friends. Yeah. David Serra. Oh, yeah. Mutual friend. David. Love David. I got you. Brought you guys both. Oh, please. One of our highly coveted, super rare, enterprise, enterprise-ready. Ben, over. Okay. The software. How are you, enterprise-ready? All right. Square, enterprise-ready. What does it mean? Well, pretty much every software company, eventually, when they get product market fit and go up market, there's a ton of stuff they have to add to their app to go sell the enterprise.

2:00:13Yes. So, the guys at Microsoft and GitHub, they did this years ago. Yeah. But if you're a new company, you have to add all this stuff to your product. Sure. And it's things like single sign-on, user provisioning, logs, security. WorkOS just does all that for you as a developer. Got it. Yeah. Okay, so you can just focus on the core products. Yeah. Yeah, and the same way you use Stripe for payments or Twilio for messaging, WorkOS is really that. This is what the interesting thing is because it feels like it's not something that you could just go through YC and sell to another startup. So who was the first client?

2:00:43How did you get into this? What were you doing before? I started working on this a long time ago, almost seven years ago. So WorkOS. Overnight success. It's kind of pain. Overnight success. Yeah. We're also like a pre-AI company. Someone called us the other day, which also kind of hurts a little bit. I know. Shook my ears. Dinosaur. I started as AI native before AI existed. Yeah. For real. I saw this problem with another company at the start. Okay. We had built an email product. Yep. Got a bunch of usage. Got a bunch of adoption. Try to go to enterprise. Try to sell it to these guys. And they said, no way we can let this cut our data.

2:01:13And you're saying, no, is the CTO, CISO? Simply engineering leaders, co-founders, VP of Eng, whoever is kind of responsible for the technology. But is it because they want those features or they need them for legal reasons? They got to have them. They usually have deals that are blocked because they don't have these features. So you'll start growing up market and there'll be some customer that says, hey, we'd love to use your product. We'd love to roll it out, you know, at Coinbase or Microsoft or something. But we can't do it unless we have these features. Has demand just been insane because people are building products so quickly and then they start, you know, employees, employees, of companies started adopting them kind of personally and then they realized compounding yeah sure so we had a lot of growth you know years ago through the kind of the early cloud era sass like brucelle is one of our customers carta plaid folks like that in the last year and a half two years in the last year and a half or two years what we found is it's actually perfect for all these ai companies yep so today we're powering enterprise off for open ai anthropic perplexity cursor sierra you know all these guys that are growing faster so i got a lot of that button yeah okay uh walk me through the thesis uh in the yc era it became uh like the yc trade was basically you could be a kid in college uh you know graduate and move to y move to mountain view or silicon valley and for a hundred thousand dollars and some cloud credits from azure or whoever uh you could set up a website and go kind of build the first era of consumer.

2:02:43And we got our Airbnbs from there. We got a variety of consumer companies. But in the AI era, it's becoming easier to go enterprise on day one. Is that real? Is that a reasonable thesis? Do you see any data to that effect? Absolutely. So I think that previous era, the privilege that those companies had is they could take a while to get to enterprise. So if you look at Dropbox, Figma, it was years. It was like three, four, five, six, seven years before they actually went after enterprise. What we're seeing today is AI businesses get pulled up market way faster. Sure. And it's way more competitive.

2:03:15So companies like Cursor or Complexity, pretty much in year one, year one or two, they get pulled up market to the enterprise. And that's why they need this. Yes, because of that competitive dynamic, but also the tools that they're building. Like the enterprise is just so ready for them. There's another piece of it as well. It's not just that they grow faster at market, but you think about these AI products, they are touching sensitive data. You have one of these things that it's only valuable if you get access to all of your stuff. You give it access to do things on your behalf. So suddenly it becomes this huge security concern.

2:03:45Maybe an old product like Figma, you could say to the design team, just don't put any sensitive data in it. But you get one of these agents or something connected, you need it to access everything. And so they're scrutinized at a higher amount, plus they grow faster, plus in their life cycle. It's a perfect storm where we come in and help them grow. Talk to me about domestic versus international. I imagine a lot of your clients are already international. So does that mean you're international? Or are you focused on making American companies enterprise ready immediately? And then maybe you'll go after the European market later?

2:04:14How do you think about that? So many of our customers are actually right here. Yeah. Like probably at the universe, literally right here. We were joking we could cut up our sales territories by north and south of Market Street in SF. Because we have so many businesses that are here that are growing quickly. What we find is their customers are international. Of course. So they're going and selling to larger organizations elsewhere in the world. So our products that are kind of our customer's customer, those types of things we localize. We just did a big project to translate everything using AI.

2:04:41We launched with 100 languages. So we do that kind of stuff. But we find that the best product, the best companies that are using WorkOS are these high growth AI businesses that are taking off. And of course, they're mostly here. They're mostly here. What's your philosophy around operating the business? I don't remember. I don't necessarily recall the last time you guys raised money. Like, I'm sure people are throwing money at you all the time when they see the logos. Yeah, we raised our Series B almost exactly four years ago, actually. That was the last financing. Yeah, that was the last financing, which is like an eon in the SaaS era.

2:05:14Yeah. Since we've just been building it slowly since then. You know, bit by bit by bit, brick by brick. You know, slowly. Actually, I do have something to announce that's pretty exciting. You know, you had Asati here previous talking about how they do a billion in revenue every day. We're very proud to announce that we just crossed$30 million in annualized revenue. Amazing. So that's our big number. Overnight success. We're seeing you today. That's great. We're a bit smaller than Microsoft, but we're coming for you. We've been compounding since then. The AI stuff has really been this huge tailwind for us.

2:05:54And it's so fun to build infrastructure where we get a C into all these companies. Like my customers are the fastest growing, most exciting AI businesses out here. Do you invest? Do you angel invest? I do some, yeah. Could I imagine seeing these companies at this like crazy fluctuating point? I've had VCs start asking me for the data. They want to invest just to get the growth data out of it. Yeah, it's a little difficult. Yeah, we don't share that kind of stuff. But just through, you know, building for developers and running events. I mean, I love GitHub Universe. This is like Coachella for like, you know, developer stuff.

2:06:26You meet founders and meet other people building stuff. And so, uh, who are some of the entrepreneurs that you look up to? Who, or when the, what the story from a founder or business person that you keep coming back to is like, oh, that one. I'll go for it. Just David Senera. Just his story. Just how he, no, I do. I think, I think I, I, I, he is like, I, I feel like I have the blessing of like, I, by being friends with David, I'm friends with history's greatest entrepreneurs. It's the most efficient role model, right? He has a specific type of business, but whatever business you're building, you can learn from.

2:07:03Yeah. I mean, yeah, there's a lot of other things you could have said, but that one's pretty good. I was trying to think for a minute. David's great. I'm just laughing at me. It's like, like, like the stakes are like, you know, Henry Ford inventing the, you know, the, the, what's it called? The actual assembly line. Yeah. The automated assembly line or like, you know.

2:07:29It's something the most impactful versus like the most valuable for you. Yeah, I guess. Who do you come back to? We're very much in a marketing business. And so I always come back to a quote that I attribute to David Senra, but is actually from David Ogilvie. You are not advertising to a standing army. You are advertising to a moving parade. Yeah. And so the question is like, why are people... Proving my point. You heard that for the first... No, I read O 'Leary on Advertising. I'm familiar with the book. I read it before David read it. But he did stick it in my brain and he advertised it to my moving parade.

2:08:07And I've always liked that idea of even if you've shown someone an advertisement once or you've sent them a message or you've given them a pitch once, like you there there is a moving army there's so many distractions there's attention all over the place that you need to be hitting again and again and it's why they don't just do one github universe and say yeah we did it we're good they do it every single year and the message is different every year yeah they're evolving yeah man there's so many there's so many to choose from i'm i'm i feel like a little bit of an old soul in that when i when i heard satya talking about like the early days of microsoft and bill i love building platforms yeah like i built all this other stuff earlier in my career.

2:08:41And as soon as I started building stuff for developers, other people making stuff, I was like, ah, that's really sick. Like you see other people make stuff with the thing you made and then build their own businesses on top of that. And to me, Microsoft is like the first big software platform company. You know, Windows enabled so many developers to build and ship these experiences to change the world. And it's, you know, previous cycles. I think a lot of people forget it, but it was this huge enabler, this huge like democratization of access to technology. Is there a, uh, a specific sales funnel that flows through GitHub with your product?

2:09:15We do just a ton of stuff with developers. I think, you know, uh, I mean, we do everything from, you know, sponsoring podcasts and newsletters and meetups and, and doing developer events. Our, we did our own conference last week. Um, we do a lot of open source stuff. We run a really popular open source design system project called Radix. Oh, that's yeah. So that's on GitHub. Um, But I think my GitHub account is probably one of the earliest personal identities online I had an account. I've had it since before I was in college. So I'm thrilled to be here. Yeah, that's very cool. Are you speaking all day?

2:09:47I am. I'm giving a talk tomorrow all about AI and identity for agents. So this is a new thing. WorkOS is kind of like an identity security company. You help people with sign in and off. And there's this big question right now of how we're going to secure agents. Yeah. You know, we have 7 billion people on the planet. We'll probably have trillions of agents running around and doing stuff for us, connecting to different systems. And security is even more important. You can think of an agent kind of like a crazy hyperactive intern. You're going to have access to all of your systems. And so there's this question of how do you authenticate them?

2:10:19How do you build security around it? Permissions, approval. Prompt injected. Right, right. Yeah. There's this old quote, you know, to err is human, but to screw up 10 ,000 times per second, you need a computer to do that. Agents are kind of like that, right? They make it really easy to do stuff really quickly, but also make mistakes. So my talk is all about that and some ideas that we have around security for. Yeah. Do you think agent security bifurcates along the consumer and business to business axis? Do you think there's a discrete enterprise versus B2B layer? It feels like we talked to the CEO of 1Password, for example, and it does feel like the password to my Yelp account, I might not be using WorkOS for that in the future.

2:11:03There's definitely going to be some blurring.

2:11:08The way I'm thinking about it is a small business might have 200 agents that are out in the world. Maybe some are selling, maybe some are doing customer support. When should a CX agent be able to share account data? Right. When when can a sales agent like provide pricing? I mean, like there's so many different things and you, you know, CEOs and they're working with teams. Right. They have like processes in place for individual people. And so I think I think it's like really important problem area. It's it's completely changing the way people think about security. I think if you go to any of these like security focused conferences, it's the topic on everyone's mind.

2:11:49Yeah, exactly. Within. Exactly. Yeah. Because within companies, previously you've had these kind of silos of information or control. You have permissioning systems that are pretty static. With agents, you know, you might have 200 today and zero tomorrow. You might spin them up and down depending on a task, depending on a project. And so that permissioning model is like completely changing. And it's really exciting. You know, we're right in the middle of it. Like us working with all these different AI businesses, they themselves are building their own agentic workflows. Whether it's, you know, stuff like Codex or Cloud Code or what Cursor's building with their.

2:12:20Is there an AI horror story in security yet? Oh, yeah. People can talk about it. Is it possible? I mean, there's ones like, yeah. Do you know about security or death? I think that steps down. It feels notable. There hasn't been a specific day on the internet that everyone was like, Oh, we went down because of AI. Yeah. That hasn't happened yet. Not yet. I feel like probably just hanging down the days. I mean, even in the latest AWS outage, I don't think no one pinned that on AI. No one pinned that on generative AI or stochastic systems. Yeah. There was one a few months ago where Jason Lemkin, you know, from Saster, he was vibe coding an app on Repplit.

2:12:57Oh, I saw that. Do you remember that? Yeah, yeah, yeah. And he was like, he was pure prompting, right? He's not writing any code. He's like, just talking to the thing. And he asked the agent to do something and it deleted the full production database. Yes. And then he was like, what the hell? And then the agent lied about it. Yeah. And it was like, no, I didn't do that. You know? The agent was like, yeah. He's like, yeah, cover it. At one point, I think the agent did just say, yeah, you got me. Mine bad. Yeah, mine bad. But I do think they were able to roll it back. So I remember getting in the thread and kind of just to close it out and not leave the...

2:13:27Yeah, and they've built a lot of stuff since then as guardrails. But that just shows you, like, early on, people pushing these systems to their limit. And they can have catastrophic effects if you don't put up these guardrails. Yeah. So that's what the talk's about. We're doing a lot of innovation and research here. But it's going to take a while to get right. Yeah, yeah. Well, awesome. amazing to finally have you on the show thanks so much you've been a long time fan it's great thank you so much have you come back on again i will take care see it we'll talk to you soon um there are a lot of posts here in the timeline that i want to share that we can't timelines we can't uh i mean we're not paying them are being held back by i want to just come back to them tomorrow okay held back uh tomorrow you have our word we're doing lots of timeline if you're new here leave us a subscription follow us on x follow us on linkedin subscription everywhere yeah sign up for our newsletter tbpn.com uh we bring you the news in text form yeah uh well it has been a fantastic day here in san francisco thank you to this everyone stunning out it is i'm gonna go try we gotta go hunt that blimp we gotta find out it's not the gemini blimp i'm telling you it's sergey's blimp his personal blimp it's not a it's not a gemini project it's not a google project i thought you said it was branded no no no he's individually been funding a blimp company i know it's so funny because i sat i sat down with logan and the gemini team we were just talking about marketing ideas and i was like the obvious thing that you should do is get a blimp wrap it with gemini branding and just fly it around uh san francisco Founder Mo, Sergey's on top.

2:15:05And we were talking, it's like, okay, finding a blimp. I was doing some research. There's like six active blimps. I was like, man, this is going to be hard to find a blimp that can get to SF, that can be wrapped. And of course, Google, incredible foresight from Sergey to create a beautiful billboard in the sky that's just waiting for branding. But waiting. Well, super fun day. A surreal moment. A lot of fun. Talking to one of the greatest living CEOs and thank you to everyone on the Microsoft team thank you to everyone on the GitHub team who helped organize this thank you to our sponsors profound linear numeral hq.com sales tax on autopilot bin.ai the number one AI agent for customer service Adio of course customer relationship magic aid sleep didn't sleep on my aid sleep last night I can't wait to get back to it tonight.

2:16:03Of course, we mentioned public.com. Also, adquick.com. Get bezel.com. Your bezel concierge was able now. Jared Palmer had a nice job. He had a nice job, Master. It was good. It was looking good. And of course, Wander. Find your happy place. Book of Wanderth. Inspiring views. Hotel great amenities. Dream events. Top tier cleaning and 24-7 concierge service. It's a vacation home, but better. Thank you, folks. We will be back in Hollywood tomorrow. The podcasting will continue. 11 a.m. sharp Pacific. Cheers. See you then. Goodbye. Bye.