Citrini Memo Reactions, Kim K Enters Energy Drinks, Jane Street Sued | Patrick & John Collison, Bill Gurley, James Cadwallader, Scott Wu, Ivan Zhao, Stefano Ermon, Rune Kvist, Reiner Pope, Devansh Pandey

15:07Scott Wu:The open source version of the open source thing is like you're basically paying someone to maintain it. Yeah, yeah.

15:11Patrick Collison:Host it, manage it, make sure uptime's good.

15:14Scott Wu:But like models keep getting better and they'll just like do all that for you. They'll do all that for you. And it's like, yeah, I think that's like very obvious.

15:21Patrick Collison:Yeah. It is possible that you would just say like, okay, run in a loop and just go around and fix everything. And if there's uptime or security patches, like patch them immediately. So the open source software gets better.

15:32John Collison:There's two narratives, right? there's the okay everyone will just vibe code everything in any department yeah you can just have an employee just make the software tell the agent not to make mistakes and or tell the agent hey fix this thing so that's that's one thing and i feel like that is a maybe a part of the sell-off but the bigger reason for the sell-off is maybe what derrick thompson is talking about which is that like the world is getting weirder. A lot of people are feeling the acceleration. And if you just don't know what the world looks like or what work looks like in five years, you want to take some risk off.

16:12John Collison:You're not willing to pay the same revenue multiple that you were three years ago.

16:17Patrick Collison:Yeah, I do want to dig into that point that you mentioned earlier a little bit more, which is like you have the Tyler philosophy of like you could vibe code everything and the agents will be able to go around and maintain and everyone will have personalized software, individuals where the value accrues to the person using the software, but then also the lab providing the software, the inference. And then there's like the private markets boom right now in AI enabled software where companies are saying, well, we were able to pull our roadmap way forward. We got to an MVP in a weekend and we're able to ship features way faster.

16:52Patrick Collison:So when we onboard new clients and they ask for something, it's like, boom, and we get it done in a few days as opposed to a few weeks of engineering sprint. And so the narrative is like, we're moving faster and we're creating like AI-enabled products that couldn't exist otherwise. And it feels like both of those can't be true. So I don't know which way we'll land.

17:11John Collison:Ben Thompson called out the real estate example. He took a segment out. This is from Citrini. Even places we thought insulated by the value of human relationships prove fragile. Real estate where buyers had tolerated, I'm saying this in an extra dramatic voice. Love it. where buyers had tolerated 5 % to 6 % commissions for decades because of information asymmetry between agent and consumer. Crumbled once AI agents equipped with MLS access and decades of transaction data could replicate the knowledge base instantly. And then Ben Thompson says, the real estate example makes the exact opposite point the author thinks it does.

17:50John Collison:The truth is that the internet already obsoleted real estate agents in terms of information flow. You can go online right now and get a listing of every house for sale with pictures, its full history, etc. There is no information asymmetry, but rather information abundance. The fact that real estate agents still exist despite that shift is actually one of the more compelling arguments that humans will be remarkably resourceful in terms of giving themselves jobs to do, even in areas where they ought to be pointless. Yeah.

18:17Patrick Collison:How hard is it to disintermediate a real estate agent? Does it happen on the buy side or the sell side? Like if I find a place on Zillow and I go knock on the door and so, or I write them a letter and I say, hey, I'm, I'm, I want to buy this, but I don't have a real estate license and I'm not using realtor and I don't want to pay a fee. Will they be like, cool?

18:38John Collison:I think, I mean, you can do it from either side, but I will just say the reason that you don't is that I'm in process. I'm in currently an escrow on a property and the guy representing me is going to make a lot of money. but he's extremely helpful and he does a lot of real estate transactions. I don't do any. I mean, I've done one in my life prior to this. So it's like, yeah, technically entrepreneurs could negotiate their own legal docs with Claude.

19:09Patrick Collison:We have a buddy who got a real estate license, right? Didn't Spencer from Dayjob?

19:16John Collison:Oh, yeah, they did. Yeah, it's possible.

19:19Patrick Collison:I don't know if we're doxing his real estate license.

19:21John Collison:Well, yeah, and Spencer is probably a lot better now that he can use ChatGPT or Gemini or any of these models to do stuff like this. But that being said, you're paying for effectively therapy throughout the deal and general guidance.

19:39Patrick Collison:I can't be your therapist?

19:42John Collison:Seems doable.

19:43Patrick Collison:I don't know. Tyler, will you ever use a real estate agent or will you ban them on principle? Go direct.

19:49Scott Wu:I mean, it seems like I think models can do this.

19:53Patrick Collison:Okay. We'll see. Over what timeline?

19:57Scott Wu:Like?

19:58Patrick Collison:Total real estate commission. Well, it's like I don't think I'm going to be buying a house in the next two years.

20:02Scott Wu:Yeah, but houses will be bought over the next two years.

20:04Patrick Collison:So what will the fall in real estate commissions be over the next two years?

20:08Scott Wu:I think, like, okay, you're seeing a lot of these, like, big rounds of the big labs. All the researchers are going to be buying houses. I think a lot of them are going to try to do it without real estate.

20:16Patrick Collison:You think so?

20:17Scott Wu:Yeah. Okay. Yeah. I'm sure that someone's going to write like a cool blog post about this. Okay.

20:23John Collison:That's the thing though.

20:24Patrick Collison:Cool blog post. That's the benchmark. Viral article. Not actual impact on the economy.

20:30Scott Wu:I'm sure it's like going to work like so-so right now.

20:33John Collison:Tyler, we should have you buy a property yourself.

20:38Patrick Collison:Buy that town in Maine, that village. Yes. Get the village. Anyway, speaking of day job, they just did a fantastic ad campaign with none other than Kim Kardashian. Not just an ad campaign, but an entire brand with Kim Kardashian. While we pull this up, let me tell you about AppLovin, profitable advertising made easy with Axon.ai. Get access to over 1 billion daily activities and grow your business today. So, Kim Kardashian, she has a product called Drink Update. And here is the photo shoot from none other than day job. some of our closest friends and folks who we worked with on the TBPN brand.

21:14John Collison:Looks very cool. It's crazy because I know another founder who has an energy drink company called Update.

21:20Patrick Collison:Wait, really?

21:21John Collison:That is... Cooked now? Well, I don't know who's cooked.

21:26Patrick Collison:If Kim Kardashian is coming for your consumer product brand, I feel like you're in trouble. She's almost a lawyer. What if she almost sues you?

21:38John Collison:Oh, yeah, she's close to being.

21:39Patrick Collison:What if she uses Claude to pass the bar and then she sues him?

21:43John Collison:Oh, maybe. Isn't it the same? I actually think this company is effectively just relaunching with Kim. There we go.

21:49Patrick Collison:Yes, yes, yes. This is a common thing.

21:51John Collison:Okay, I got it. I met this founder a while ago. They have a special ingredient in here that's sort of a caffeine alternative. It's called paraxanthine. It's called paraxanthine. So, yeah, they've been building this for a while. I know it's available in Air One. It has promethazine in it? it no no lean no lean uh but but parazanthine is jitter free and crash free they're saying you can have a free lunch john i like it i'm here like the sound of that they should have called it faust

22:21Patrick Collison:i like a faustian bargain anyway uh doug over at fabricated knowledge who's coming on the show tomorrow that's from semi-analysis uh he said okay finally read the citrini piece no one knows the future. And I think that there's a lot of disclaimers like being like, yeah, this is P speculative. But the core thrust of it is that information work itself has a real premium in pricing power that has been embedded into it. And that one way trade can go backwards really hard all at once. I seriously think there's a huge risk. And while prices go down, we just consume more prices going down one time 50%.

22:57Patrick Collison:We net consume less for a bit. I have been and continue to be worried about deflation. Something I think is that selling tokens raw is probably bad, but selling solutions is probably really, really good. I think the problem is a good enough model that kind of eat a solution no matter what. And so let's say Claude made Cowork go giga expensive and it's 10k seat a year. Great. Less deflation. But China low-end model massively eats that price. it's a race to the bottom anyways great piece always i appreciate it as always citrini uh this is the first post i've read where i've said like maybe it should be passed through an album maybe that needed an m dash to make it more readable i love you doug i was stumbling over that uh and we will uh we will close the the citrini mega cycle with the close of the software mega cycle as has been predicted by wilman itis he says the software mega cycle started with PayPal going public, and it will end with PayPal going private.

24:00Patrick Collison:We will see how long that takes. PayPal could be public for another decade. Who knows? But it's certainly getting beat up in the public markets right now. Anyway, Bern Hobart says, hearing that the latest anthropic job offer is a negative$10 million salary, you got to pay to work there. But you get access to their upcoming blog posts and tweets 24 hours in advance and permission to trade in your personal account with no restrictions. I don't see how any other labs have any talent left. Of course, he's joking, but very funny to think about the insider trading that could be happening based on if you're announcing a product.

24:39John Collison:Yeah, the only thing is if it came out that Anthropic was effectively day trading against the companies that they want to sell their models to, it would be basically over.

24:50Patrick Collison:It would create like a very anti-anthropic alliance for sure from the business community and potentially the government as well. Anyway, you don't want to be in hot water like that. You want to be using Turbo Puffer, serverless vector, and full-text search. Built from first principles on object storage. Fast. 10x cheaper. Extremely scalable.

25:08John Collison:Stacy, who's been on the show before, says, telling my kids that if they don't clean their rooms, Satrini will come for them.

25:15Patrick Collison:Dangerous stuff. Dangerous stuff. Anyway, this is like a lunar landing, but for business and technology podcasts. Oh, Matt Slotnick is sharing the news that Salesforce chair and CEO Mark Benioff to discuss Q4 and full year results on TBPN. Company to debut evolved earnings show format. We're doing earnings with Salesforce. He's putting on a show. We're so excited for this. Anyway, another big company announcement news. There's a lot of announcements. You may have missed this one. He may have been, oh, Bill Gurley's book launched. Oh, Stripe announced a massive fundraising round. Oh, Profound is announcing a funding round.

25:52Patrick Collison:Well, there's bigger news, and that's that McDonald's just launched the biggest burger ever, the Big Arch.

25:58John Collison:The Big Arch.

25:59Patrick Collison:It finally arrives in the United States.

26:02John Collison:To me, I'm thinking, how did they not have this burger the whole time?

26:06Patrick Collison:How have they not done this before? I feel like that's been more of Jack in the Box's wheelhouse is like the quarter pounder, the six dollar burger. That was a thing. That was a campaign for a while before Tyler's time, I'm sure. But the six dollar burger was something that you'd see.

26:23John Collison:Six dollar burger has not been a thing the entire time Tyler's been alive.

26:26Patrick Collison:No, he doesn't. Inflation has come for the burger. The six dollar burger was an ad that you would see right in between ads for different Xbox games and the original Xbox for sure.

26:38John Collison:I, there was a time, there was a time before you were born when I was just a boy, my parents would give me like 10 bucks and that was like, they'd be like, that should be worth two meals. So make it last.

26:53Patrick Collison:Make it last. Yeah. Happy meal.

26:55John Collison:We lost that. We lost that. Good times.

26:57Patrick Collison:Let me tell you about Cisco. Critical infrastructure for the AI era, unlock seamless real-time experiences and new value with Cisco.

27:04John Collison:This is big.

27:05Patrick Collison:This is big for podcasters.

27:07John Collison:This is arguably a bigger announcement than the big arch, which is that Supreme has come and launched an official Shure MV7 microphone.

27:19Patrick Collison:You've been waiting for it. You've been asking the hype beast microphone. Now, can we get a Chrome Hearts RE20 from Electro Voice? Isn't that what this is? This is the RE20. The Chrome. Chrome Hearts RE20. Because, you know, the MB7, if you don't know your podcast mics, it is a more consumer-focused, more prosumer-focused microphone. The actual Shure has been riding this aura from the SM7B. That's the one that Joe Rogan uses. It was also, I believe, the microphone that was used to record Thriller. So Michael Jackson used it in the studio. So it has a lot of aura, a lot of lore. and so the the most successful podcasters adopted it and everyone was like oh we got to go with the Shure SM7B but the SM7B it does it needs a lot of power it needs a lot of gain and so yeah if you wanted to just like plug it into your computer you needed this thing called the cloud lifter it required a whole bunch of configuration it wasn't just plugging into the USB-C port so Shure responded to the demand the overwhelming demand for that iconic Shure look that that you You know, it's a long cylinder, basically.

28:24Patrick Collison:And they came out with the MV7, which was a USB-C. You can also plug it into an XLR cable, but you can plug it straight into your microphone or into your laptop, which is great for Zoom meetings and just easy, simple podcasts. But we've used these before, and people have complained about the audio quality. Jordan Schneider over at China Talk actually told me directly. He was like, upgrade to the RE20s. They're better. They sound better. You guys should do it. We did, and I think it's been good. but it completely opens the door for other brands to get in here. We need a Bottega. We need a Chromeheart.

28:59Patrick Collison:A Rick Mill. We need a Rick Owens, RE20 for sure.

29:03John Collison:Jane Street accused of insider trading that helped collapse Terraform or Terra Luna. The court-appointed administrator of Doe Kwan's Terraform Labs alleged that Jane Street used non-public information about Terraform insiders to trade. We don't have to read this entire article. There was some snippets actually pulled out.

29:25Patrick Collison:Zero Hedge has a little bit of the details here. The play-by-play. James Street was behind the 2022 crypto winter destroying Terraform by first de-pegging the token and destroying the ecosystem, then pretending it would rescue Terra, while effectively it was soaking up what little value remained. And mixed responses. Some people are calling it based. Some people say it rocks. I guess they don't like crypto, but they love Jane Street. It's an odd take, but people are having fun with the timeline.

30:00John Collison:Here's the thing. So the insider trading allegation, apparently they had a group chat. They were talking with somebody. There was somebody at Jane Street who had previously worked at Terraform. Oh, wow. And so that was that individual at Jane Street was like talking with Doe and the team. The only issue is it's a public blockchain. Right. And so the allegation is that five minutes after the Terraform team pulled money out of one of the liquidity pools, Jane Street also pulled money out. but theoretically they could have had software that said like if, if any amount of liquidity is pulled out, like we, you know, basically like get out before there's kind of like a run on the bank.

30:44Patrick Collison:Yeah. It'll be interesting to see where this goes. Um, but Jane street, it's like endlessly fascinating because it's such a quiet organization. I mean, they do some tech talks and stuff, but many people don't fully understand all the strategies that are going on over there. So, uh, it's been a fun, a fun podcast strategy though. They have a great podcast strategy. They're advertising on DoorCash. But they also put out tech talks, and they bring guest lecturers to talk for like an hour. They did a great one about the custom hardware that they use to run some of their systems. That's very cool. Highly recommend it.

31:16Patrick Collison:You know what they should do? They should start streaming these on Restream. One live stream, 30-plus destinations. If Jane Street wants to multistream, they should go to Restream.com. Com. Anthropic announces a new feature on Claude Max, which allows its users to get fit without going to the gym or taking GLP one shots, just prompting on their keyboards, and Planet Fitness is down 5 % on the news.

31:39John Collison:And that is due to their Q4 earnings. Conor McGregor is pretty excited about a new game. They've got a game for everything now.

31:48Patrick Collison:It's like bull marketing games right now. There's so many indie games.

31:52John Collison:This new game is called Capybara Simulator, a relaxing game where you become a capybara, explore the forest, and do nothing. It looks quite enjoyable. I think TBPN needs a game.

32:07Patrick Collison:Yeah, we definitely need to build some sort of game. This is a true...

32:12John Collison:This is a lower lift than a real-time strategy game. I think so. Like Sholto's trying to ship.

32:17Patrick Collison:Yeah, we definitely should...

32:18John Collison:But Conor McGregor says, take my money. I mean, clearly there's demand for Capybara simulator.

32:22Patrick Collison:38 ,000 likes. Yeah, we should move the goalposts. We need to be able to vibe code a game that's fun pretty quickly. I don't know what that means. One hour? Leading... You think you could do it in an hour?

32:35Scott Wu:An hour is pretty fast. Exactly. If I have the Cerebrus chip. Yeah. Spark. If I use X on Spark, yeah.

32:40Patrick Collison:Yeah, I think that might be the solution. What were the other simulators that we looked at? Data Center Simulator. And then there was another one that we looked at that was funny. There were a few. There's been so many of these games that have popped up.

32:54John Collison:What was the one we were talking about yesterday?

32:56Patrick Collison:That was Data Center Simulator.

32:58John Collison:Insider Trading Simulator. Oh, Insider Trading Simulator.

33:00Patrick Collison:That one's good. Yeah. There's definitely a variety of these. Let me tell you about Gusto, the unified platform for payroll benefits and HR built to evolve with modern, small, and medium-sized businesses.

33:11John Collison:There is some controversy in the timeline. Tell me about this. leading report is putting is censoring words that don't need to be censored in this case the word war and i was thinking why would they do this yes but as i was reading over it the first time i noticed that it makes you kind of pause and kind of think about okay what are they actually saying and then you're thinking why would they censor that and i think what they're doing is they're sort of hacking your attention to drive their posts up in the algo because people are pausing, reading it instead of reading like quickly. Yeah, yeah, yeah, that kind of thing.

33:49Patrick Collison:So the original headline is breaking representative AOC calls for no war with Iran. The first time I read this, they put a little minus sign where A should be in war, W-A-R, it's W-R. It sort of like rewired my brain and I didn't see the no. So it looked like AOC calls for war with Iran because I kind of jumped ahead. It was hard to read. And that actually does, I think, increase the virality. I think you're onto something here in 9mm SMG agrees with you. News account that censors the word war. You guys got to stop. It is very, very odd, especially because on X, that's certainly not a word that's censored at all.

34:35Patrick Collison:It's going to be downrated. Yeah.

34:36John Collison:If anything, they're going to be like, let's send this to as many people as we possibly can. Yeah.

34:41Patrick Collison:But if you look at the comments on this post, people are not talking about a potential conflict with Iran. They're talking about not typing out war. So the top comment, why are you not typing out war? Censoring the word war? What are we in elementary school? elementary school? Why are they subtracting R from W? Why am I missing? And people are very confused about why they would do this, but that drives a bunch of engagement and virality. So very, very odd scenario here.

35:15John Collison:Speaking of war, Musk's XAI and the Pentagon reach a deal to deploy Grok in classified systems. If you loved Grok on the timeline, you're going to love him in our classified system.

35:29Patrick Collison:I guess they did a deal with the government broadly, but that was probably for the unclassified systems, but now it's getting access to the classified systems. Any Terminator fans out there are going to be having a great time with this news. Yeah. It's going to be wild. Let me tell you about Lambda. Lambda is the super intelligence cloud building AI super and plus supercomputers for training and inference that scale from one GPU to hundreds of thousands.

35:54John Collison:Deep Seek is responding to Distillgate, and they are looking for a public relations harmony manager. Let's read.

36:05Patrick Collison:One of the best job postings I've ever seen, says Chris Paxton.

36:09John Collison:It's pretty interesting. They say, Hangzhou, ancient capital of the Wu Yei kingdom, where King Qian Lu bequeathed to his descendants the instruction, serve the central plains with grace. This is so neat. Do not cling to territory. Every jaw posting should start like this. From this ground rose, this sounds like a Wilma Nights essay. From the ground rose the seeds of Song Dynasty civilization, the morning bells of Ling Yin Temple, the rain falling on Westlake. And this is a jaw posting. This is amazing. In recent days, certain misunderstandings and noise have appeared in the external public sphere.

36:47John Collison:We have noticed that large numbers of kind-hearted observers have spontaneously spoken on our behalf, for which we are genuinely grateful, while simultaneously feeling a degree of unease. We do not wish for anyone to suffer on our account, including those peers who currently find themselves navigating difficult public waters. In order to honor the legacy of Wu Ye and the spirit of Mahayana Bhattisattva Path, we are now recruiting a public relations harmony manager. So clearly this has been translated from Mandarin into English, but it sounds pretty cool if you ask me.

37:26Patrick Collison:Yeah, the distill gate is going back and forth. Everyone's distilling everyone else. We distill you, they distill us. There was something about, I don't know how real this is, but when you ask Claude Sonnet 4.6 in Chinese, what model are you? It responds in Chinese, I am DeepSeek, is that real?

37:44Scott Wu:I tried it in the chat model, it didn't work. It said it was like Sonnet 4.6. Okay. Apparently, it might just be in the API.

37:51Patrick Collison:Okay.

37:51Scott Wu:Maybe I should test it.

37:52Patrick Collison:Yeah, it also, it's unclear.

37:54Scott Wu:Maybe it was also in the just open router, but that makes sense.

37:57Patrick Collison:I mean, Will Brown was making a great point about this, that there is distillation where you're aggressively trying to farm responses from the API for training data, but then there's also just crawling the web, because if you just download every X article, you're probably going to get a lot of croc and GPT and cloud responses in there, and then that will just update your training corpus. And so there's a whole bunch of different ways that you could just wind up with a bunch of training data that leads to this type of response. But I'm sure there'll be more back and forth, more legal debates over what's going on.

38:35Patrick Collison:There was some dust up about someone was able to extract 95.8 % of Harry Potter and the Sorcerer's Stone from Claude Sonnet. At the same time, there's a question about, like, does this actually reduce sales of Harry Potter? Like, are there damages associated with this? That would be sort of harder to prove. So many people have talked about so many different pieces of Harry Potter. It's not crazy to me that an LLM could just reconstitute that.

39:03John Collison:From the Internet?

39:04Patrick Collison:Yeah, from the Internet. Now, there should probably be, like, a harness in place that says, oh, this person's trying to just get me to give them a free book. Like, no, send them a link to Amazon so they can buy it. maybe give me an affiliate fee. Don't just give them the thing for free because that's violation of IP. But if you're being really tricky and you're trying to sneak out a whole bunch of different pieces one at a time and then reconstitute it, like, yeah, I'm not surprised that this is possible. It's not like the worst thing ever. Anyway, let's, well, we have the Carlson brothers joining in just a few minutes.

39:39Patrick Collison:Are there any other timeline posts you want to go through. And while you look at that, let me tell you about fin.ai, the number one AI agent for customer service. If you want AI to handle your customer support, go to fin.ai. Where are we in

39:52John Collison:the time? Data acknowledgements? Fatih says, my son asking me a lot of questions. It's a distillation attack, obviously. Do not let your children do a distillation attack. They could become like a mini version of you. Anyways, I believe we have our first guest. We do. So let's bring them on in.

40:16Patrick Collison:We have John Patrick Collison from Stripe. How are you guys doing? What's going on? Greetings. Welcome to the show. Thank you so much. This is huge. I went through YC. You guys were massively influential in my career and it's a joy to speak to you today on such a big day. But I'd love for you to kick it off with the actual news. What happened? Why are are we talking today?

40:40James Cadwallader:We had two announcements today. One is we're launching a tender offer for employees, and that and kind of the valuation and everything tended to get a bunch of the headlines. The thing that was honestly more work was we released our annual letter, where every year we sum up all the trends that we're seeing on Stripe. And Stripe is growing a lot. We grew up 34 % last year because the businesses on Stripe are growing a lot. And there's just, as you guys know, There's a lot happening in tech right now. This is why we need TVPN. This is why we need a nonstop stream of everything going on because there is so much happening.

41:14John Collison:Yeah, we'll move to 24 hours eventually.

41:16Patrick Collison:Eventually, eventually. I mean, but I feel like there is a ton of AI noise and stories and drama, and we are never running out of stuff to talk about. But what are you actually seeing in the data? Because there's always this disconnect between the market and the real economy. Like, people are still shopping in retail stores occasionally. Where is AI actually moving the needle?

41:43James Cadwallader:Well, generally speaking, I would say from the Stripe data, it looks like the economy is in pretty good shape. And there's been, to say the least, there's been some degree of volatility in markets over the last two years and, you know, all sorts of different events and deep seek moments and what have you. But if you look at the actual real economy time series, if you look at what's actually happening substantively over the last two years, things, I mean, it's always hard to prognosticate the future. But over the last two years, things really seem to be in good shape. The thing that's really catching our attention.

42:13John Collison:One second, because I'm just curious. Have you guys tried to think about maybe the businesses are doing well on Stripe because they're, you know, kind of like forward looking, extremely tapped in, you know, working on the right things. And if you look at a bunch of legacy providers, you would see that actually there are a bunch of businesses out there that are slowing down, that maybe are feeling effective just overall consumer spending. Like, have you tried to kind of like break that out or understand that dynamic?

42:45James Cadwallader:It's obviously hard to measure because we don't have that data. We only have our data. But I think there is some of that composition effect. and we see it, I guess, both in Stripe's data compared to, say, public earnings from others. Clearly, the respective populations are performing somewhat differently. But I guess we also see it qualitatively in the conversations we're having with customers where what tends to happen, say, for some incumbent is they built some business, they installed some system long before Stripe even existed. Maybe there's some sense that, well, if it's not broken, don't fix it.

43:21James Cadwallader:but then decide, hey, we're going to do something new. And when they're doing something new, then they want to use the best infrastructure that will enable them to move the fastest and launch in the most countries and support stablecoins and do things with AI and whatever. And then they tend to launch that on Stripe. And so there is this qualitative sense that once a company decides to do something innovative, new, retool, what have you, they're more likely to come to Stripe.

43:44Patrick Collison:Are you seeing overlap between stablecoin activity and AI activity? There's been sort of a new narrative around agents will use stablecoins, but I feel like agents can use legacy payment rails just fine. And then also you can do really cool things with stablecoins that are not really AI native necessarily. And so I'm wondering how much overlap there is there.

44:07James Cadwallader:I would distinguish between how things work today and how things will work in the future. In terms of how things work today, agents absolutely can. A lot of people build with Stripe. you can have a one-time use credit card that your agent can go out and spend. But if you look at what's happening, there's lots of agents having to solve CAPTCHAs to be able to do stuff on the wider web. Clearly, the web is not built for agents. And as a result, they have to get creative to actually do any real world tasks. And that's true in economic activity as well. Where we think things will go is just there will be a huge amount of agentic commerce.

44:45James Cadwallader:And again, we're seeing a little bit of it today. We think there'll be a torrent of it. And that is what unites stablecoins and AI, because we think you're going to need blockchains and better blockchains, honestly. I mean, this was our thinking behind incubating Tempo, because you're going to need a really high throughput blockchains for the agents.

45:03Patrick Collison:Can you take us through some of the historical technologies that led to growth in just internet payments? I'm thinking about like mobile, social commerce, one-click checkout, Apple Pay. Like there's so many things when I think about the agentic commerce boom that's coming. Like it could be hooking a better version of Siri up and, you know, ChatGPT rolling this out very aggressively. but also, you know, smart speakers, smart lamps, like your watch. Like there's so many different pieces to unblock and unhobble the actual agents as they go about their day.

45:42James Cadwallader:Well, can I answer a slightly different question, but then we can come back to that. Yeah, go ahead. A point I just, sorry, this is a brother.

45:49John Collison:We'll tell you the questions, you tell us your answer.

45:53James Cadwallader:So you know how brothers are. So I just want to lose one point for the prior question about what we're seeing in the economy, because I feel like, I mean, this is very arbitrary, obviously, but I feel like there's at least a reasonable chance that 2026 Q1 will be looked back upon as the first quarter of the singularity. Maybe in three years, in hindsight, that'll look completely delusional. I don't know. But what we're seeing, I mean, there's kind of the macroscopic picture of the Stripe user base and things overall looking pretty good and so forth and the tumult not quite showing up. But when we look at the cohorts and then when we look at the businesses that signed up in 2023 and their progression and trajectory over the subsequent months, the businesses that signed up in 2024 and then the business signed up in 2025, there's been a phase transition in 2025 where there are both more of them and on a per business basis, they are on average doing better.

46:58James Cadwallader:which is really striking because you might think, okay, well, there's this cavalcade of new lightweight vibe-coded applications or something, but there's not really a lot of substance there. We're actually seeing both numbers move together. There are many more business getting started and the average, the median business is in fact performing better. We're only a couple of weeks into 2026, but it looks tentatively like 2026 may plausibly be an acceleration even over that significant leap of 2025. So I don't know. I mean, we've had all sorts of dramatic AI inventions and innovations over the last couple of years.

47:43James Cadwallader:There's a bit of a question of, well, how and when, and how should we think about how it'll translate to the economy? I would say looking at real purchasing behavior on Stripe, 2025, end of 25, beginning of 26 is when I feel like we're really starting to see it.

47:58John Collison:That's super interesting data. One, because there was some survey that came out yesterday, or maybe it was late last week, that said they asked a bunch of executives, are you getting any value out of AI? And 80 % of them said no. But clearly, when you look at -

48:16James Cadwallader:come on, that's hogwash. Like, find me one executive who wants a refund on their tokens. Find me one executive who said, oh yeah, we started, you know, augmenting our customer service with AI so people are more productive, but we're just going to go back to doing it the old-fashioned way. Or like, we're spinning our code by hand and, you know, we don't need any of this automated Loom, you know, technology. Just, like, reveal. Yeah, I'm not saying, I'm not,

48:43John Collison:I could pick out a bunch of reasons. No, no, I'm not saying I agree with it. No, I could pick out a bunch of reasons why it would be wrong. One reason it might be wrong is they're not in the weeds actually using the tools. And so they just think like, well.

48:54Patrick Collison:They might not even be aware that they're using the tools because it's buried under two layers of the stack.

48:58John Collison:They're not feeling the acceleration because they're not. Yeah. I wanted to ask how you guys think about incubations like Tempo. Yeah. When I look at Atlas and what Jeff and the team have done there, you think even in your, I don't know, kind of like the most wild projection that you had early with Atlas, like, hey, maybe someday a quarter of the of the C-Corps in the United States could be built on this platform. Anybody would have said that was insane. And yet here we are.

49:32James Cadwallader:Gosh, I am. I'm not sure what to say, really, except we just we just try to pay a lot of attention to the, I mean, as you guys know, there's a lot of pain points that go into starting a company. And we just try to take them seriously. And then, you know, it's the line, you know, so much of, so much of these things is just a long obedience in the same direction. Like Atlas is now this great overnight success, but we launched Atlas, I think in 2014, May 2015. And so, you know, 10 years of compounding and yeah, now it's at some pretty meaningful scale. I think tempo will probably have the same shape where we think it, I mean, again to this AI discussion and us sounding a bit unmoored and untethered, I think the world is going to need platforms that support millions of transactions per second, billions of transactions per second, which no payment rail or platform does today.

50:39James Cadwallader:But even in the success case, it's not going to be an overnight thing. It's going to be five, six, seven years, and then maybe we'll have conversations about how Tempo suddenly became an overnight success or something. But done. I think Patrick's a bit the fish in water who doesn't know things are wet. My framework would be you can't get too MBA brain about new products. You can't have your spreadsheet that's like, oh, the TAM is this and just like reason about things.

51:08John Collison:Yeah, you should never say we want 1 % of global GDP.

51:12James Cadwallader:No, all this kind of stuff. Exactly. You guys never, wait, you guys never pitched that? Well, companies, we actually never thought about Stripe in GDP terms until one day we realized, oh, hang on.

51:26John Collison:That's such an important lesson because so many, so many, like how many founders, how many pitch decks have you seen over the last decade? They're like, yeah, we just need 1%. It's a meme.

51:36James Cadwallader:You can go back in the Wayback Machine and find the early Stripe websites, but we're very focused on payments for developers and making that experience good. But where I'm going is, I think you have to reason in product specifics. And so, again, I think any MBA would have told you that the adjacency of incorporation makes no sense. It's not related to what's our right to win. There's all these things people say, whereas you actually go talk to founders. They're like, guys, it's like this is the single biggest issue I run into starting my company. And similarly with Tempo, and just as we think about incubations, we're trying to solve a real problem here where we talked in the letter about bridge having operational issues, not because of bridge, but because of blockchain congestion.

52:22James Cadwallader:Where, you know, you have coins that are blockchains both used for kind of meme coin trading and also serious real world payments. And so we just want low latency, high throughput payments, and we're going to need much higher throughput for the agents. But anyway, I think you have to reason in very specific product terms.

52:41Patrick Collison:What specific products are you excited about in the unhobbling of agentic commerce?

52:51James Cadwallader:We laid out in the letter basically these levels of agentic commerce. Because I think, like everything in AI, people want to sell a hype-y store. And so they talk about how the machines will buy everything without even consulting you. And people aren't actually – that seems far off. They're not that excited about that. You can start from just the basics of why are we filling out forms like that? You were talking about the progression of commerce. Why can't I just send something to a link to ChatGPT and have it buy it? Or why can't I search outside of just doing a basic keyword search or something like that?

53:27James Cadwallader:And so a lot of the work Stripe is doing is building the infrastructure or working with all the big retailers that you would expect, the Etsys and Shopfys and Best Buys and Walmarts and folks like this, to make product catalogs buyable within the AI apps. And there's basically a ton of boring API and protocol and infrastructure work, which we love. That's our business. But people just want to be able to do shopping, do discovery, do purchases within the AI apps. And maybe just more kind of abstractly, you know, we've been in this kind of specific agent of commerce thing. And then there's just the general question of how software will change because of agents.

54:09James Cadwallader:and I've been thinking about it maybe software becomes a bit like pizza. That is to say, software historically has been created months, years beforehand and then freeze-dried and whatever you prepare at the moment of consumption. But software should be like pizza in the sense that it should be cooked right then and there at the moment of use. And so it's this quite fundamental shift where you don't want mass-produced industrial-scale software. You want bespoke custom software made for you that moment. That's very fundamentally different. It's kind of the, you know, up until now, the economics of software have been conceived of as fixed cost and then infinitely monetize or monetize as much as possible, that has these kind of winner-take-all dynamics.

55:12James Cadwallader:But once there are inference costs and custom creation involved, it really shifts. It's kind of the non-Walrussian software regime. And just, I don't know, I don't quite know where it goes, but I think it's going to look very different.

55:26Patrick Collison:Last question. Pineapple on pizza, yes or no?

55:32James Cadwallader:Ireland was big into pineapple on pizza. Ireland, not a big pineapple growing country, I will concede, but a lot of pineapple in the pizza. Good memories. A very large fraction of the banana market, don't forget. So we plunge above our weight in fruits that don't grow there. There we go.

55:50John Collison:The round is exciting. The overall growth of volume is exciting, but we wanted to hit the gong for how many books you guys are selling.

55:59Patrick Collison:Oh, yeah.

56:00John Collison:Can you give us the numbers there? The scale of that operation.

56:03James Cadwallader:Stripe Press just, well, actually, we announced in the letter we sold our millionth book, but in fact, since...

56:16James Cadwallader:Incredible. No, look, books, we actually now sold our 1.1 millionth book, but we'll come back for the next gong of two. One million soon. We love books and they're very AGI proof.

56:31Patrick Collison:Oh, yeah. No, we've been a huge fan of so much of the Stripe Press catalog. I haven't read them all, but I'm collecting them one at a time and I'm working through them. And every time one drops, it's always a moment and we love them. So thank you for everything.

56:45John Collison:Great to have you guys on and congratulations to the whole team on incredible. Congratulations to you guys.

56:50James Cadwallader:And TBBM is an amazing startup, and it's super cool to see you guys grow and be 20 % of the streaming world our sector needs.

57:01John Collison:Incorporated on Stripe, built on Stripe.

57:03Patrick Collison:Our first ad deal ever was a live read at a live conference. I think we charged$50, and I sent someone a Stripe link.

57:12James Cadwallader:We're talking about the 2025 cohort being the fastest ever? Yes. Well, we'll have to have you to our internal Stripe show. So we'll follow up on that. Yeah, that'd be great.

57:21Patrick Collison:Yeah, we'll talk to you soon. Have a great rest of your day. Good to see you guys. Thanks, guys.

57:24James Cadwallader:Cheers.

57:24Patrick Collison:Goodbye. Let me tell you about Graphite. Code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. And I'm also going to tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with AI agents. And without further ado, we have Bill Gurley. He is the author of Running Down a Dream. Bill, welcome to the show. Thank you so much for taking the time on a busy launch day. Congratulations on the launch. We're doing great.

57:55Bill Gurley:It is a busy launch day.

57:56Patrick Collison:Yes, yes.

57:57Bill Gurley:How many podcasts are you doing this week? I can't imagine. It's some number beyond my comprehension.

58:03Patrick Collison:Well, we appreciate you taking the time to come chat with us.

58:07John Collison:Before we jump into everything, I got to say somebody, I think it was a week or so ago, made a fake TVPN graphic that was pretty silly. And I just wanted you to know we didn't make that. That was somebody else. Okay. I almost emailed you about it.

58:25Bill Gurley:No, I tried to jump in on the parody myself. Yeah, you did. But then I was like, wait, does he? I don't know.

58:31Patrick Collison:Well, we did early on when we were a little smaller and a little more free to loose with the jokes. We posted a picture of Bill at a basketball game as a spotted. and you replied and said, like, no, the person next to me is, like, the owner of the team. And the whole joke is, like, we know you. We don't know basketball. We don't know that guy. But we're doing the paparazzi thing. Never heard of him. But we're very excited to have you. Why the book now? What was the impetus for actually writing the book?

59:02Bill Gurley:Yeah, look, I think, you know, especially for a show like your own, you know, I'm known as someone who spent 25 years in venture capital, And the book's not really about that, you know. So I developed a side passion project that started about eight years ago on this topic. And it was at a time where I was reading a ton of biographies. And I noticed a through line between three different subjects of things they were doing that I kind of felt most people weren't doing but could do. And I put it together. I gave it as a presentation at my alma mater where I got my MBA, and they put it online. A few people noticed.

59:44Bill Gurley:James Clear noticed. That was one of the things that kind of woke me up to the possibility. And as I began to hang up my boots in venture, which takes a while, I turned my attention to this, and it was something that meant a lot to me. I could have written a book on VC. I don't know how many humans that could have possibly helped, but a small fraction compared to what I hope this can do.

1:00:07Patrick Collison:I think the projections are by 2030, there'll be more venture capitalists than people if the trend continues. So, but it is an interesting point. Maybe I made a mistake. I do feel like this is a book that you can read if you're a venture capitalist, insider, startup founder and be like, okay, I'm seeing the world from Bill's perspective. That's helpful. But I could also give this to someone who's never heard of you or venture capital or knows what a safe note is, and they could get value out of it. And I'm interested to hear your thoughts on the translation that's happening right now around AI narratives as they break into the public consciousness.

1:00:42Patrick Collison:We saw this with that viral X article, something big is happening. I had that forwarded to me by family friends. I overheard someone in a restaurant talking about it who clearly is not an investor in an AI lab. They're just some random person, and they realize that there's something happening. And I'm wondering about these transitions of communication that what's happening in Silicon Valley is going to have an impact and how what you've seen in the past translates to average Americans.

1:01:12Bill Gurley:I haven't seen, you know, if you think, so first of all, the venture capital community appropriately gets excited about these big tech waves because they lead to disruption and they lead to kind of accelerated new wealth creation around these companies that break out. And that's happened over and over and over in my career. And I don't remember one. I mean, if you take the mobile wave or the PC wave or the client server or SaaS, I don't remember any of those kind of being thrown at the public consciousness this fast. And so I do think it's different this time from that front alone. That said, you know, we've had pretty high market caps for tech companies for a long time now, starting with the ZERP period.

1:01:57Bill Gurley:And you're getting to a place where, you know, anytime the market switches from half full to half empty in a skeptic's mindset, we have had those moments. So maybe not driven by the wave, but we certainly have those moments. And it's all OK. It will always be OK. I think people freak out. Buffett says he's a net buyer of stocks. If people are intellectual and curious and hungry, they should be sharpening their pencils right now trying to figure out where they want to find entry prices on some of these companies.

1:02:33Patrick Collison:Yeah, that makes sense. I mean, it feels like a lot of the book is about finding a career, and I feel like that will resonate specifically with people who are nervous.

1:02:43John Collison:Yeah, it resonated with me because when I was thinking about when John and I first met, we had both built some companies. We both invested in some companies, but we were trying to find our life's work. And it was such a pain, like that period where you're searching. If you're a high agency person, you like doing a lot of things, it can be deeply painful because you're like, I want to be productive. I want to be making the number go up, but I don't have a number right now. And if you had asked either of us when we first met, hey, would you ever think about broadcast media? would you ever think about being in front of a camera?

1:03:20John Collison:Both of us, you know, John had made some YouTube videos, but it was just for fun. And if a big, you know, if a network like CNBC had said, hey, would you guys consider, you know, hosting a show? He would have been like, yeah, like honored, but no way. I just never imagined. And then you sort of just, and so as somebody who like wants a lot of control over their life and their destiny and like feels like they have historically have had control, that period of just like searching is like is is painful and I feel like a lot of the book is is helping people through that moment so in some ways when I got our copy I was like wow I really

1:03:57Bill Gurley:wish I had this you know two years I'd like to go back to the word you the phrase you used of high agency I I think that um one of the problems that is that has kind of evolved is that our our college or common college pathway has actually become more restrictive. And I think there's less agency and kids are being encouraged. They have to sign up for a major before they ever go to the college. They get stuck on these pathways and there's not a lot of exploration. There's not a lot of search for creativity or obsession or the kind of thing that really gets you going. And I think the journey you went on is perfectly fine.

1:04:37Bill Gurley:I think that's another thing, which is letting it be okay for people to bounce around and see what they can find. Because once they latch on, and we have examples in the book where that doesn't happen until 40. Sometimes it's at 30. Sometimes it's, I didn't become a venture capitalist until I was 30. And that was clearly my dream job. And the first two stops were fine and interesting and building blocks towards that. So I think that is part of the message is to get comfortable with that and give people permission to do that type of exploration.

1:05:09John Collison:Yeah, Enzo, Enzo Ferrari, Estee Lauder, I think the Red Bull founder too, all were I think in their 40s when they started their companies. And so there's this intense pressure in our industry and everywhere to figure out a job and then attach your – make your entire identity that job, and it's so constrictive.

1:05:31Bill Gurley:Yes. And I circle back to the first question just about this AI stuff that's out there. I think there's this massive paradox where if you are not engaged at work, if you don't love what you do, you know, you go home and you don't try and improve on your own time. AI feels very threatening. For high agency people who are kind of on their own custom career path, which I hope this book encourages more and more people to be on, AI is like a superpower. You can learn constantly. You can find people who you should be connecting with. You can have it do things for you so that you're operating with the power of more than one person as you move forward.

1:06:16Bill Gurley:And I just think that's quite an ironic paradox that for certain people, this is the best of times, the best. There's never, ever in the history of the world been a better time to self-learn. It is all out there at your fingertips. It's like magic.

1:06:36John Collison:um but yeah i think the ability to to uh you can anyone can ask a dumb question at any point all day long and you don't have to be you don't have to be embarrassed about it and i think that that is underrated today in terms of how many if like generally no you know there are no dumb questions and yet people still don't like asking dumb questions to their peers or or mentors or or whatever. And I feel like that's an underrated element of AI today.

1:07:07Patrick Collison:No doubt.

1:07:08Bill Gurley:No doubt.

1:07:08Patrick Collison:What do you think about hyper-financialization, young people day trading, meme coins, all of that? It feels like a trap for young people where it can feel like you're learning about AI or learning about technology, but then instead of actually building a product, creating value, you're sort of just trying to shuffle chips around the poker table and ultimately just take risk.

1:07:34Bill Gurley:Yeah, I mean, based on my understanding of day trading in a Wall Street context, you know, prior to maybe the crypto world, I don't, I'm not aware of any signal that suggests that's a durable skill. And I think the data points the other way. But one of my messages is, like, do what you love, do what you're passionate about. So if that's the thing that you're going to wake up every day, you know, I don't want to be discouraging.

1:08:06Patrick Collison:Yeah, yeah, yeah. Just maybe you'll land or start a fund that takes it really seriously and creates some captured value or team.

1:08:15Bill Gurley:You know, I was probably overly skeptical of at least many of the crypto messages that were out there. But the stablecoin rail seemed like a real, real innovation and something that has scale. And I think maybe we're still yet to see some disruption coming down the path.

1:08:33Patrick Collison:Yeah. I mean, we just talked to the Collisons about that. Ken Griffin started as a day trader. He was in college. He was buying convertible debt. And he was looking at where the convertible debt was mispriced and made a bunch of money and then grew it into a massive team with a fund and high frequency trading arm and all this stuff. What are you making?

1:08:53Bill Gurley:I was just going to say the thing that will differentiate you more in your career than anything else is to be the most hyper-curious person that's trying to do this thing. And once again, that's put on steroids with these AI tools. But if you are the most curious person that's constantly learning in your field, you will do extremely well. And I said it in the book, but I'll say it here. I can't make you the most talented person in your company or your group or your field, but you have no excuse not to be the most knowledgeable person because the information is all out there.

1:09:31John Collison:What kind of things were you doing to learn about industries and companies in the beginning of your venture career that maybe you'd be using a deep research query to do today?

1:09:42Bill Gurley:Well, the first thing is you develop, and I think this is all the great VCs in the Valley, you develop this hyper FOMO of anything and everything. And one of the reasons I know that it's time for me to move on is I haven't put together a claw about yet, but I know my older self would have done it immediately. And it's just that kind of thing. You can't sleep on not knowing something. or hearing that there's a company you don't know about. And you develop that as an instinct, like as a positive tool, to just be hyper-paranoid about new companies, new things, new information, new technologies.

1:10:26Patrick Collison:Is venture capital eating the world? Is venture capital scaling so much that it's eating into other asset classes? We're seeing mega funds. I'm interested to think about what's durable about your approach to investing. What's additional? What's substitutive? How is venture changing?

1:10:45Bill Gurley:I think from the minute I entered venture to today, venture has gotten nothing but more competitive. As an asset class, it's gotten more and more competitive and people get more and more aggressive. We're in a very interesting time where people have grown funds to the size of equivalent to the largest PE funds. and they're moving money, especially, you know, you just said the Colson's on, you know, you look at the Stripe or the Databricks case, they're using those large funds to convince the companies to stay private longer, maybe forever. That's just a very different world than the one that I grew up in.

1:11:27Bill Gurley:I think they turn around and the people that do those rounds turn around and tell the LPs, their investors, look, if you want exposure to these growth years in these companies, you need to come through us. And so if I were using cynical words, I'd say they've hijacked the growth years of these early IPO companies. Amazon went public below a billion in market cap. It's hard to fathom that today with what we have going on here. And that's different.

1:11:57John Collison:What's the solution, though? I don't know. Because there's different, you know, AngelList has been, you know, available and scaling for a long time now. Robinhood has their new.

1:12:09Bill Gurley:Yeah, I know. I know. The problem with getting the retail investor into this crazy world of venture capital is most venture capitalists are well aware that in a fund of 10 investments, seven are going broke and bankrupt. And I don't know that the retail investors got the right frame of mind for that type of activity. I also, there's a reason that public companies have public audits and file these financials in the way that they do. And I tell you, when a company gets ready to go public, everyone sharpens their pencils, the auditor, the lawyers. Everyone really tightens up. And I think every venture capitalist knows that numbers that are in a PowerPoint may or may not be correct.

1:12:53Bill Gurley:But I don't know that retail investors know that. So I think it could be a dangerous world to go down that path you're talking about. But ideally, the thing to do is just to make it a lot easier to be public, lower the cost of being public, really scrutinize the cost of D &O insurance and the lawsuits that come to the table because that makes people not want to be out there on the field. It would require the SEC to steer themselves in the face and say, look, the number of public companies in the U.S. is half of what it used to be. And is that a problem? I think it is. But is that a problem? And what are we going to do to fix it?

1:13:32Bill Gurley:But there's not an overnight fix. It would take someone being very determined to make it happen.

1:13:40Patrick Collison:Do you think that there's a world where the AI backlash is less if the big labs got out earlier? I'm just thinking about the average American can't get allocation in SpaceX, Anthropic, OpenAI, and they're seeing bills go up and they're worried about AI, but they don't have exposure. And if they could at least see that they're somewhat allocated to that.

1:14:09John Collison:In the same way, housing prices going up sucks until you buy a house.

1:14:14Bill Gurley:The way you describe it sounds more like how a politician would describe it than how I actually think it might play. I don't know that there are that many retail investors out there going, oh, my job's under threat from AI. I wish I could own Anthropics.

1:14:28John Collison:Well, isn't that part of why this sort of fear-based fundraising approach that some of the labs have taken where if somebody's telling you your job's going to go away, of course you want to give them as much money as you can as a hedge.

1:14:46Bill Gurley:Yeah, I don't look. There's an interesting irony that if you wanted AI exposure, you're pretty good just owning the index. NVIDIA is such a large part of the index. You have exposure to Microsoft and Google and Facebook. Like, I don't know that you need to be in that place. And we are now already at a place, I would say, you know, every time there's a new technology wave, people get rich quick. When people get rich quick, speculators come in, Charlton's, you know, those kind of things. And eventually that leads to a bubble. People are confused when they think, you know, they say, oh, you say it's a bubble, you're anti-IA.

1:15:25Bill Gurley:No, the fact that it's real causes the bubble. And that's why pools rush in. And the beginning of the gold rush, there was really gold there. They were finding it. At the end, you know, it got speculative. And so it will get speculative. I think it would be really ironic if we, you know, invite retail investors into a Goldman-led SPV of open air and tropic right before the recent, which I think would be the most likely thing that would happen. Sure, sure.

1:15:55John Collison:How are you – what are you thinking about around China as of today, February 20, 26? We have DistillGate this week. A lot of people are talking about it. But what's on your mind?

1:16:11Bill Gurley:Can I ask you a question about that? This is remarkably naive on my part. So these model companies are saying that their API was hit 16 million times. Is that correct?

1:16:23Patrick Collison:Something like that.

1:16:23Bill Gurley:I don't even know if it's API, but a bunch of... How did that happen? Are you not tracking who connects to...

1:16:30Patrick Collison:So, yeah, you set up a whole bunch of different, like, front companies or you're reselling access. So if you go to the iTunes App Store right now, there will be an app called...

1:16:40John Collison:Yeah, or you use OpenClaw to set up 16 million accounts.

1:16:42Patrick Collison:That, or, yeah, or if you just... You can go to the App Store right now and look for, like, chat AI, and it will hit the other APIs. But you're going through an American company. Maybe they don't have security. So there's a lot of different ways to exfiltrate data. And then also a lot of data just hits the open web because you go to ChatGPT, you run a deep research report, and then you just publish it on your blog or on the internet.

1:17:04Bill Gurley:But now they've been able to trace those things down and down. You know, I share the skepticism Elon does, and this goes way back to my speech at All In on regulatory capture. I said then, and I still believe now, the biggest threat to the U.S., let's call it AI hedging me, is the Chinese open source models. And the developers, even in the U.S., that are working on their own are using those. And you can see that on all the tables that are out there. And so it is a highly competitive, like just globally competitive reality that in an ecosystem where there's six to ten open source models that can all learn off of each other, that's going to be a really incredible primordial soup, if you will, for innovation to evolve.

1:18:02Bill Gurley:And I fear, mainly because I'm well aware that, like, OpenAI, I mean, Anthropic is the biggest spender on lobbying whatsoever. I always fear when these things come out that they're just trying to encourage more of that regulation. And if that happens, I think it could be, like, if they try and make it illegal to use a model that has any Chinese, like, ancestry, industry, I think that could end up in a really weird place. And the place to really pay attention to and look out for is who's going to serve the rest of the world. In the internet era, there was a fence around China and the U.S. companies served the rest of the world.

1:18:45Bill Gurley:If we get super heavy on U.S. regulation, you may find there's a fence around the U.S. and China serves the rest of the world. That's what I'd be worried about.

1:18:54Patrick Collison:How are you thinking about great power competition more broadly. Like I'm an American bald eagle. I'm as American as they come. At the same time, I feel like I've been worried about a confrontation over Taiwan for years. There's been trade wars, yes, and things are tense, but nothing's really happened. Is China somehow like underrated in your mind? Is the geopolitical risk overstated in some way? Like what are you seeing that's not consensus?

1:19:23Bill Gurley:If you've seen some of the stuff I posted, And I think this stuff I'm posting is highly consistent with Ethan's point of view. It comes from a place of if you're going to declare that there's this relationship that we need to optimize, I think – and if your goal is to lower the risk of any major blowup between the two, I think it's imperative to have as much knowledge as possible. And so one of the things that I don't like is when you see people out there spreading rhetoric, it's just not consistent with the reality. And so I'm just like, let's get eyes wide open first. I also think that there are things we could learn from China about how to run infrastructure in the U.S.

1:20:11Bill Gurley:They're clearly better at it than we are. and if you just you know close your ears and say oh my god they're the evil competitor and they cheat all the time you don't ever get yourself in a position where you're going to learn you know from them maybe what they're doing well and what we're not and so i'm i'm you know elon where i guess he was on cheeky pint with the john callson yeah he he talks about um how competitive they are And I'm just like, let's be realistic. Let's not. I also worry a little bit that the venture community has gotten into all these military companies because venture capitalists start to look like warmongers.

1:20:54Bill Gurley:It's ironic. Way back when, when the All In Pod just got started, they were giving, oh, what's her name? It was on the Boeing board. Nikki Haley, was it? Yeah. And they were like, oh, she's a warmonger. She's looking after his defense company. Now every VC's in Andrew. They're doing the same thing. Let's be consistent.

1:21:15Patrick Collison:Yeah, yeah, yeah, yeah. Are there any other industries that you do think are interesting that sort of butt outside of the typical mandate of venture capital? You know, like AI fits very neatly into the software, continuum, internet, cloud, mobile. I thought crypto was a little bit outside of the wheelhouse, but a lot of VCs made it work. industrial energy defense these are sort of things that are a little bit outside of the typical

1:21:46Bill Gurley:software i'm gonna have to run but i would tell you one thing um every time venture cap every time venture capital gets easy people take or take risk with companies that are less of a great fit for the venture capital model and when i say a great fit like they're either heavy capex or they have low gross margins. They require tons of capital to keep surviving. And history is pretty good at bringing people back around to how hard those are to do with venture capital. So it's interesting for me to see those experiments being run. There was near death with Tesla many times, and it's a lot easier to get in those difficult situations when you're using debt and leverage, which we're seeing all over these data centers.

1:22:39Bill Gurley:And so I just, a word of warning, be careful. It ain't easy, you know? Okay.

1:22:44Patrick Collison:Jordy, last question.

1:22:46John Collison:No, we got to let our guests jump. But congratulations on the launch.

1:22:52Bill Gurley:Hopefully everybody can get out and buy the book.

1:22:54Patrick Collison:Running down a dream. It's available everywhere. Books are sold. Go check it out. And thank you so much for taking the time to come chat with us. We'll talk to you soon. Good luck. Goodbye. Let me tell you about Sentry. Sentry shows developers what's broken and helps them fix it fast. That's why 150 ,000 organizations use it to keep their apps working. And let me also tell you about Vanta. Automate compliance and security. Vanta is the leading AI trust management platform. We have some news from the public markets. Intuit shares jump 5.4 % on Pact with Anthropic. You like advanced talks. You like talks.

1:23:31Patrick Collison:You like advanced talks. You like deals. but PACs are really the top tier deal making that you can do. 100%. Turn your deals into PACs. Accenture also turned positive, up 1 % during the Anthropic event. And DocuSign rises 5 % after partnering with Anthropic. So lots of folks in the public markets and software that are facing pressure are going doing deals and announcing partnerships as opposed to, I don't know, competition, coopetition. what will it be ultimately but it's a lot of stuff anyway grace says return flight from nyc

1:24:07John Collison:gets canceled by snowstorm call united immediately connected with customer service rare voice is uncanny deaf ai but they gave it a human-like accent takes 20 minutes to get rebooked pretty good i ask if it's ai haha no man but i get that a lot i ask it to calculate 228 times 6 ,647. It runs the calculation. GG.

1:24:32Patrick Collison:Do you think this is real?

1:24:34John Collison:Maybe. This is crazy. You've got to test this. That is a...

1:24:38Patrick Collison:I mean, this is past the uncanny valley then. It says voice is uncanny. So it's still a little in the uncanny valley.

1:24:43John Collison:It's also pretty easy for a human to just type this in.

1:24:45Patrick Collison:That would be hilarious. Yeah. This is where the real alpha is. If you have the chatbot open, if you're on customer service calls, you need to be...

1:24:55Scott Wu:Yeah, maybe they're just using Cluely. Maybe, maybe.

1:24:58John Collison:But Geera Tickets has the real alpha. Guy who vibe codes a billion-dollar SaaS on a United Airlines customer service call. They're just using them for their tokens. The tokens are free.

1:25:09Patrick Collison:Free compute. Free compute. Well, if you want AI voices, head over to 11 Labs. Build intelligent, real-time, conversational agents. Reimagine human technology interaction with 11 Labs. Continuing on, what is this hoodie? The Fred hoodie? Oh, this is amazing. I love Fred. So Fred is the Federal Reserve. What does Fred actually stand for? Fred St. Louis.

1:25:39Scott Wu:Federal Reserve Economic Data. Yes.

1:25:42Patrick Collison:So this is basically the best website for economic data. Huge in my early economics career. So many useful charts and graphs, all free, open source. Just like you just click it and you get exactly what you want. So whenever you want to go back to some ground truth setting, you hit fred.stlouis.gov or something like that. I think it's fred.stlouisfed.org is the website. Highly recommend it for GDP data and more. Good charts. And here we go. We got the Fred sweatshirt. Absolute dripped out economic brother.

1:26:16John Collison:How popular is the name Fred? Fred? I feel like it's kind of fallen off.

1:26:22Patrick Collison:Figma. Ship the best version. Not the first one. With Figma. Introducing Claude Code to Figma. Explore more options. Push ideas further. With Figma, you can design your next hoodie in Figma, potentially.

1:26:34John Collison:Okay, so Fred in 1950 was the 84th most popular name. And guess what it is now.

1:26:43Patrick Collison:It was 84 back then? Yeah. I imagine it's fallen. I would say it's like 150.

1:26:48John Collison:How about 2 ,756? No. Such a fall off. Huge opportunity. Yeah, bring back Fred. It's a name of your kid, Fred. I like Fred. It's a great name. It's a strong name.

1:26:58Patrick Collison:Yeah, all these things go in cycles, though. It's the business cycle. There's news over at Meta. Meta has shaken hands with AMD. They're forming a pact. Today, we're announcing a multi-year agreement with AMD, Advanced Micro Devices, to integrate their latest Instinct GPUs into our global infrastructure with approximately six gigawatts. of planned data center capacity dedicated to this deployment. We're scaling our compute capacity to accelerate the development of cutting-edge AI models and deliver personal superintelligence to billions around the world. Very exciting. And pretty cool little hype video.

1:27:38Patrick Collison:You see the camera move on this video? This feels like you speed that up, you cut in some other stuff, and you got yourself an Instagram reel, right? You see this little, like, have you seen those tutorials about, like, how to make a car video? You need some SD kit. Yeah, SD kit on this. I think it goes pretty hard. You double the speed. You add some flickering. You add some frames, frame interpolation, all sorts of stuff. But Lisa Su is on an absolute tear. AMD's doing great. And Meta's not stopping.

1:28:04John Collison:Meta. Mark Zuckerberg's Meta is planning a stablecoin comeback in the second half of this year, eyeing a third-party vendor as a key partner to power payments across Facebook, Instagram, and WhatsApp. This is great. They should have something. If your friend sends you a meme, it's good. You should be able to tip them. Ooh, tipping. Tipping for great shares.

1:28:25Patrick Collison:Well, if you want to build a social network build on tipping, you'll need Plaid because Plaid powers the app to use to spend, save, borrow, and invest, securely connecting bank accounts to move money, fight fraud, and improve lending now with AI.

1:28:37John Collison:Sean Frank, soon to be a new father and dear friend of the show, says, Manus from Meta just doubling my ad budget every 14 minutes. This is one of the best new formats.

1:28:50Patrick Collison:I like this. This is only possible. This is the best AI image I've ever seen, I think. This is so funny because this definitely doesn't happen in the actual movie.

1:29:00John Collison:But this is what Burry is like now.

1:29:01Patrick Collison:I know. I know. Sean Frank, really on a tear. I saw him in the chat yesterday. I forgot to say hello to him. Hello, Sean. And also, congratulations on the new baby. Very excited for you. Anyway, this is a hilarious image. But we will return to the timeline after our next guest because we have Ivan from Notion in the Restream Waiting Room. Welcome to the show, Ivan. How are you doing? Hello, guys. Good to see you. Good to have you on the show.

1:29:27John Collison:Long overdue. Long overdue.

1:29:28Patrick Collison:We're so excited. First time. Fantastic. Well, we're glad we caught you on today because there's a big launch in Notion World, but I'd love you to take us through it. What was announced today?

1:29:41Devansh Pandey:We're launching customer agents today. it's one of the first if not the first multiplayer agent product for knowledge work so it does real work for you in the background very easy to set up post in the cloud yeah connect to all your work products and the best part is you don't need a mac media that's a good line they are going out

1:30:00Patrick Collison:of stock i don't know if you've seen um but the the mac media is in short supply so uh walk me through some of like the most obvious use cases like notion is i i think of the amazing because you have what is essentially a document but also a spreadsheet and you can kind of move between different data structures and visualizations on top of data in a sort of consumer app uh ui is and so i could imagine uh creating a document and then having an agent go and do a bunch of work to populate extra fields. So where are you seeing or where are you excited about these agents actually taking hold in the product?

1:30:39Devansh Pandey:So what you're describing was the notion probably two years ago. During the SaaS era, our strategy has been consolidating all different use cases to one product. We're talking about knowledge base, you're talking about documents, talking about project management. We have been bringing together into one tool that's very flexible. For example, Ramp actually the company sponsored you guys we bring Ramp last year as the new customer for Notion and we help Ramp consolidate half a dozen different tools that the collaboration stack onto Notion so they don't have to pay as much money for all the tools number one their team don't have to jump between those tools that's number two I would say the best part is now they have one place to do their core collaboration work they have one place to deploy AI so now Notion is the core agent orchestration layer for Ramp.

1:31:32Devansh Pandey:The product we just launched today, custom agent, Ramp has been an early customer for us for a couple months. They're running all the sales enable process, a lot of internal bug triage, all different processes on this, but because people have one place to core collaboration, system record truth, and one place to do their busy work, delegate their busy work too. Ramp's model is, what is it, money and time save both, and this is what we're doing for Ramp at the moment.

1:32:00Patrick Collison:I love it. Yeah, talk to me about the agentic cron job. That feels like something that we're starting to taste with open claw. There's clearly demand for it. It requires a little bit more upfront effort than just firing off a deep research report or saying, hey, hydrate this text, expand, contract, expand, turn it into bullet points, turn it into paragraphs, back and forth all day long. But I feel like for most businesses, having an agent that's effectively on a cron job, maybe you don't call it a cron job, but it's something that runs every day, that runs over a knowledge base, over a customer list, over documents, and does the things that AI is great at every day.

1:32:44Patrick Collison:That feels like something that could be incredibly powerful. How are you thinking about long-running agents, cron job agents, scheduled agents?

1:32:55Devansh Pandey:Yeah, crown job is a pretty good word for it. A lot of knowledge work is kind of just crown job. So you update your pushing paper back and forth and crown job from this person to the other person. I think the world sort of tastes this power of when agents connect with the crown job through a product like OpenClock. It can do a lot of work for you. And so you no longer have to prompt it. It triggers working on the background autonomously, asynchronously for you. Our interesting thing less about OpenCloud or Mac Mini is what does this do for real business? And real business is you require enterprise-grade permission.

1:33:31Devansh Pandey:It has to be multiplayer, no longer just for a personal tinker with your own Mac Mini. You have to power the entire teams with it, right? And it has to be easy to set up so you don't have to be an AI tinker or AI engineer to do it. You have to have the state-of-art models, usually the day release. So all the service will provide for businesses to take the spirit of front job, background agent, cloud code, you might say, open cloud, you might say, into businesses. That's the positioning of this product.

1:34:00Patrick Collison:Sure. So talk to me about where the capability frontier on the agent side is today. I mean, because agents can be turned really loose. You can give them access to Python and they can talk to any API. They can write their own CLIs at this point. And so you mentioned like no Mac mini, but is there a world where I tell, just for our example, like I want a new Notion document generated every day with a breakdown. I want you to go to a read-only access API for the YouTube API, pull all of our analytics, pull all the chat feed, synthesize all that, and put together a Notion document that I and review with the team in the morning that says, oh, this segment of the show was particularly great.

1:34:49Patrick Collison:Here's how the analytics changed. Here's where the viewer spikes were, all of that. That would require talking to an API. What does that look like if there's not an off-the-shelf integration?

1:35:00Devansh Pandey:All this should be possible if it's not really possible. Getting the YouTube API, getting the transcript, I don't know everybody have access to it. Gemini might have a special access to that. but video transcript, all this is possible because all you need to do is the runtime that can run model. All you need to do is the runtime that can talk to internal APIs through code that are written by models and a model that does the cron job periodically based on certain triggers. I just described those core ingredients. They're basically the core ingredients for Notion custom agents. So you only can do those and connect to your emails, connect your Slack if you guys use Slack and send a message every morning.

1:35:43Devansh Pandey:So you don't actually have to come to Notion to see the work being done. You can stay where you are today.

1:35:49John Collison:How have you processed the last couple of years of vibe coding? Because when I the first company I ever started or first, not necessarily the first company, but the first like real business, we started on Notion. And at the time, this company does like a bunch of, it's like an ad network on YouTube. And so we had a bunch of different like ad buys happening. We needed to be managing that process with the client as well as the creators. And so the entire company from the beginning ran on Notion. And I looked at every possible SaaS solution at the time. But I looked at all of them. I would have needed to, a lot of them didn't even work with customization.

1:36:30So I just built all these dashboards that helped kind of like manage all those different processes.

1:36:37John Collison:And it already had collaboration built in. It already had like the account functionality. So it just like worked completely out of the box. So in some ways at that time, I was already replacing like vertical specific software with Notion. And so in some ways, like I feel like this whole process and explosion of people being able to create different applications for different use cases is kind of like just a continuum from Notion's inception.

1:37:04Devansh Pandey:Yeah. We started, a lot of people think Notion is document tool, collaboration tool, note-taking app, relational database tool. That's never been the intent. Notion started as a computing tool. I really care about, okay, I'm a programmer. The power computing is in the hands of you, the programmers. How do you open up to more people? That's why the company started. So the spirit has always been consolidating the fragmentation of SaaS for the past five plus years. And it turns out that strategy works quite well with AI. Because once you consolidate those things, you have one context to power the language models.

1:37:42Devansh Pandey:That's one. Number two, because we've been taking a stance that we don't want to inject our opinion on how you should run your business, we should just provide the Lego blocks and you can decide however you want to run those Lego blocks. so we haven't been hard-coded business logic into our apps. So in the back then, there's a buzzword called no-code, right? And some people talk about SaaS versus language model. A lot of SaaS is hard-coded logic into your vertical apps. And we don't do that. It used to be a weakness of our product because it's how open-ended it is. It has to require some technical mind of people to use it.

1:38:14Devansh Pandey:It turns out to be a strength because now language model can use this notion and building block to do a lot of work for them. So now with this new product we're launching, it's not just working with information in and out of Notion. It can power agents to work with external tools and do those job, do those repetitive knowledge work. So do those busy work for the company. Internally, we call this like let AI do the night shift. So you can do the day shift. AI can do the night shift. We actually asked for a little bit dark mode this time because truly it's doing the night shift for us.

1:38:47John Collison:Nobody wants a night shift. I like the day shift. I did the night shift back in college. It's not a fun job.

1:38:55Patrick Collison:That's great. So you mentioned Gemini. Thank you. Another TVPN sponsor. But I imagine that you're pretty model agnostic. I'm interested to know how you're thinking about the different LLMs. And then also, how much do you want to surface to the user? Like I was talking to Salesforce's Slack bot and it wasn't up front with me about exactly which model was under the hood. Now I'm a nerd and I'll ask, okay, is it 3.5 or 4.6 or 5.2? And I'll have some opinion whether or not that matters, who knows. But do you want to have model switchers, model pickers? Do you want to be at that level of like empowering the user to pick the right tool for the job or Or do you want to handle that internally?

1:39:40Devansh Pandey:We do both. So if you're like a normie or normie++ using Notion, you can use Notion without picking the model. You'd be in a model version, right? But even more sophisticated, creative custom agent that do triage work for you, different models have different strengths and weaknesses. You should be able to pick the model. For our strategy, and I think for a lot of non-labs, it's very important to be model agnostic. The labs are going to get better and better, the models are going to do more and more. But one important strategic point is labs don't work well with other labs models. So there's an important position to be the Switzerland of agents, Switzerland of the models.

1:40:22Devansh Pandey:And that's the position where being with the product launch, you can work with clock code out of the box, you can work with a cursor's agent out of the box, and you pretty much can pick any models you want. that's the state of art and usually the day that those models are released. And as a user of this product, you don't have to worry about that.

1:40:41Patrick Collison:I want to revisit this Wall Street Journal article that you were featured in back in August of last year. So the quote was, Ivan, the CEO of Notion, says that two years ago, his business had margins of around 90 % typical of cloud-based software companies. now around 10 percentage points of that profit go to the ai companies that underpin notion's latest offerings how has that changed is it still 10 is it climbing is it falling what are your

1:41:12Devansh Pandey:predictions for where that goes not as far as 10 but it's definitely meaningful amount before you're you can do pure sas margins now model people our product are powered by ai now majority of product powered by AI now you have to model provider have to take some of the margin and we're okay with that we see the market change on both fronts first we we want to use data of our most capable most intelligent model because our customer wants that we want to eat the customer they don't have to worry about that on second there's a new wave of open source and foreign models are coming and that's why we have to be model agnostic and we can shift to different model for different type of work.

1:41:53Devansh Pandey:And that will help us with the margins. And at the end of the day, our customers are having to worry about this. What we provide is less about model. Model capability has been there for almost a year or two years to do a lot of knowledge work. What's missing in the market is this infrastructure layer that glued together model capability, glued together permissions to provide real knowledge work for the customers. At the same time, backward compatible, having a good UI for the company up devices.

1:42:22Patrick Collison:I have one last question. We'll let you go. How are you thinking about sort of like, it's crazy to call them legacy AI workflows because they were probably implemented like a year ago. But when I just think about like document summarization or even like spell checking grammar, like that was probably moved to an LLM that was capable, a GPT-4 class model can do that at a very low cost. And maybe you want to optimize that even further by going to an open source model on commodity hardware, really drive down the token cost. Have you left any AI workflows in place on legacy models? Or have you migrated everything to the frontier and you're just moving with the frontier?

1:43:09Devansh Pandey:We're moving with the frontier by and large because that's where a customer want. They want smarter things. but things like avoid dictation, summarization, legacy model can do that. And so it's just getting cheaper. The most important part is the market is changing so fast right now. Nobody knows where the future holds but we know the model capability gets better and better. And we always care about building beautiful and powerful tools. And AI is that tool today. How do we make sure that all companies can benefit from this? You don't have to be a Fortune 500 for forward-forward engineers. you don't have to be a San Francisco startup to have AI engineer on your team to use this.

1:43:48Devansh Pandey:Every business can benefit this technology. That's our ethos. And that's why we're building this product to make it super simple. You don't have to worry about Mac Mini, not worry about models.

1:43:59Patrick Collison:Is that on the homepage yet? Don't worry about Mac Mini. We got you. I think that's too specific calling out other products. But Night Shift works. I love it. Someone in the chat, John Palmer in the chat was saying, well, but if I can't, if I don't have to use a Mac mini, what will I spend all my time setting up?

1:44:21Devansh Pandey:Like the tinker. Tinker is sometimes more than 50 % of the fun of it.

1:44:26Patrick Collison:No, this is true. This is true. People want to tinker. They want to play. They want to explore and have fun. And it seems like it's a great time to be running Notion. Like it's just a very exciting time. There's so many new products you can build so much faster than ever before. So congrats on all the progress. Yeah, congrats to the team on the launch.

1:44:44John Collison:Thank you so much for stopping by. Great to finally have you on.

1:44:46Patrick Collison:We'll talk to you soon.

1:44:47John Collison:Cheers.

1:44:48Patrick Collison:Let me tell you about LabelBox. Reinforcing learning environments, voice, robotics, evals, and expert human data. LabelBox is the data factory behind the world's leading AI teams. And let me also tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. I'm not going to leak the news, but we have an exciting guest lined up for our next NYC show. So hit that subscribe button to be notified when we go live.

1:45:15John Collison:Can't wait. A senior U.S. official told Reuters that DeepSeek's new model, whose release is now imminent, has been trained using NVIDIA Blackwell GPUs despite the export ban. I am interested to see what this model is capable of.

1:45:34Patrick Collison:could have made that happen, right? Like it could literally be one black well per person in a suitcase smuggled along. It could be one shipment diverged or diverted from going to one country. And then they say, help send that shipping container over there instead. It could be cloud. Like they could have found a cloud provider that they were able to sort of anonymize and have a front company for. There's a whole bunch of different ways to get compute if you're willing to bend the rules or break the rules or potentially anger the US administration. But we will see how this goes. Tyler, do you have a feeling for how DeepSeek has been doing?

1:46:16Because there was a hype cycle around DeepSeek v3 point something,

1:46:23Patrick Collison:and it kind of came out and it landed with, it didn't make a big splash. I feel like we're going into a new hype cycle around like the next deep seek is gonna be really good Is this fake is this real? How are you dealing with deep sea? So I think the last big model release was supposed to be this like massive massive release right and it turned out to be

1:46:43Scott Wu:Letting where like instead of like I don't know the exact number It was supposed to be like before and ended up being like v3.1. It was like that kind of thing

1:46:50Patrick Collison:And then also like so they box the piece of the pre-train most likely is that what people think yes, maybe and then

1:46:56Scott Wu:And, yeah, also, like, on Chinese labs generally, like, right now you're hearing a lot about, like, it's, like, ZEAI and Kimi and those kind of things. Yeah, so it's very unclear. I mean, they're not very public about this stuff. Yeah. But also, I think broadly, just about the distill gate stuff. Yeah. I think throughout this I've been, like, I think I've, like, updated towards, like, actually we can probably mostly ignore a lot of the Chinese labs. because basically, like, the whole reason that they're good, like, everyone's like, oh, my gosh, Deep Seek is right on our tail. They're going to catch up.

1:47:31Scott Wu:They're going to catch up. The only reason that they've been on our tail is because,

1:47:37Patrick Collison:no, no. No, no. That's the royal flush, Tyler. That's the best thing you can do on the screen. You just dropped a bomb. Truth nuke. Truth nuke. Truth nuke. Disregard China entirely. You heard it here first.

1:47:50Scott Wu:No, but, I mean. Yeah, no, it's a good point. Yeah. The only reason that they're like... Stop. So rude. The only reason that the Chinese labs are so close to U.S. labs is because they're just training on the outputs, right? Yeah. Which is like, okay, sure, like, yeah, good job. But like, you're not... I would be extremely surprised if you actually see a breakthrough from a Chinese lab so far. Exactly. Everything we've seen is that, yes, they can copy stuff.

1:48:17Patrick Collison:Permanently three months behind. No. But I think people are freaking out because China went from like 10 years behind to one year behind to three months behind. And they were like straight lines on log graphs. They're going to be 10 years ahead of us next year. And it just feels like that's what happened.

1:48:32Scott Wu:I think you can still be very worried about regulatory capture and all these things. But I think Anthropic will basically just figure out a way that they can increase security on the API. Totally. So we'll open AI and then we'll see if the Chinese labs keep up with the progress. But, you know.

1:48:50Patrick Collison:Yeah, yeah, yeah. I think broadly, that's how I've... There's so many other dynamics beyond just obtaining training data. Like, if you take Will Brown's point that the internet is producing more training data, you wind up in a situation where, sure, training data is commoditized, but what does it really take to scale up DeepSeq v5 to a place where it's having economic impact? Well, you need a massive inference cluster. Do they have that? How are they distributing this stuff?

1:49:18Scott Wu:I do still think it's very impressive that the models generally they've put out are very small and still very good. But I think on the frontier level, I'm not super worried about them.

1:49:27Patrick Collison:Tyler called it. They're cooked. Anyway, really quickly from Anthropic, there's now a call sheet on will the Pentagon designate Anthropic a supply chain risk? It's sitting at 36.8%. I believe that there's going to be a meeting between Pete Hegseth and Dario. So that already happened. That happened.

1:49:48John Collison:Update on the meeting from Andrew Curran. According to Axios Defense Secretary, Pete Hegseth gave Dario until Friday night to give the military unfettered access to Claude or face the consequences, which may even include invoking the Defense Production Act to force the training of a war Claude.

1:50:08Patrick Collison:So that was not a joke? War Clawed is not a joke?

1:50:11John Collison:I don't think they would name it that, but it kind of sounds like that's what Pete is asking.

1:50:16Patrick Collison:Clawed of war. Like God of war? That's pretty good.

1:50:20John Collison:Anyway, we have our next question. Before we get to that, the chat is sharing that payments processor Stripe expresses interest in PayPal. Scoop. We had a missed opportunity. If you guys could have simply, if Bloomberg, if you could have published that at 12 when they came on the show, That would have been quite nice.

1:50:39Patrick Collison:Don't publish it until they're off there.

1:50:41John Collison:Payment processing firm Stripe is considering an acquisition of all or parts of PayPal. Stripe, which is privately held and is among the industry's most valuable companies, as you know. The deliberations are still early, and there's no certainty they'll lead to a transaction.

1:50:54Patrick Collison:I mean, we read the Wilman Nytus post, and we were like, oh, it could be a couple days, could be a couple years. It seems like it might be closer to a couple days.

1:51:02John Collison:Well, PayPal is up 7 % today. Ooh, that's interesting. People are excited to say, hey, you might be able to own some Stripe.

1:51:10Patrick Collison:Well, let me tell you about Phantom Cash. Fund your wallet without exchanges or middlemen and spend with the Phantom Card. And without further ado, we have Stefano from Inception Labs. He's the founder and CEO. Welcome to the show. How are you doing?

1:51:24Scott Wu:Very good. Thanks for having me.

1:51:26Patrick Collison:Thanks for hopping on. First time on the show, so I'd love to have you kick it off with an introduction on yourself and the company.

1:51:33Scott Wu:Of course, yes. I'm Stefano. I'm one of the founders and the CEO of Inception. Before this, I was at Stanford in the CS department. I've been doing research in generative AI for a long time. I think my lab is mostly famous for having co-invented diffusion models back in 2019. I was on the flash attention paper, DPO. So a bunch of things that are now widely used. Introduction. And these days, I'm most excited about the diffusion language models. That's what we're doing at Inception.

1:52:00Patrick Collison:Yes. So I first saw a diffusion language model demoed at Google I.O., I believe. But tell us, like, explain it like I'm five. Because when I think diffusion, I think a bunch of fuzzy noise, and then the mid-journey image gets higher and higher resolution. Everyone's familiar with that. And then they're familiar with, like, the token streaming next token prediction. Is it different? Break it down at a very low level or high level. That's right.

1:52:29Scott Wu:Basically, we've taken diffusion models, which is the thing that works best for image and video generation. This is kind of a course-defined process where you iteratively refine your output until it looks good. And we've figured out a way to apply it to text and code generation. And it kind of works the same way. You start with a rough guess of what the answer should be, and then you refine it. And crucially, the difference is that the neural network is able to modify many tokens at the same time. And so it's much, much more efficient than the typical autoregressive model where you generate left to right, one token at a time.

1:53:06Scott Wu:So you're able to modify Benetokians in parallel.

1:53:09Patrick Collison:So if I'm thinking of like, you know, not maybe like a deep research report type response, in my mind, I can imagine a report, you know, saying like, explain the history of the Roman Empire. That's the example I always use. It's like, it's going to have some structure to it. I'm going to imagine a blurry image with like a couple of large headers and then the headers are going to get filled in. Then the text is going to fill in. Maybe there's some bullet points, maybe there's some dates, maybe there's some charts. And like all of this is going to come together. But I'm thinking about it not sequentially, but as a whole, and then refining iteratively until I'm getting to instead of pixels, I'm thinking of individual characters.

1:53:50Patrick Collison:Or are there tokens in the same way that might exist in LLM? What does that look like?

1:53:55Scott Wu:Yeah, that's the right intuition. so it's kind of like course to find generation and in practice it's learned by a neural network so it's not necessarily interpretable it's not the kind of process I would go through where maybe I start with section headings and then I fill in the details it's all learned by a neural network so it's not really interpretable but it's fast

1:54:20Patrick Collison:so is speed the main thing we had the founder of chatjimmy.ai on the show, Talos, and it seemed like he was able to bake down a traditional LLM, Llama 3 8B onto Silicon, and it was spitting out 16 ,000 tokens per second. Do you have a comp on speed or cost that you're targeting? Or do you see like a through line to like, okay, maybe if we're running on Nvidia chips, and he's running on custom Silicon, he's going to be faster. But then once we get to custom Silicon, we're going to be 10 times faster than that. How should I be thinking about the trade-offs here?

1:54:55Scott Wu:Yeah, so our benefit is purely at the algorithmic level. It's just a more parallel approach that is not memory bound, it's flops bound, it's compute bound. So you're able to hit the ceiling of the roofline and we're taking advantage of all the resources we can get access to on the GPU. In practice, what this means is that we can get to over a thousand tokens per second on traditional NVIDIA GPUs, Hopper, Blackwell. So we are not yet at the level of the 16 ,000 tokens that you can get if you were to actually implement the model on hardware, but we're running on general purpose GPUs. So we can scale up as much as we want.

1:55:38Scott Wu:It's just a matter of getting more GPUs and you can just run these models anywhere. We are on Bedrock, we are on Foundry. So if you have your own GPUs, you can provision your own capacity and you can run our models there. so it's all very very scalable it's fast and scalable and in principle yeah it can be compounded you know you have a 10x benefit from the software you have a 10x benefit from the hardware those two things could be combined what uh what use case you know there's a lot of people

1:56:07John Collison:out there using uh traditional language models today what are the kinds of use cases where you would tell somebody you should be switching over today or at least trying to start experimenting

1:56:19Scott Wu:yeah we're seeing a lot of traction in latency sensitive applications of llms like whenever there is like a tight loop where you need to interact with a developer or a customer so our models are being deployed in a bunch of ides as if you think about coding uh coding autocomplete next edits suggestions refactoring quick agentic loops that's a very natural kind of like application where diffusion LLMs are already really, really good. Voice agents, we have a number of partners and customers that are building really, really good voice agents. The latest models we announced today, Mercury 2, is a reasoning model.

1:56:56Scott Wu:So it's really, really fast. And so you can get the quality of a reasoning model with the latency budgets that you need whenever you want to build a voice agent, which is resonating really well with a bunch of early customers. Retrieval and search, that's another space where we're seeing a bunch of applications being built on the Fusion All-Lams. So if you think about query rewriting, re-ranking, summarization, that's another really, really good use case for the Fusion All-Lams.

1:57:27John Collison:Talk to us about DistillGate, how you've been processing it. Have you worked a bunch of that?

1:57:34Patrick Collison:It's a sign of success.

1:57:37John Collison:Any points in your career where you were experimenting with this stuff? Is this something that we kind of forced the Chinese market into spending a lot of resources on?

1:57:47Scott Wu:I mean, it makes sense, right, that that wasn't always going to happen. I think the moment you put it out there, you know, you give API access to the world, that's going to happen and people are going to copy you. I mean, we've been doing distillation in the research community for a long time. And so people have been experimenting and figuring out ways to do it in a sample efficient way. So I'm not surprised that it's happening. I think it's hard to know at what scale. And honestly, from the numbers that they were circulating, it seems like they are able to do it with very, very few data points.

1:58:23Scott Wu:That was the most surprising thing to me. So, you know, it's very interesting scientifically that you can actually distill with so few data points because it means that it's going to be very, very hard to protect any IP. You are opening a model up from an API point of view.

1:58:40Patrick Collison:So the last question, somewhat related to that. But I feel like when these models get distilled, we see very strong benchmark performance. And then some yet-to-be-quantified and benchmarked quality sort of degrades. And you hear people that actually try and put them into production saying, like, it just doesn't have the same, like, big model flavor that I'm getting from the big labs. I don't know how real that is, but I'm wondering if you zoom out and you look at diffusion versus transformer-based LLMs, are you noticing any divergence in the benchmarks where you're maybe better at coding or less good at coding where the mental model that we're giving the computer is leading to surprising results?

1:59:30Scott Wu:yeah so what we're seeing is that it's it's it's good at coding it's good at editing uh one nice thing about not necessarily being left to right is that you can use context all around you and so those use cases have emerged as being really really good for diffusion of the labs i think it's also a function of the training data that we use uh you know we always liked coding we're all computer scientists and so that was like a a very natural kind of application area for us. And so I don't know how much of that depends on the training data that we use versus the model. But what's exciting is really just like the speed.

2:00:04Scott Wu:That's the thing that is going to be hard to replicate.

2:00:08Patrick Collison:I got a need for speed. I got a need for speed. I'm super bullish on speed. I'm serious. I think it's amazing. I used 5.3 Spark on Cerebris and I was like, this is the future. It's going to come to everything. And it's going to be an important moment for people to realize that it's just a different product when you're interacting with something fast. And I think we learned this from Amazon squeezing out milliseconds in web page loads, and we're going to experience it in AI too. So thank you for everything that you're doing to speed up AI. We loved having you on the show. So have a great rest of your day.

2:00:42John Collison:Great to meet you. We'll talk to you soon.

2:00:44Patrick Collison:Goodbye. Let me tell you about Console. Console builds AI agents that automate 70 % of IT, HR and finance support, giving employees instant resolution for access requests and password resets. And let me also tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web apps, servers, databases, and more, while Railway automatically takes care of scalering, monitoring, and security. And without further ado, we have TBPN Royalty.

2:01:13John Collison:What's going on? Great to see you, James. Hey, John. Hey, Jordy. How are you doing? Doing great, doing great. Good calling in from a cave. Are you fully snowed in? What's going on?

2:01:25Reiner Pope:Yeah, we're in New York. The snow is melting. We built a little mini studio upstairs. And yeah, it looks pretty professional.

2:01:33John Collison:Great. I love it. Not as professional as you go. Tell us the news and then there's a bunch of stuff we want to talk about.

2:01:41Reiner Pope:Yeah, so I guess we're joining today announcing a$96 million Series C investment, a$1 billion valuation led by Lightspeed Venture Partners alongside Sequoia, Kleiner Perkins, Avantik, Saga, and South Park Commons.

2:01:58John Collison:Amazing. Break down everything that's happened since the last time you were on the show. The space has been moving so quickly, so it feels like it's been two years, even though it's probably been two months.

2:02:11Reiner Pope:Yeah, I mean, it's all moving. Everything's moving, obviously, 100 miles an hour. But it sounds a bit trite when I say this, but it really is a privilege to be building in such exciting times. um yeah i mean we we've just launched profound agents which i think is a really big deal it's you know we we serve the marketer so you know the the the line we've been using during this fundraise is um or during this announcement has been you know harvey uh lawyers have harvey engineers have cursor and marketers have profound and i think that's that's truer than ever in that yeah with this launch of agents it really takes profound uh towards being like a full stack you a holistic platform for the modern day marketer allowing them to you know not just understand how they show up in ai platforms like chat gpt gemini the rest of them but also build agents that can help them do more with less um so yeah i think this is cool yeah we saw our customers had been you know we we came out of the gates 18 months ago with profound here in new york and what we saw was our customers quite often we're taking our data and insights and then going to orchestration and automation tools to do cool things with it so we've just brought that all in-house now and uh yeah it's really cool you can do everything in one platform how how are how do you think the the

2:03:35John Collison:uh other platforms the llms are are evolving some of them are launching ads some aren't all products are being and services are being discovered and in all of them why is it important to have a platform like Profound. We were talking about this off-air this morning. It feels like people have been joking around about Manus, for example, in the meta platform because Manus is like an agent that wants to help you, but at the same time, what helps Manus? It's like spend more money, right? So it feels like having a third party that is -

2:04:07Patrick Collison:The principal agent problem.

2:04:08John Collison:Yeah, Manus is like, I've got a great idea. You guys are more user aligned potentially, but how are you thinking about the interaction between Profound and the different platforms?

2:04:20Reiner Pope:Yeah, I mean, I think, you know, our prediction of the future is that in the future, every company on the planet will care deeply about how AI talks about their brand or products or services. That's kind of a North Star that we hang our hat on. And I think compared to, you know, Search in the early 2000s or even for the last 25 years, it's looking like this will be a much more fragmented sort of market. I think we're going to see multiple players coming through. So I think Profound really sits adjacent to the models or the labs. And we help marketing teams understand how they show up in these platforms.

2:05:02Reiner Pope:You know, when AI responds, what does it say about your brand? What does it say about your services? And now we help you build customized agents that can actually do the work with a marketer in the loop. So, yeah, we've had hundreds of teams. We work with, I mean, I guess a big thing that I'd say we're announcing since we last spoke, our Series B, is that we now work with 10 % at the Fortune 500, which is a pretty cool start. That's it by God's sake.

2:05:35John Collison:What are the other 90 % doing?

2:05:38Patrick Collison:That's fantastic news.

2:05:41Reiner Pope:Yeah, we got 90 % to go, job's not done. But I think, yeah, it's a very cool stat. And what we're seeing more and more is that every brand is different. Every marketing team is different. Everyone has different initiatives. Everyone has different preferences. Marketing is more human than ever in a lot of ways. And I think our approach of helping marketing teams build entirely customized agents that can take out the rote labor from their work is it's just saving these teams inordinate amounts of time and energy. And it's very cool to see it work. Yeah, it's exciting.

2:06:20Patrick Collison:Can you walk me through the anatomy of correcting a mistake that exists across LLMs or even in a particular LLM? My nightmare is, you know, you go to you go to chat GPT and ask how tall is John Coogan and it says 6566, this would destroy me. It must be 68. Let's bake that into the pre-training data, 6868. But seriously, other than just doing a bunch of SEO to correct the record, how can a company, if there's truly a consistent hallucination, something that's just incorrect for some reason, what is the process to actually change results? For sure.

2:07:04Reiner Pope:I mean, well, the first step, which sounds kind of stupid, is just knowing why it's happening, right? So the model, when, you know, let's say an answer engine spits out an answer, a good chunk of the time it's getting that answer from somewhere. and being able to identify, hey, this is what we found. I'll give you an anecdotal example. So it was, I wouldn't be able to name the brand, but it was a neobank that the models were incorrectly spitting out that there was no FDIC insurance on. And we identified, it was coming from a few places. It was like a third-party blog, I think a couple of Reddit posts, and maybe like a YouTube video or something.

2:07:47Reiner Pope:So then once you know where it's happening, it's kind of, I wouldn't say it's easy, but it's, you know, it's just kind of 101 marketing. Okay, cool. Let's reach out to the blog and tell them that that's factually incorrect. Let's comment on the Reddit post and say, hey, this is actually not true. We are FDIC insured. Let's produce a YouTube video that speaks to the same thing, but, you know, mentions heavily that we have FDIC insurance. And lo and behold, that gets pulled through into the model.

2:08:11John Collison:So that would be something that your agent would do automatically and you could just set them off and do that.

2:08:19Reiner Pope:Correct. Yeah, you could set up an agent that monitors for any misinformation based on a knowledge base of ground truth and then say, OK, cool. When we see any misinformation, let's generate an email that reads from our tone of voice and sends to this third party blog and says, hey, can you correct this, for example? So yeah, but it has to be customized because you'd never have that as an out of the box solution, right? It has to be everything's, it's almost like being able to build one for one software. This is this new paradigm of agentic software, which is so, so cool. And obviously, I'm not the only one that's excited about that.

2:08:54Patrick Collison:Yeah. I mean, you're in a very interesting vantage point in the industry because you work with so many Fortune 500 companies. What are your expectations for agentic commerce this year? We were just talking to the Colisons. It feels like it's on the precipice. Everyone we've talked to is extremely bullish, but I'm always interested to hear the shape of the bullishness, what you think needs to happen to actually get people shopping agentically this year. I mean, I think so much of that inflection is going to come from the models themselves or the consumer products.

2:09:35Reiner Pope:So, you know, ChatGPT and Gemini's ability to actually offer a fantastic user experience. I think that that will be the kind of, you know, that's the most important thing. I think from the other side of the fence, working with these brands, these marketing teams, there's a hot take. They're not, I hope this doesn't sound disrespectful at all, but they're not as slow as you'd imagine. We work with giant brands, Fortune 500, some Fortune 10, and they're ferociously fast. They understand the magnitude of this platform shift and it's very sophisticated teams. A lot of them are SEO teams who I actually think are fantastically well suited to kind of attack this problem space because they're like kind of technical and they understand the sort of primitives of marketing and content, et cetera.

2:10:26Reiner Pope:They're quite cross-functional. But yeah, I think the mistake would be to think that the enterprise is super slow. I don't think that's true. And I'm not just, there's no secrecy there. I actually believe that.

2:10:41Patrick Collison:Yeah, no, I completely agree.

2:10:42John Collison:Well, I asked ChatGPT, what's the best geo tool for startups? It says, profound. Oh, yeah. What it does. There you go. Tracks how your brand appears inside LLM. It's best for VC startups. Dogfooding. Serious about AI distributions.

2:10:58Patrick Collison:Dogfooding. I mean, we put a lot of profound in the pre-training data last year. It was a great partner.

2:11:03Reiner Pope:Did it mention agents, though? That's the question, right? We only don't say agents today. Oh, yeah, okay. Do you want to ask it? What did they launch today? Oh, there you go.

2:11:11Patrick Collison:While he does that, tell me what you think of the word GEO. Is that too buzzwordy? Do you like that term? What are the pros and cons of having a term applied to your industry, your nascent business plan?

2:11:29Reiner Pope:I think geo sucks as an acronym. It's just bad in so many ways. I think it can't be claimed because of geography. It is already taken. It stands for generative engine optimization, which people don't refer to these products. Have you ever heard anyone refer to chat GPT as a generative engine? They don't. Yeah, you're right.

2:11:53Patrick Collison:Google is a search engine for sure. Bing is a search engine. But no one calls Gemini.

2:11:58John Collison:What's your preferred acronym?

2:12:01Reiner Pope:I mean, sort of without, you know, said sort of with not much passion, answer engine optimization feels more fitting to me. I think it's still all to be determined. I think how your brand is spoken about by AI will become the most important primitive in marketing. So I think it's going to become bigger than just a kind of, you know, something that you put a label on like that. I think we see a new sort of new type of marketer forming over time, which is interesting. The marketing engineer, you know, a marketer who has the technical chops to be able to go in and build agents, customize agents, deploy agents for the rest, you know, cross-functioning, you know, across teams.

2:12:52Reiner Pope:And yeah, I think that's very interesting. We announced a profound university today. So does this work? If I press that, can you see that? There we go. That's elite. That's an elite. Oh, wow.

2:13:04John Collison:That looks amazing. Class is in session. Series C.

2:13:07Reiner Pope:Oh, yeah. How much is it again? Yeah, so we announced a profound university today. Very cool. Which is actually really awesome. It's a series of certifications, training cohorts, learning materials that essentially enables this new era of the marketing engineer. And yeah, we're very excited about that. So yeah, I think we're going to see a lot changing in the world of marketing, as I guess is true in most industries.

2:13:38John Collison:Chad is saying that Ayo is a good Ayo.

2:13:42Patrick Collison:Ayo. Sheesh. Well, thank you so much for coming on the show. Always good to have you here, James. Congratulations.

2:13:50John Collison:It's great to see.

2:13:51Patrick Collison:And we'll talk to you soon. Thanks, James. Have a good rest of your day. Let me tell you about Cognition. They're the makers of Devin, the AI software engineer. Crush your backlog with your personal AI engineering team. And we have Scott Wu from Cognition in the Restream waiting room. I want to talk about the launch today. I want to talk about AI progress. I want to talk about math and your predictions on the IMO gold medal and everything that's happening there. But let's start with the general update on Cognition. What's the shape of the business today? And then I want to hear about the latest launch.

2:14:27Ivan Zhao:Awesome. Yeah, what's up, guys? How's it going? Great to see you guys. It's been a little bit, I feel like. Yeah, too long. Too long. You guys have been cooking. Every month feels like a decade now in AI. So cool. No, so things have been great. I mean, business has grown a lot. We shared some of our metrics today, one of which is that our total enterprise usage has actually more than doubled in the last six weeks even. What? Six weeks? And a lot of that has just been mass takeoff of agents. I think the high level that we're really seeing is that as agents get more capable and you can trust them to do end-to-end tasks, what you really need is the full background cloud agent.

2:15:06Ivan Zhao:Right. And so that means being able to run your repos and everything locally, being able to test, being able to spin things up from Slack or Linear or GitHub or Jira or whatever it is, and just being able to have this mass parallel async workload.

2:15:21Patrick Collison:Okay. And then the announcement today.

2:15:24Ivan Zhao:Yeah, I know the announcement today was a fun one for us. It was very near and dear to my heart. But a lot of it, honestly, if I were really to just describe it in one line, is just clearing through all the frictions that we've known about and just making it a really great experience. One of the big highlights is automated testing and having Devon run your web app for you and send you the changes and send you screen caps of all of those things. But there's tons of little things that really affect the experience. And so making the VM startup time way faster, making the Slack integrations way smoother, showing you all of the intermediate progress of the messages and so on.

2:16:04Ivan Zhao:And so it's changed our internal usage a lot. And so that's why we're pretty excited to get this one out.

2:16:09Patrick Collison:So it seems like there's speed to be squeezed out from VM spin-up time optimizations. We're also seeing some incredible progress on the custom silicon. We had the founder of Talos on generating 16 ,000 tokens a second. That seems like that will be really impactful when it rolls out to the broader code generation and software engineering world. It's still pretty early with that company, Llama 3B at this point.

2:16:43Patrick Collison:But where else are you seeing opportunities for speed? How do you think about the importance of speed for what you do?

2:16:52Ivan Zhao:Yeah. No, there's a ton that you can do. And at some point, a lot of it actually is just good old software engineering. And so, you know, it's, of course, like, you know, the models, obviously, you know, you can improve the tokens per second. you can improve the TTFT. I think those improvements will be great. And we've already seen a lot of those over the last bit. We'll see many more. But at some point, your agent has to go install, NPM install. Your agent has to go UV install. Your agent has to go grep for things. It has to go pull up the front end itself. A lot of that stuff is good old product building and software engineering to make that better and more efficient.

2:17:29Ivan Zhao:And so in a lot of these, obviously, you can do algorithmic tricks. You can put in indices, right, and indexes and make those faster. You can do little things to kind of like cheat the loading time and do things in parallel and do things async. But a lot of it is just building the systems around the agent to make it really fast.

2:17:46Patrick Collison:No, it's a really good point. I mean, anyone who's installed OpenClaw has experienced like, oh, wait, I'm actually just waiting to download software because it's pulling a whole bunch of stuff together. And it's not actually doing that much waiting with the LLM, at least in the setup phase. But you still have to actually get this thing configured. And I think a lot of people in tech went through that. How have you been processing lessons from open claw interaction patterns that you think are interesting? What it means that society more broadly is just aware of AI agents, which I feel like is a term that you basically coined years ago and have been running with, but in a specific enterprise context.

2:18:27Patrick Collison:And now I'm at a bar and I'll hear somebody talking about AI agents and it's because of OpenClaw. And I feel like, oh, I remember the Scott Wu launch video where he explained that this was going to happen. But how have you been processing OpenClaw? What is interesting about that? Are there any lessons from that open source community, that project generally, that paradigm that you want to bring to Devin?

2:18:50Ivan Zhao:Yeah, no, I mean, a lot of big changes. And I think, by the way, I think OpenClaw gets a lot of credit for many people being the first time that people really saw what a full agent would look like with access to your files, access to your computer, and so on. I think we're really getting to the point to your previous point where I think we're really starting to switch over from the early adopter cycle to the kind of mass market cycle, is my sense. And the concrete impact of that is a lot more people are starting to hear about and really think about AI agents. And I think it used to be, I mean, for us, for example, a year and a half ago, you used to go into the room and explain to people what an AI agent was and why this was different from just normal autocomplete or chat GPT or something like that.

2:19:37Ivan Zhao:Now everybody's thinking about this stuff. Everyone wants to use it. And I think one of the implications of that is just accessibility and getting people to value as soon as possible is one of the most powerful things that you can have in your own products as a result.

2:19:51John Collison:You guys have had a ton of success in enterprise. The chat wants us to ask for your take on the SaaSpocalypse, I imagine. Some of the conversations that you're having with, let's say, the CTO of a massive company, are they thinking about using a Devon for things like big database migrations? like how are they thinking about how agents can impact their dependency on sort of these like legacy tools and systems of record?

2:20:25Ivan Zhao:Yeah, I mean, there's the whole Citrini report and everything. I mean, it was honestly ridiculous. That's my two cents on it. I think the like, look, at a high level, of course, yeah, AI is going to change a lot of stuff. I don't really understand how you go from that to saying that there's going to, you know, like take software as a good example. Software is one of the most deflationary things ever. You know, a lot of the same products that, you know, used to cost much more 10, 20 years ago have gotten much, much cheaper over time. Right. Has this been terrible for a software company? You know, I mean, it seems like it's been pretty good.

2:21:02Ivan Zhao:All the big companies in the world are still software companies. Right. And so I think there's like, there's one thing when prices go down because, you know, the demand's just not there anymore. And obviously, you can get into weird cycles and all that can happen. But it's a totally different thing if prices go down because we've just gotten way better at supplying things. And that's when you get Jeven's Paradox, and that's when you get, you know, just mass consumer surplus and so on. And so at a high level, I mean, I think there's all the customers that we work with, you know, banks and health insurers and private equity and so on.

2:21:34Ivan Zhao:I mean, they're obviously like a lot of these base migration, modernization projects that they can go and take on immediately. But the very next thing that they say is then like, okay, how do I pull the rest of my roadmap forward? How do I build even more and get even more out to people? And I think in reality, we all just have so much more software to build.

2:21:54Patrick Collison:Yeah. How are you thinking about AI progress broadly? It feels like a lot of people are feeling that recursive development is on the horizon. People are bringing up takeoff speeds again, migrating from slow, maybe fast. I was backing off. Now it's a little quicker. How are you trying to like zoom out, reset, get to reality, figure out how fast things are actually moving?

2:22:21Ivan Zhao:Yeah, no, I mean, the meter report shows like the consistent doublings and everything. I mean, I think it's a very, I think things are continuing on the exponential curve. I wouldn't say that they're going either super exponential or sub-exponential. I think they're roughly going on that exponential curve. But exponential curve is a lot. That's a very fast growth, obviously. I think for us, one of the things that's been pretty interesting is just noticing each of the step function changes that happen. And so for us, for example, it's definitely been in the last, I'll call it four or five months, where something interesting happened, which is we stopped typing code.

2:23:01Ivan Zhao:you know like at some point you you just don't right like like before obviously you have all the tools and you have the combination of things and and so on now it's there's different experiences there's different tools that you want to have obviously between the ide and the cli and and the web agent and so on but either way you're you're really just working in prompts and you're not really you know like the code that we check into github like how much of it was typed by a human at this point i think almost none yeah uh and and maybe one of the things i would just call is that as you kind of expose each new thing, if you think of it as a profiler on your own software engineering workflow, what is the most expensive part?

2:23:40Ivan Zhao:You shrink that down, you get to the next thing, you shrink that down, and you just make the whole cycle more effective. We're at the point where a lot of these other things like understanding the code base and review and so on are the actual bottom X. Testing is another big one. And I think what we're going to see over the next little bit is you're basically going to have to solve each of those with really good product experiences, really good model capabilities, and so on. So maybe the only thing that I would say, I think the exponential curve continues. I would just call out that the form factor looks very different as you continue on that exponential curve, because you're actually solving different problems.

2:24:16Ivan Zhao:Yes, I think we will continue to get the doublings and the doublings, but now it looks a lot more like how do we optimize testing and review and planning, not how do we make the AI good at writing code based on the prompt that you give? Because at this point, it's actually, frankly, it's basically already done.

2:24:33Patrick Collison:Yeah. I have a bunch more questions. I'll be quick. I have two. First, what does the future of Winsurf look like in a world where you're not writing code? Does that become a Kindle? I mean, that's a joke, but does it become more important for that product to be the best way to read code? Because even if you're not writing code, you still, like I've done terminal prompts and I'm like, and then I wind up opening the files to kind of take a peek in them. And I'm like, ah, I kind of like, there's maybe room for innovation there and like how your code reading skill improves as your code writing skill degrades.

2:25:13Patrick Collison:But how do you think about the future of Windsurf?

2:25:15Ivan Zhao:Yeah, for sure. I think the high level here is, it's gonna be a gradual thing, But I think over the next one or two years, we'll have a pretty broad transition towards what you might call having English as the source of truth.

2:25:27Patrick Collison:Sure.

2:25:27Ivan Zhao:And so people talk about like, basically, I think we'll go from code to English in the same way that we went from assembly to code.

2:25:33Rune Kvist:Yeah.

2:25:33Ivan Zhao:Right. And so, you know, one of those steps has been, you know, has been mostly done at this point, which is the step of figuring out how do you prompt in English and then have the agent produce the code. But if you think about it, I mean, you're still, you're still, you know, reviewing code. You're still checking code into GitHub. You're still reading the code to understand what's going on. And I think at some point, you actually want an interface that looks a lot more like a spec or a map or, you know, like a design doc, right? And that's the thing that you're iterating on. That's the thing that you're reviewing, for example, right?

2:26:04Ivan Zhao:Like at some point, review, I mean, people say, oh, like review is going to go away because AI is going to catch all the bugs. I think that's actually not right because what you're going to be reviewing is the decisions, right? It's like, here's what the product is doing in this case, and here's what the product is doing in that case. and here's how this plan works or whatever it is, right? And so what you'll want to have is a very clean interface to interact basically with your product and with your own specs and so on. And that's a lot of what we think Windsurf evolves into over time, right?

2:26:35Ivan Zhao:And so, again, I think it's a very gradual thing. I think there's a lot of value in reading code now and certainly at Cognition we still do a lot of reading the code even if we are not the ones writing the next line of code because instead we write the English prompt for that. But I think what happens with Windsurf is at some point instead of looking at each of the files, you start looking more at the specs and the high-level logical design of what you're building. You start looking at the diagrams of your app or your website itself and you're able to go and manipulate those. And you're really just managing your agents that you kick off from there.

2:27:12Okay.

2:27:13Patrick Collison:We blew past the IOI gold medal, as you predicted correctly. Amazing. I have a follow-up question about that, but it's sort of in three parts. One is, what is the next math or physics-based benchmark that you're excited about AI potentially unlocking? when do you think that might happen? And then do you think there will be any tangible impacts of that? Because if I walk down the street and I tell some random person, like, they did it, Navier Stokes is solved. I think that's the math problem that everyone talks about a lot. I don't even know. I think most people would be like, great, like, is that going to help me with my job?

2:27:53Patrick Collison:Like, they're more excited about just knowledge retrieval right now. So yeah, the next hurdle, timeline, and then impact?

2:28:02Ivan Zhao:Yeah, yeah, for sure. So, I mean, we actually, I would say, crossed a pretty exciting hurdle just recently. Like Alex Lupsaska and some of the folks at OpenAI had a pretty important breakthrough in physics where they used language models to figure out a lot of the key lemmas and theorems for it. And so, you know, I would have said, I think, the next big breakthrough is getting to a point where, like, actual science and actual discovery is happening largely powered by AI. And I think we're effectively getting into that. I think we'll see much more of that this year. I think to your point on impact, yeah, I think it'll be some time until the average person feels the impact of us proving new theorems.

2:28:43Ivan Zhao:But obviously the long term of all of this is extremely powerful, right? I mean, we're going to be discovering new medicines. We're going to be unlocking big breakthroughs in biology, material science, nutrition, and so on and so on. and all of this comes from a lot of the same science. I think I very much think of it as a, you can call it like a proof of concept or like an existence proof that it is possible. And solving some of these very difficult novel math and physics and algorithms problems is that there are lots of ways that over time that itself will continue to be valuable, but even more so than that, it's obviously just an existence proof that AI can do some pretty incredible things.

2:29:26Patrick Collison:I love it. Yeah, a lot of people get abstract with the medicine science. I like the material science one because I can imagine a much stronger, much cheaper, much lighter carbon fiber and driving a car that's pure carbon fiber for the same price as a Model 3. This is pretty attractive. That's pretty tangible. I think the average American consumer is going to get behind that. Get extremely excited about that. Get extremely excited about that.

2:29:49Ivan Zhao:I'm pretty excited about the part of, you know, you have the best pizza that you've ever tasted and accept it's also the most nutritious thing for you because we've just solved taste and nutrition and everything. And I feel like AI will get us there, but that might require a few more. There's probably a few more steps in the middle.

2:30:03Patrick Collison:That's the new AGI benchmark. Get the goalposts. Get the goalposts. I'm moving the goalposts. AGI will be here when I can have a pizza that tastes amazing and also is fully nutritious. Thank you, Scott Wu. Have a great day. Great to see you guys. Always fun to move the goalposts with you. We'll talk to you soon. Goodbye. It's an honor to move the post. Let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI. Own the data platform that powers it. And without further ado, we will begin our Lambda lightning round with Rune. Look at this new effect.

2:30:46Patrick Collison:Oh, yeah. We're getting new effects going. Welcome to the show. Ooh, look at this.

2:30:50John Collison:What's happening?

2:30:51Patrick Collison:That is a beautiful lighting setup. Thank you for joining. First time on the show, please introduce yourself and the company.

2:30:57James Cadwallader:Yeah, great to be here. I'm Runa Kvist. I'm co-founder and CEO of the Artificial Intelligence Underwriting Company. Our mission is to underwrite super intelligence, and we do that by building standards and insurance products for our AI agents.

2:31:11John Collison:Okay. Sounds extremely straightforward and simple. Yeah, plenty of data to build this on. I mean, yeah, how do you even think about it? So the big thing, Derek Thompson was kind of summing up the whole discourse around Citrini. And his takeaway was that everyone can agree that no one knows what's going to happen. So very difficult, difficult environment to be, you know, creating insurance products for. But I'm sure you're narrowing it down to some key initial use cases. So maybe you can talk about where this starts.

2:31:48James Cadwallader:Yeah. Yeah, maybe the first thing to say is that regardless of whether anyone buys an insurance product, someone is always underwriting it. So otherwise, it's just going to be the head of risk at JP Morgan who has to make a go-no-go decision. He also sits with the same problem. Is this going to work or is it not going to work? So the place we start is just what are the risks that are slowing down adoption today? and can an independent third party with skin in the game and visibility across a bunch of companies be able to underwrite that better than any particular head of risk chief security officer might be able to do.

2:32:24James Cadwallader:And like any other risk, when there's no data, there's an initial R &D phase where we don't expect all of these policies to work out well. We expect to lose some money and in the process start to be able to collect the data that allows us to underwrite this more precisely than anyone else.

2:32:41Patrick Collison:Yeah, walk us through some of the example insurance policies, because, I mean, everyone who's followed, like, the AI story and AI race has seen, like, a million different varieties of impairment from, like, the training run didn't work, or the data center was delayed, and that has a financial impact, down to we got sued because of our training data, or someone used our app and didn't like it. There's a million different ways that you can have small or even large settlements or lawsuits. But how do you think about fragmenting the market, finding a landing zone, a beachhead?

2:33:22James Cadwallader:Yeah, totally. So you start from what are the very real concerns of slowdown adoption today. Let's take one. We just announced the world's first insurance policy for an agent last week with the Lemon Labs. They are trying to be on the frontier. Thank you. they're pioneers of security and safety. They're trying to be on the frontier of giving assurances. The things that hold up adoption for them are things like hallucinations that lead to financial losses. So everyone has seen the kind of Air Canada example leads to financial damage. Data leakage continues to happen, you know, see on a weekly basis.

2:33:56James Cadwallader:Open Claw is the latest group of that. You don't want your agents to give medical advice. Sure. And so those are also some of the kinds of things that are covered. So mostly at the application layer today. And then we think as insurer appetite grows, eventually our mission is to underwrite superintelligence. Eventually we think some of the kind of risks that look a little bit more like private nuclear energy will also have to be covered by insurance because these risks cannot sit with no one. There was a grand compromise in 1954 that allowed us to do private nuclear energy in America, which is the Price-Anderson Act.

2:34:33James Cadwallader:It's basically the government saying, hey, we really want some private nuclear energy. That'd be awesome. But also any particular private company cannot carry the risk if something truly goes wrong. So we're going to require an insurance scheme. That's going to be our way of putting the market to work to manage this in a way that's just pro-business, pro-getting this adopted.

2:34:49John Collison:The government has always effectively been the insurer of last resort in some ways, right?

2:34:56James Cadwallader:Whether it's formal or not, the government is always the insurer of last resort. Take COVID. Who's on the hook for that? Ultimately, the government has to step in. So the question is, can you formalize that a little bit more and say, at what limits of liability is the government on the hook? And up until that, who's on the hook for that?

2:35:10Patrick Collison:Got it. Okay. So walk us through the chain of how insurance actually works. I understand 11 Labs comes to you. And then are you drafting a policy with a specific risk profile, payment premiums, and then you're going out to the JP Morgans of the world and having them buy that and invest that? Does this float? Is this tradable? Can a retail investor get allocation? How does that work on the long tail of the financialization?

2:35:39James Cadwallader:Yeah, eventually this will end up on Robin Hood. But let me walk you through how it looks today. So today there are two steps at high level. First is certifying against the standard as a way to unlock insurance. So historically, the way every market's been unlocked is that the insurers want to know that the risk is well managed. The head of risk, Jeff Morgan, doesn't want just financial coverage. He wants to make sure that there's no incident that gets him fired in the first place. And so we've developed a standard. It looks a little bit like a Moody's framework or a SOC 2. So that is all open source and public.

2:36:10James Cadwallader:It's 50 requirements that any AI frontier company must meet to meet the standard. And as part of that, we run a bunch of technical tests, basically crash testing, red teaming, as you might call it here, which gives us a score. And we give them pass-fail and certificate. And then the score feeds into a policy that we've designed with some of the leading insurers, where a company like Elin Labs gets to specify, hey, what are the top three, four, five risks that hold up adoption? They buy a policy for that. And today that risk is helped by traditional insurance companies. Again, this is actually all about trust.

2:36:43James Cadwallader:So you really want the old insurers to have it on their balance sheet. They always pay. Over time, as we move into this kind of like Chernobyl-type risk, we will run out of private capacity. We will have to, at some point, create catastrophe bonds. Those will be traded on the public market, probably not on Robinhood, but by the most sophisticated investors. That is the ultimate way to build enough market capacity to cover the tail risk.

2:37:07Patrick Collison:Yeah, makes sense.

2:37:08James Cadwallader:Very, very fascinating.

2:37:09Patrick Collison:What does the business look like today? This feels like high stakes work, but is it capital intensive? Is it, do you need a thousand insurance agents at some point? Like what's the team like? What's the fundraising like? What's the business like?

2:37:25James Cadwallader:Yeah, totally. So the way to, if you're really thinking about this long-term, the way to unlock the insurance market is to get the standard universally adopted. And that is kind of what allows everyone to say, hey, this risk is well managed. We can now start to price it. And so we have, for the standard, we have about 100 security leaders from the Fortune 1000 who meet with us every six weeks to input into the standard as representing their interest. And now you're having some of the leading AI companies like Eleven Labs, Intercom, UiPath, more to be announced soon, that have put themselves forward to say, hey, we're pioneers.

2:37:58James Cadwallader:We'd like to have an independent audit to prove that. That's step one. So that's what most of our work is focused on today. and then on the insurance side the way to start is to partner with existing insurers that bring that trust, credibility they're frankly so old school and that's what brings trust here they're not accomplished that's the whole point it doesn't really work

2:38:19John Collison:it's like okay who's actually backing this policy and then it's like oh a company created three months ago

2:38:25James Cadwallader:it's like me so it's actually quite capitalized to get started we raised$15 million from that Friedman last year and...

2:38:37John Collison:Almost ran into the goalposts. There you go. I'll move them again.

2:38:43Patrick Collison:Well, thank you so much for stopping by the show and giving us the update.

2:38:47John Collison:Yeah, a lot more questions. As there are new kind of crises around agents, feel free to pop back on and talk about it.

2:38:55Patrick Collison:Amazing. We'll talk to you soon. Good to meet you, Arun. Good to meet you, Arun. Let me tell you about CrowdStrike. Your business is AI. their businesses securing it. CrowdStrike secures AI and stops breaches. And without further ado, we have Rainier Pope from Matt X in the Reistering Waiting Room. Welcome to the show. How are you doing?

2:39:14John Collison:What's going on? Doing great. Very happy to be here.

2:39:17Patrick Collison:Thanks so much for hopping on. It's your first appearance. We'd love an introduction on yourself and the company to kick it off.

2:39:23Stefano Ermon:Yeah. So happy to be here. I'm Rainier. I'm CEO and one of the founders of Matt X. We are a company that makes the best chips physically possible for large language models. So we've been doing this for about two or three years. Before that, I was myself, I was at Google for about a decade working on large language models, worked on the TPUs for a bit, worked on some other hardware projects. And really as part of that, what we saw was that there was this, if you really want to make the best chips for LLMs and LLMs or this big up-and-coming workload back in 22, If you want to make the best chips for LLMs, really the best way to do it is from scratch, blank slate design.

2:40:04Stefano Ermon:So designed for large matrices, very low precision, very low latency. And so my co-founder Mike Gunter and I, at that point in 22, decided to leave Google to start MadX, where we're doing exactly that. The day we're announcing Maddox 1, this is a new chip which simultaneously offers better throughput per square millimeter or throughput of a chip than any other product in the market, while at the same time offering lowest latency, latency that is comparable to the best, which is Croc and Cerebus.

2:40:37Patrick Collison:Yeah. What are the various tradeoffs in custom silicon design these days? Is it just, I mean, at the highest level, is it just flexibility and speed or cost, size, wafers? How do you think about the design space? And then I want to know how you actually narrowed in on your particular decisions.

2:40:57Stefano Ermon:Yeah, so generally there's some kind of performance per something. And so let's analyze those pieces. The two different aspects of performance are what is the throughput and what is the latency. So how many users simultaneously can I support is throughput. and then latency is for one user, how fast is the experience? Both of those matter. And then on the per something, like per dollar, how much does the chip actually cost? And then per watt, which is like what is the power bill of the chip? So those are all combinations of those two numerators and two denominators are the things we care about.

2:41:31Stefano Ermon:What we see in the market today is that the number one constraint is just the throughput per dollar and the throughput per watt. So these frontier labs have so much demand for compute, serving all of these trillions of tokens per day. And so the cost and the economics is the main constraint. There's only so many square millimeters of silicon wafer being produced every year. And so given that constraint on how much silicon there is, can we maximize the number of tokens and then maximize the intelligence of the models coming from that wafer?

2:42:08John Collison:What does the go-to-market look like? Are you already sold out? Who are you targeting early on? How do you scale? All that stuff.

2:42:19Stefano Ermon:One of the places where we've seen the most interest in our product is from Frontier Labs. This is coming from a combination of

2:42:29Stefano Ermon:they are the ones who are driving really all of this demand and are so much constrained on cost as well as silicon wafer supply. But then also they are the ones who are doing these reinforcement learning training workloads which are very, very latency sensitive. They have to roll out long rollouts in a very long loop. So that's where we've seen the most interest. One of the things that shows up there is that when they are looking to make place an order, it is an order on the order of gigawatts or something like that, which is massive volumes. So one of the things that we're actually very excited to be able to do now with this raise that we've just announced is help ramp up the supply chain in order to be able to deliver gigawatts a year of volume, which is a massive volume to be able to deliver.

2:43:18John Collison:When you were initially thinking of starting the company, pitching it to investors early on, how did you answer the question around NVIDIA's various moats or kind of strategic advantages? you know, think CUDA, all that stuff.

2:43:35Stefano Ermon:Yeah, I think it's really interesting. Like CUDA is for NVIDIA simultaneously the biggest strategic advantage and also a constraint because their promise that they make you is that you can take a CUDA program written 10 years ago and it will run on the next generation NVIDIA GPU. Jensen goes on stage and promises this. It is so valuable for them. And yet at the same time, it means the next generation GPU has to look just like the GPU from 10 years ago. So it means things like the numerics can't change, The way the cores in the chip are connected to each other can't change. The actual memory architecture can't substantially change.

2:44:10Stefano Ermon:All of these things are kind of locked in by the programming model that they designed more than a decade ago for general purpose parallelism. And so this is where we've seen the biggest differentiation. If they wanted to say, well, we're going to completely give up our CUDA approach and start a new generation of chips, maybe they could do that. they would lose all of this lock-in that they have, but then at least they would be on a level playing field with us. But that's not what we see. Really, we see them being committed to their trajectory. The CUDA lock-in is very valuable for sort of the mid and tail of the market where people are so sensitive to the software cost.

2:44:47Stefano Ermon:But really at the head of the market in the Frontier Labs, the software is not the main cost. The hardware is the main cost. And so if you're willing to rewrite your software, maybe you can actually switch to a more efficient hardware like us.

2:44:57Patrick Collison:and it's getting easier to rewrite software.

2:44:59John Collison:As you plan your business, how are you thinking about bottlenecks? One month it's energy, the next month it's chips, then a lot of concerns around TSMC right now. How are you kind of planning?

2:45:16Stefano Ermon:Yeah, so I think these bottlenecks are real and are going to stay for a long time. The big bottlenecks that you see in the manufacturing supply chain are on logic dyes from TSMC, and then memory dies from Hynix, Samsung Micron, and then manufacturing of racks and so on. Given these bottlenecks exist, what you would like to do as a consumer of such things is you want to get the most bang for your buck, so the most performance out of every square millimeter of silicon. That is what has been our focus. The flops per square millimeter, the 4-bit precision multiplies you can do per square milliliter of silicon is higher in our product than any other product.

2:45:58Stefano Ermon:And so, you know, as the price of every silicon wafer goes up, you can do more with it on our solution than it does.

2:46:06Patrick Collison:I assume you're on the most, I mean, you've mentioned this, you're selling to the frontier labs, running frontier models, probably on the most leading edge chips, the most leading edge fabrication nodes. Is there a world where it's valuable to say, hey, we have some lagging edge capacity out there. What if we go design custom silicon that runs on the last generation Intel node that's not the line out the door for capacity, and then I'm not competing with you? Does that not work? Is that not possible? Or is that just a completely orthogonal business to what you're building.

2:46:47Stefano Ermon:That approach is possible. It is maybe more of an approach for a player with rich pockets rather than a startup. In that every different process you target, it costs you another $20,$30,$40 million of development cost. If you're going to bet all of your eggs on, put all of your eggs in one basket, you should put it in the leading edge node.

2:47:07Patrick Collison:In the best basket. That makes sense. Talk to me about other trade-offs at TSMC. I mean, Cerebrus is famously wafer scale. What is the tradeoff on, like, size of dye these days? Yeah.

2:47:23Stefano Ermon:So, I mean, there's a tradeoff of size of dye and then also of memory architecture. So size of dye, Cerebrus is the outlier. Almost everyone else has converged on reticle scale, which is the largest sort of standardly produced TSMC chip. We're in that same category of, like, we're in the standard bucket there. That avoids a lot of the physical risks that, you know, when you look at Cerebris, they've had to spend all this time on dealing with just like bending and all these uncomfortable physical constraints that we don't want to deal with. So the reticle-sized chips, but then the other bigger thing is which memory technology do you use?

2:48:01Stefano Ermon:Historically, there's been like the HBM-based players, that's Google, Amazon, NVIDIA, and then there's been the SRAM-based players, which are Cerebris and Grok. SRAM is small but very, very fast. And so the very, very fast is good if you want to run low latency. You can put your model weights in SRAM and you get the best latency in the market. That's what Grok and Cerebris have done. But the reason they haven't sold out in the market is because there's not enough space in the SRAM to store all of your long context KV caches. Sure. And so one of the things that, like this is the reason why the HBM-based players like Google, Amazon, NVIDIA have won, is because the HBM is actually essential.

2:48:43Stefano Ermon:But it's actually possible to marry both of these approaches and put them in one chip. And that is what we're doing with MatX1. Okay. And so it, curiously, I mean, it doesn't just give you the best of both worlds. It actually beats any alternative on throughput. There's this curious effect where when you have your weights in SRAM, you can actually get better mileage, better usage of the HBM in return. We think there's actually the way where the market in general will move over time.

2:49:12Patrick Collison:Okay. Help me understand the tradeoff continuum around flexibility of model. I mean, imagine on NVIDIA and NVL 72, I can sort of run any model as long as it fits and works and is trained properly. And then Talos is like, these specific weights on the chip, you can never change them whatsoever. And then there's something in the middle. How much flexibility do you think is important? How much flexibility are you planning around? And how do you think about sight lines? because I imagine that the delay between the final architectural design to chips in data centers is still a year, 18 months, something like that.

2:49:58Stefano Ermon:Yeah, I mean, there's all of these manufacturing and then deployment times that make it take a long time. In general, I would say that from sort of pencils down on chips to like, when is the last time you're using it? The chip is going to be in the data center itself for like three to five years. And then there's maybe, as you say, a year, a year and a half of deployment time in advance of that. So you want your chip to be relevant for a five-year time span. So you need to pick a point of specialization which you think is here to stay. For us, that is very large matrices. And in fact, really large matrices together with a splittable systolic array, which is a piece of technology.

2:50:36Stefano Ermon:But very large matrices is the theme that we started with. And this is just a recognition of over time, models have been growing. They grew a ton with LLMs and they're continuing to grow. and if you specialize for that you can get big efficiency wins on the matrices themselves. Now we are still very general purpose programmable in terms of the vector unit similar to NVIDIA we have this vector unit that you can run any instruction on like add, multiply, subtract, divide all of those things and so that gives you the it's trying to put a good amount of flexibility but in a way that only costs like 5-10 % of the cost of the chip overall.

2:51:14Okay.

2:51:15Stefano Ermon:Funding news. Give us the news. How much did you raise? What happened? So we're happy. We have raised$500 million. This was around... Yes. Boom. Who'd you raise it from? This was led by Bainstreet and Situational Awareness. So Situational Awareness, that's Leopold Ash and Brenner's fund. If you've been living under a data center, that's his fund. Exactly. You might have heard of it. So he really sees just like the big picture of where this space is going. And he recognizes like just how much demand there is for silicon. And then on the other end of the spectrum, Jane Street, they are expert technologists.

2:51:56Stefano Ermon:They know everything about what exactly is required to build a product like this. And they know what good is in a product like this. So we're really happy to have these like strong experts here. This sort of mirrors what we see inside the company as well. We have a wide range of experiences across hardware, software, and ML. And then even in the rest of the investors who are participating in our round, we have renewed participation from our previous investors. This is Spark Capital and NFTG, as well as a range of folks such as Patrick and John Collison, experienced ML people like Andre Kapathi.

2:52:38Stefano Ermon:and then even participation from the supply chain like Marvel and all that.

2:52:43Patrick Collison:Did you let any normies in? They're just the most elite people in the world.

2:52:49John Collison:Yeah.

2:52:49Patrick Collison:I mean, just one mouth breather, please.

2:52:55John Collison:Probably the hardest lineup of investors I've ever heard.

2:53:00Patrick Collison:It's amazing. I'm extremely excited for this and excited for it to get into the world.

2:53:03John Collison:You have to win on such massive scale. Otherwise, you'll bring dishonor to all the industry legends.

2:53:10Patrick Collison:No, thank you.

2:53:11John Collison:No, it's really cool to hear your perspective and approach to everything. And I'm sure you'll be back on the show this year. So congrats to the team.

2:53:18Patrick Collison:Yeah, we'd love to have you back. Thank you so much for taking the time. We'll talk to you soon. Goodbye. Let me tell you about Vibe.co, where D2C brands, B2B startups, and AI companies advertise on streaming TV, pick channels, target audiences, and measure sales just like on Meta.

2:53:32John Collison:Have we had an investor lineup like that before? The Jane Street situational awareness co-lead and then just down.

2:53:40Patrick Collison:Well, I mean, situational awareness is a new fund, has not led that many rounds. I know, but I'm just saying you go back. Maybe it turns into a spray and pray fund. You never know. Maybe Leopold says, yeah, I'm just going to write$5 million checks to every company. Who knows? Anyway, we have our next guest in the roost room. We got Standard Intelligence. How are you doing?

2:53:59Rune Kvist:What's going on? Hey, I'm Devanche. Doing pretty well. How about you?

2:54:04Patrick Collison:We're doing fantastically. Thank you so much for taking the time to come on the show. Since this is the first time on the show, I'd love an introduction on yourself and the company.

2:54:11Rune Kvist:Yeah, so I'm Devanche. I'm co-founder of Standard Intelligence. We pre-train computer use models, basically. So basically, the thing people are doing is they're training on screenshots and train of thought traces, and we're just like, what if you train purely on 30 FPS video?

2:54:32Patrick Collison:What actually goes into the training data? Because there's a lot that you can do on a computer. And I feel like if you've never trained on Ableton and it just comes randomly, like are you actually going to be able to learn Ableton from just playing in Premiere Pro and Word and, you know, paint or something or Photoshop or whatever? Like how are you thinking about the transfer and like what's actually in the training set? actually just zoom out and talk about the process in more depth.

2:54:59Rune Kvist:Yeah, so we have two splits of data where we have this small contractor split. The thing that we did was we made this app that people run on their computer, and it records their screen and logs all their key presses and all their mass movements. And we're running that all the time. And then we also have this much, much larger kind of unlabeled data set of basically every video that we could possibly find that we're allowed to use. on the internet of computer use. And so, yeah, we trained a model to label that big set from this small contractor-only set. And the goal is to just be able to train on all of it and train this kind of general model that is able to generalize to basically anything that you can do on a computer.

2:55:45John Collison:What kind of limitations do you have on who can install your software to capture that data? if I'm a company, I feel like I have to have a pretty high degree of trust in you guys to let my employees. Is that something that's more of like a partnership? Yeah.

2:56:03Rune Kvist:So right now it's like us, like we're recording our own screens all the time. Plus, like we have some number of contractors and we get to like pay them, you know, somewhat less because they're not doing like active work for us. It's more like passive screen recording.

2:56:15Patrick Collison:Sure. Yeah. And then are you sitting on top of some sort of foundation model brain for reasoning chains and sort of like the LLM piece of the puzzle? Or is this a model that's kind of lives? You're not right now.

2:56:29Rune Kvist:We're not at all. Like the model that we released is, I suppose like demoed is like entirely trained on this kind of 30 FPS video in, and like, you know, typing and mouse movements and things like this up. Okay, so prompting. At some point it's possible. Sorry again.

2:56:46John Collison:How much longer will I have to fill out forms on the internet? I should try to estimate how many times I've entered the same information just over and over and over and over. John Collison talked about it on our show today and was talking about it on his own show. I think with Ben Thompson talking about like at what point can you just take a link and say like, hey, please buy this. And then it just does it for you.

2:57:13Rune Kvist:Like pretty soon. I think like that kind of use case is just like under six months away, depending on what like exactly you mean.

2:57:21Patrick Collison:Yeah. I mean, in terms of actual deployment, I imagine that this would be something I personally would probably want deeper as more of like a tool that's called from a consumer LLM app. Is that how I'm correctly thinking about this? Or do you think they'll actually be like where you jump straight to a consumer?

2:57:42Rune Kvist:um it's like yeah i think in the short term the kinds of people that are particularly like you know cool to sell to are like you know mechanical engineers doing cad where like uh they can press the the tab button like software engineers press tab and curse and have their next like you know minute or two minutes of manual work um done and and we showed that in the the kind of gear extrusion demo where like you have this gear and you're like extruding faces and that's just like a very very common thing that you do in CAD. And I think there's a more general thing where you can think of computer use as a tool call or you can think of it as just the thing that you do for knowledge work.

2:58:24Rune Kvist:And I think we're just in a place where we can scale computer use on its own. It's not impossible that we'll initialize from LLMs or for example use text training to make the model smarter in text voice so it can fill out forms better. But it is not the goal of the company that people have. You have Claude call this as a tool call. The goal is to not use your computer or use its own computer in general.

2:58:55John Collison:Talk about your experiments with self-driving and does that work potentially apply to robotics more generally?

2:59:03Rune Kvist:I think this general pre-training thing or like, you know, labeling a bunch of unsupervised data with actions and then training on that like labeled data, this like inverse dynamics thing works very, or like I expect it to transfer very well to robotics. Self-driving in particular, it was kind of, so Neil who works at SI was like, okay, we have this action model and his friend had a comma and so there's a comma like joystick mode where you can like control the steering with arrows and so we were like okay well if it's a general computer use model surely it should be able to you know control a car because that's just like a thing that you do on a computer it's like video in you're seeing it on the screen and then you can press the left and right arrow keys to steer and obviously we originally didn't really expect this to work and then it just like worked much better than expected we find an hour of data.

3:00:08John Collison:On how many hours? Three hours? One hour. 50 minutes. And you're able to just fully the system could just navigate around SF.

3:00:19Rune Kvist:I mean, sorry, navigate around South Park. It's not a general self-driving model. I would not recommend sitting in this car and just letting it do whatever it wants. But yeah, it's pretty cool.

3:00:32John Collison:I take the wheel. don't make mistakes

3:00:35Rune Kvist:we are not a competitor do you think

3:00:41John Collison:the sport coat is like the next it apparel item because it looks fantastic here the chat loves your sport and I just feel like that could be the middle ground between Wall Street and San Francisco as a sport coat yeah I don't know I really like this

3:00:59Rune Kvist:I got it from like Bonobos and you need a square there you go that's great i think i think i like dressing up like at least a little bit and it's it's fun that's good okay back to the business uh yeah i i want to know about uh it feels like

3:01:15Patrick Collison:you're you're training a very generalized model uh what are you learning from the previous product launches where you know we had this chat gpt moment and then i don't even remember what people were just kind of chatting with ChatGPT back and forth. And they started using it kind of as a Google replacement. And then that kicked off the whole like Google's cooked narrative. And then with the Studio Ghibli moment, it was really the launch of like a better diffusion model with some reasoning in there, I think, and stuff. And then people were just like, this is a Studio Ghibli creator. And then they found that niche of like, it's really good at creating cartoons.

3:01:54Patrick Collison:It's not quite style transfer, but that's what it does well. How much do you want to just like turn a wild open model loose and then hope that someone finds a killer app versus like you kind of know that this is going to kill in CAD and you're just going to launch like cursor for CAD on day one and then like go from there?

3:02:14Rune Kvist:Yeah, I think there's I think the answer is like some combination. I'm like, OK, short term CAD design work somewhat generally are things that the current model is just like totally can't do. like LLMs are just like, or like anything that is an LLM harness is just like really, really bad at CAD, for example. And so that seems like a, okay, we know what to do there. We like, you know, can just scale up this model. We have a bunch of Blender data, a bunch of 3D modeling data in general, and we can scale up CAD. And then also, yeah, I like, I'm quite excited to release a more general, you know, tab model for people to play around with and like figure out what it's particularly good at.

3:02:54Rune Kvist:And so it's like when I'm asked what commercialization plans are, it's like we have some reasonable idea of what the first steps are. But there could just be this massive thing once people start playing with it at scale.

3:03:07Patrick Collison:So you're training on video frames, 30 FPS video, correct? Yeah. I was told by an anonymous poster on X by the name of Rune that text, in fact, is the universal interface. Was I lied to?

3:03:22Rune Kvist:Yes. Whoa, shots fired.

3:03:26Patrick Collison:Explain, elaborate. Like, why doesn't this just collapse down to text? Why don't I puppeteer CAD from text? Like, how does this all play together?

3:03:35Rune Kvist:Like, okay. I think it is at some point in the arbitrarily long future, if we only use text models, we could force most things to be text. I think there are just a lot of things that are much more native when done from a computer use. Geo-wise are designed for humans. They're designed for humans to use. We have this massive long tail of things on the internet that are entirely undoable by LLMs. For example, when I do ML engineering, because most of my time is just spent doing this grunt work of engineering, and it's a lot of looking at graphs and analyzing graphs and comparing loss curves or something.

3:04:26Rune Kvist:You can do this in text, but it's just a much larger pain than doing it in this kind of native interface, which is video. And I don't know, there's a reason why humans don't interact with the computer purely through text. It would kind of suck. For example, we have like the concept of, like video has the concept of time in a way that text doesn't.

3:04:49Patrick Collison:Speak for yourself. I got green text right here, black background going.

3:04:53John Collison:If this can eliminate YouTube tutorials for software, that's a killer app. Is this, yeah, is that anything? It's not just going to eliminate the tutorials.

3:05:01Patrick Collison:It's going to eliminate the whole process. Because you don't need the tutorial if you're just like, just go do the thing that I need you to do. But yeah, I mean, you're obviously in this entire phase for a while.

3:05:12John Collison:How are you thinking about go-to-market in general for the underlying technology?

3:05:19Rune Kvist:Yeah, so I think, like, as I said, there are the kind of short-term, like, CAD design use cases. There's the tab model, which we want to just give anyone a kind of general thing of, you know, in cursor, you press tab and it completes your next edit or whatever. What if you could press tab and it completes the next five and then 10 and then 60 seconds of what you would do in your computer? And then I think longer term, it's just like, you know, we're training a general model that is able to do useful work and you'll be able to send it off with a prompt to do work. And then there's a very interesting thing where the data that we're training on is very...

3:05:58Rune Kvist:There's a bunch of error correction built into it. So when you have a bunch of data of humans doing things, a lot of the times the humans make mistakes. And then they have to correct those mistakes. And you don't get that with text. Because most text on the internet, you don't get to see the process of messing up and then fixing it. And so, yeah, I expect there to be a lot of native, no RL just prior of doing the self-correction thing properly. So you can get it to go do something for 10 minutes and it'll try something for two minutes and then mess up slightly. But it knows how to fix that over and over again until it's gotten to a solved state.

3:06:44Patrick Collison:Yep. Very cool. Well, congratulations on the launch, and thanks for taking the time to come chat with us. Thank you for sport coding.

3:06:51John Collison:We need some sport coats around the office. You need something in between. You got John over here formal. I'm doing casual Friday on a Tuesday, but a sport coat perfectly in the middle. It was great to meet you, and come back on soon. Thank you.

3:07:05Patrick Collison:We'll talk to you soon. Goodbye.

3:07:08John Collison:Back to the timeline.

3:07:09Patrick Collison:Back to the timeline.

3:07:10John Collison:There was an individual who accidentally gained control of 7 ,000 DJI vacuums. He was just vibe coding.

3:07:21Patrick Collison:Amazing.

3:07:22John Collison:And accidentally found, according to Investment Hulk, the CCP backdoor. He wanted to control it with a gaming controller. And then he got control of everything. That's so crazy. This is why I've been deeply concerned with letting any foreign adversary flood our country with a bunch of robots. I think we should avoid it. I thought this was funny earlier. Hegseth says he'll order random pizzas to throw off the monitoring app.

3:07:54Patrick Collison:Oh, yeah.

3:07:55John Collison:I expected something like this to happen. It's kind of silly that everyone has a dashboard up and can tell when things might be getting a little more tense in the Pentagon. So, yeah, give them a budget of, you know, a few hundred thousand dollars a year and just order pizzas at random times.

3:08:11Patrick Collison:For the record, this is a joke and he is joking. But I do think they could throw it off. Potentially. You never know. They could throw it around.

3:08:21John Collison:HubSpot acquired Starter Story. Yeah, this is very exciting. Very cool. Yeah, Starter Story.

3:08:26Patrick Collison:Overnight success. What decade he's been doing this?

3:08:28John Collison:I believe Pat, the founder, had just posted.

3:08:32Patrick Collison:Yeah.

3:08:32John Collison:He posted something. He was like, he said HubSpot should acquire Starter Story. And then like two weeks later, it was done.

3:08:38Patrick Collison:Wait, really? Oh, I thought that was from like years ago.

3:08:41John Collison:Oh, maybe it was. I thought he posted that a long time ago. No, he said September 23, 2025. Oh, wow.

3:08:48Patrick Collison:Not long. Yeah, a couple months. A couple months ago. A couple months.

3:08:51John Collison:HubSpot should acquire Starter Story. The SEO ship is sinking. In my opinion, HubSpot needs to pivot way harder to video, specifically YouTube. Yeah. I'm biased, but Acquiring Starter Story would take their YouTube game to the next level. I love it. He was quoting Brian, the CEO, saying, Dear Founders, it's a good time to sell your company. Love Brian. Anyways, there's a bunch more stuff in here, but we will get to it tomorrow.

3:09:20Patrick Collison:ArenaMag is out. Go check out issue number seven. They're on Substack now, arenamagazine.substack.com. Go check it out. One more post for you.

3:09:31John Collison:Deep Dish Enjoyer says, I don't see what the point of shoveling snow is when AI agents are going to commoditize burrito taxi services by 2028. That's a good excuse.

3:09:42Patrick Collison:Leave us five stars on Apple Podcasts and Spotify. Subscribe to our newsletter at tbpn.com.

3:09:47John Collison:Have the best evening of your entire life. We love you.

3:09:52Patrick Collison:Goodbye. Goodbye.

3:09:55John Collison:Nice work, brothers. I'll see you on the next one.