The Exchanges

Q Okay. Um, so, yeah, what, what else was, uh, like, key in the thesis, or maybe potentially overlooked, that you think people, uh, should be really focusing on?

A I think the problem, A, the first thing, the most important thing is just the labor market dynamics. Uh, we've just been in a really weak labor market for a while, uh, and that's before these things roll out, but then you put that together, uh, with the fact that, you know, we just have a very structural, uh, environment where what, what is the thing that drives our entire economy? It's wages. Uh, most of those wages that are ultimately driving all the discretionary spending is coming from, The white collar worker. And the problem with that is that we're now entering this place where you made all these assumptions on like loaning money to all these companies, uh, you know, to mortgages and everything else, like white collar economy is our economy. If you all of a sudden just take a leg out of that economy, it has a contagion effect into basically every asset, uh, in the world. And so that I think is the part that people haven't thought about. Cause when, you know, people were making these loans, No one ever conceived of a world in which, wow, okay, now like white collar jobs are in sort of permanent decline, right? If that's at two percent a year, then I think we can skate through, but if it's at four or five percent a year, then, you know, we have to, we need action a lot more quickly.

AI assessment note: “that I think is the part that people haven't thought about.”

Q are not writing your piece saying like, you know, this, we, we believe we will stake our entire reputation on, on this sort of, uh, narrative, but what do you think? What, what, how much did you pay attention to like the nineties, early 2000 internet predictions? What do you think they got Wrong. Why is this time different in terms of how a new technology will diffuse the economy?

A Uh, I think the difference is if you just plot, uh, what's happening to technology, it's all just going exponentials. These are all just continuous timelines of, like, we have microcomputers, we have the internet, we have mobile phones, and today, you know, we have very, very powerful AI. And so I think most of the predictions that you, you ticked off there, it's kind of interesting. I would, you know, just looking at them today, you know, I would say they couldn't really happen until you had proper AI, because, like, if you have the ability to Just freely, you know, have commerce the way you do today doesn't work if you still have to do all the work. Ultimately, like, you have to go in, you have to log in. Think about the amount of friction there is in buying a product for most people today, right? You still have to go to the website. You have to put your credit card in. It's all work. We only have gotten to kind of the tech required for those predictions, I think, this year. And that's why this is the year that I think it really begins, because now it is completely seamless. You just, and no one's really doing this yet, but it's going to happen, I think, you know, in the next six months. Is just tell your agent, you know, tell Gemini, tell ChatGPT, go buy these things, it has your credit card, and now that world that they were describing is truly going to come to pass.

AI assessment note: “they couldn't really happen until you had proper AI”

Q Yeah, it feels like, was this the most, like, stretched or controversial prediction?

A It seems like it was certainly the one that got, you know, that's getting the most chatter. Uh, and I think we did it for a reason. We wanted to be a little provocative in thinking, thinking it through because, you know, it's an amazing business and they're gaining a bunch of market share, but the fundamental idea that you're, cause what did the lock in, right? Like, do the drivers have lock in on DoorDash or on Uber? Not really, right? They're, you know, most, most drivers are doing Lyft and Uber, so they're, they're not locked in. The real lock in, the real business value, The franchise value of an Uber or a DoorDash is the customer lock-in because the customer gets comfortable. They've got everything saved. They want to hit a couple buttons. They don't, they don't price shop. Agents are happy to price shop as much as possible. And so if you take that away, then it's a real problem for businesses that, you know, are ultimately built on customer lock-in.

AI assessment note: “It seems like it was certainly the one that got, you know, that's getting the most chatter.”

Q So do you think the future is, uh, what, Anthropik's head of sales position in France? The company, uh, will be spending 530,000 euros per year, the government will get 340,000 euros, and the employee will get a 190,000. Is that, is that the level of taxation do you think we're, we're headed for?

A Uh, I think when we're at, you know, France's level of government spending, then, you know, the math probably means roughly that. Uh, I, I would say that, you know, government spending would be at France's level. I'm guessing like, you know, five, seven years from now, if this, if this scenario kind of comes to pass. Uh, and so I think we'll head there over time, but I think it's less a question of the percent of spending and how much goes to the employee versus goes to the government. And ultimately what is the size of the total pie? So the bet here is that the pie If done properly, can just increase multiples of what it is today, and, and thus, you know, there's, it's just a win-win.

AI assessment note: “when we're at, you know, France's level of government spending, then, you know, the math probably means roughly that.”

Q What do you think is most misunderstood about Hyperliquid today?

A Oh man, there's so many things. Uh, most misunderstood. One thing I would say is that it's, well, the, it is a network. I think people think it's like a purpose exchange. Some people think it's like a centralized thing. It's really not. It's a blockchain. The validator sets permissionless. They're currently 24 validators. You know, anyone can spin one up. It's like top, top 24 by stake. Uh, the, the validator set, you know, every validator executes every transaction, including all the orders. So like this kind of like base layer of like, just like what it is, I think is kind of lost because people are so They're focused on the product basically, which, which maybe is good. Maybe it's nice that the tech is abstracted from users.

AI assessment note: “I think people think it's like a purpose exchange... It's a blockchain.”

Q capabilities of the model? You mentioned that you've seen some fantastical scenes. I'm interested to know about specific, like, uh, specific breakthroughs that you've noticed that Sora to the model is particularly good at. I noticed one about, uh, reflections being great. Obviously people love the cameos, but what, what has surprised you in terms of just like, technically the model can do something now that it couldn't do before?

A This model is a huge leap forward in terms of physics IQ. So pretty much all past video generation models really struggled with prompts that, you know, involve like backflips, gymnastics routines, et cetera. And this is really the only model that exists today, which can reliably handle these kinds of really complicated dynamics. Um, one of the big features that people have really loved on the app so far is the steerability of the model. So, you know, if you give it like a really simple text prompt, that's maybe even only a few words. This model is really good about kind of telling a coherent story with like a beginning, middle, and end, and doing this like automatically in a way that doesn't require like a lot of direct steering from the user. If you want to like go into, you know, a ton of detail about exactly how your prompt should be laid out and how the story should unfold, it supports that too. So it can kind of meet you wherever you're at in the creative process. But really this model is just like so hyper steerable and it's like just vastly higher physics IQ just makes it able to do things that were like not possible a few months ago.

AI assessment note: “This model is a huge leap forward in terms of physics IQ.”

Q You have been extremely busy. Give us the update. What's new in your world?

A Gosh, what's new on our world? Well, the, the big news is we, we launched dynamic tech defense reform, which I know we're going to get into today, um, which is a huge initiative on behalf of all of defense tech, all of the defense tech community. Um, but also it's, I just returned from Washington last week, had a host of conversations across the Pentagon, across Congress, and it feels like for the first time in probably 10 years since I've been doing this, that everyone is aligned, that we need to see Major, major changes happen to defense procurement, and not only is that sort of things that people talk about, but it feels like it's finally going to happen in this year's NDAA, um, and that we're on the one yard line, um, sprinting to a finish that I think will, will hopefully, um, end up great for, for all of the, the companies in the defense tech community, so we can go further into what's changing, but I always, like, I, I feel like I'm bringing, like, actual good policy news, like, probably, like, It's like the one thing that everyone in Washington can agree on, Republican or Democrat, uh, you know, House or Senate, Pentagon, everyone is aligned.

AI assessment note: “the big news is we, we launched dynamic tech defense reform”

Q Page has said internally at Google many times, I'm willing to go bankrupt rather than lose this race. Everybody is focused on ROI, but the people making the decisions are not basically saying that, uh, scale, it's all about scale. Uh, you just, they're going full set, right? They're not thinking they're, they're optimizing for winning, not necessarily, you know, uh, in this spreadsheet, optimizing purely for, uh, immediate ROI.

A Yeah. Yeah. I mean, if you're in the position to generate a lot of cashflow, it is kind of the per the game theoretic optimal thing to over invest in a new huge technology. You have to, uh, because if you don't invest in the technology wave happens, you look really stupid. You lose your job. If the technology wave happens and you invest, then you're good. And if the technology wave happens, uh, if the technology wave doesn't happen, but you didn't invest, you don't really get as much credit for that. Because then you're just kind of stable where you are. Uh, and so, it's this, it's this uncapped opportunity with a lot of downside if you don't invest, so people are pushing it. What you laughing at?

AI assessment note: “it is kind of the per the game theoretic optimal thing to over invest”

Q I mean, the fact that you can Buy this in a couple weeks is just insane. Talk about, ah, the, the team's foresight around the, the intersection of glasses and AI, because now it seems incredibly obvious, right? This, like, always on, this live AI. But it wasn't that long ago that people thought these were, like, two different, sort of, like, tech trees, and they didn't see the convergence.

A Yeah, I mean, look, every new important technology needs a new class of devices in order to make it first class, and I think glasses have three main advantages that I think are gonna be, just make them the ideal candidate to be the next major computing platform. One is that they help preserve this sense of presence when you're there with another person. I mean, you take out your phone, you're gone from the moment. Glasses have the ability to bring that back. Two is that glasses, I think, are the ideal form factor for AI, It's the only device type where you can let an AI see what you see, hear what you hear, talk to you throughout the day. Soon it's going to be able to just generate a UI visually for you and your vision in real time. And then the third thing that glasses can do, it's really the only form factor that can bring together, you know, the physical world that you have around you with, um, realistic holograms and blend those together. And, you know, I think it's one of the, Crazier things about living in the modern world is that we have this incredibly rich online world, and you access it through this, like, five-inch screen most of the time. So I, I just think that it's, it's like only a matter of time before these two things are, are basically fully merged, and glasses enable all of that. So that's kind of been the plan all along. I mean, we, um, when we started Reali…

AI assessment note: “glasses, I think, are the ideal form factor for AI”

Q I think David Ogilvie said, like, simple ideas are powerful ideas. That's certainly something I've seen in mischiefs work throughout your entire career, but is there some sort of contrarian or nuance to that concept?

A I think there could be the, the line that we use here is a good idea should slap in one sentence and then slap even harder in three. And that just means there, there are layers, right? Like you appreciate it for the headline, but if you like really dig into it, you're like, wait up, is this thing making fun of me? And do I still like it? And if I buy it, am I a chump or am I a patron of the arts? Right? That's like, uh, The layers are good. The ambiguity is good, but at the end of the day, like from a marketing perspective, just being practical right now, like people's attention spans are so low and honestly, like people are just, we're all like very distracted. So you don't have a lot of window to like nail it. So you do gotta have that one lighter, like pretty dialed in.

AI assessment note: “a good idea should slap in one sentence and then slap even harder in three”

Q I think David Ogilvie said, like, simple ideas are powerful ideas. That's certainly something I've seen in mischiefs work throughout your entire career, but is there some sort of contrarian or nuance to that concept?

A I think there could be the, the line that we use here is a good idea should slap in one sentence and then slap even harder in three. And that just means there, there are layers, right? Like you appreciate it for the headline, but if you like really dig into it, you're like, wait up, is this thing making fun of me? And do I still like it? And if I buy it, am I a chump or am I a patron of the arts? Right? That's like, uh, The layers are good. The ambiguity is good, but at the end of the day, like from a marketing perspective, just being practical right now, like people's attention spans are so low and honestly, like people are just, we're all like very distracted. So you don't have a lot of window to like nail it. So you do gotta have that one lighter, like pretty dialed in.

AI assessment note: “a good idea should slap in one sentence and then slap even harder in three.”

Q I think David Ogilvie said, like, simple ideas are powerful ideas. That's certainly something I've seen in Mischief's work throughout your entire career, but is there some sort of contrarian or nuance to that concept?

A I think there could be. The, the line that we use here is, a good idea should slap in one sentence, and then slap even harder in three. And that just means there, there are layers, right? Like, you appreciate it for the headline, but if you, like, really dig into it, you're like, wait up, is this thing making fun of me? And do I still like it? And am I gonna buy it? And if I buy it, am I a chump, or am I a patron of the arts, right? It's like, uh, The layers are good. The ambiguity is good, but at the end of the day, like from a marketing perspective, Just being practical right now, like people's attention spans are so low and honestly, like people are just, we're all like very distracted. So you don't have a lot of window to like nail it. So you do gotta have that one lighter, like pretty dialed in.

AI assessment note: “a good idea should slap in one sentence, and then slap even harder in three.”

Q to, to another, uh, topic that I'm, I'm sure you guys Uh, so one, I don't think AI is gonna take your guys' jobs anytime soon, uh, despite it being able to maybe piece together, you know, random ideas, uh, so much about, uh, taste and things like that, but, uh, how excited are you to leverage that technology for, for some of, um, for some of these various plays?

A Honestly, TBD, and I think, I think it's, I do, I will say, like, unlike the whole, like, crypto craziness of the last couple years, um, we didn't get into that because we didn't feel like it was here to stay yet. That, that's like the truth. Like, when you see the herd run so fast, it usually means they're running towards the end of a cliff. There's a little bit of a similar sentiment right now. That said, I, it is rooted in something real. Um, but we're not in a rush to, like, Use it necessarily, right? If it's going to be here forever, we've got time. Like it's mischief doesn't win as a first mover. If you look at all of our work, it was never about being the first to react to something or the first to like make use of a format. Um, so we'll see. We just have to see.

AI assessment note: “Honestly, TBD... we're not in a rush to, like, Use it necessarily”

Q that had negative gross margin and are all now pouring cash into both the like AI application layer companies and like, you know, foundation models that all have like ridiculously, I mean, I forget, I think it's girly is nonstop, you know, not my favorite person in the world, but girly is nonstop talking about like, you know, what is going on here? They're selling a buck for 90 cents.

A So I think, I think it's, it's an important example because you had that, um, uh, You know, plenty of examples of these chained losses during that, uh, that era where a restaurant was selling something below cost to a platform that was selling something below cost to a logistics provider, an individual contractor that like maybe wasn't actually making money if you factored in depreciation and fuel costs of their vehicle. And that ultimately worked out, right? DoorDash is a, is a massive, fantastic business, uh, based on the power of the American, uh, consumer. But when you compare that to today, where a lot of the conversation on the timeline this week has been the margin profile of this new generation of software companies that has to pay a lot for sales and marketing, but also inference. And so I think like the debate should really be, you know, continue to be around just how quickly will, uh, the cost per token fall. And I think a lot of people are, have a lot of confidence around that, but I think that That is the key thing that Everett's sort of, like, broad investment thesis right now is dependent on.

AI assessment note: “the debate should really be, you know, continue to be around just how quickly will, uh, the cost per token fall.”

Q this model is gonna behave weird in this weird way. Uh, like the Manchurian candidate problem. We haven't found any of that, but it certainly seems like something we'd want to keep an eye on. But, and from your perspective, like, what, what, what are the, what are the risks that we need to be aware of going into a world where China is really pushing hard into open source?

A Yeah, there's two, there's two, and you identified them, but let's, let's, let's talk about both of them. Um, so the, so the phone home thing is the, is the easy one, which is you can put up, you know, you can packet sniff, uh, you know, a network and you can tell when the thing is doing that and you, and plus you can go and you can go in the code and you can see what it's doing that. And so you can validate, you can even validate that that's either happening or not happening. And I think that, you know, that's important. Um, uh, but I, you know, I think people are gonna, people are gonna, are gonna figure that out. You, you, you can kind of gate that problem practically. Yeah. Um, the, the, the bigger issue is, um, we, we have this term in the field, uh, right now called open weights. Um, and, um, open weights is a loaded term. Uh, it, it uses the open term from open source, but of course with open source, the thing is you, you can actually read the code. Um, you know, with open weights, you have, you know, just a giant file full of numbers, as you said, that you, you can't really interpret. And then what you don't, what you don't have, what, what most, what most of the open source, uh, open weights models don't have, including, you know, deep seek specifically, What they don't have is they don't have open data, right? Um, or open corpus, right? So you, you, you can't actually…

AI assessment note: “Yeah, there's two, there's two, and you identified them, but let's, let's, let's talk about both of them.”

Q Did you see the roadshow? Did you go to any of the roadshow? I heard there was like a long product demo there, which is kind of uncommon.

A Um, yeah. So I think it's, I think this is true for, you know, uh, many CEOs, but especially for Dylan, like Dylan, Cares a lot about the product and it's important that everybody who is considering investing in the company really knows what figment, not just what figment stands for and design and creativity and all the things you see, but really how the product works. Um, and it was not a five minute flyby, you know, look at the cool, you know, shiny features. The product demo was, you know, they were in the weeds. Um, it was, yeah, I was, uh, I, in fact, uncovered parts of the product I didn't know existed. Um, but it was good. I think it's important, again, like part of, you know, that's who Dylan is, right? He actually cares so deeply about every single pixel and every single feature, and then that shows through.

AI assessment note: “it was not a five minute flyby... The product demo was, you know, they were in the weeds.”

Q You think it would a hundred percent have killed the deal?

A Almost certainly. I think one of the key deal points In the negotiation with Google was we do it via Google's comms and legal strategy because Google does not want to get hit with some crazy FTC lawsuit. That's going to not just be some multi-billion dollar settlement, but also a huge headache for years and bad press and stuff like that. So Google says, Hey, we, we want to do this our way. It's going to be awkward, but we're going to take care of everyone. And, and yes, we will. Yes, we will leave. Enough money on the balance sheet to pay out all the unvested options. And, but in exchange, you got to do the comms our way, which was rough. You have to create a terrible, you have to be scapegoated by the entire, you have to get hazed.

AI assessment note: “Almost certainly. I think one of the key deal points In the negotiation with Google”

Q You're thinking, this day I turn 18, that's gonna be me. Wait, is it, is it 18 or 21?

A It's 21 now. That also changed recently. Um, so Enjoy comes out with one of the first big e-cigarettes. It's, it's an example of that, uh, you know, older technology that has the big vapor cloud. Um, and the FDA hits them with a lawsuit and says, hey, you are selling an unapproved medical device. This is a device, it's electronics, and it's a medical product because you're making a medical claim, and that medical claim is this product helps you quit smoking. Smoking addiction, cigarette addiction is a, Is a disease. And so, by selling a product that helps you quit smoking, you need to be regulated by the FDA. ENJOY fights this back and forth. It goes all the way to the Supreme Court. ENJOY wins. Because they were kind of not saying that, at least they made the argument that they were not saying this is to help you quit smoking. They were just saying, this is a cool thing to do, separately. Don't worry about it. Don't ask us about the relation to smoking. This is classic. Um, but in the interim, The FDA is able to put an import restriction on the company, so they're making the product, I believe, in China, probably overseas somewhere, because it's electronics, it's equipment. They bring it in at the ports, and the FDA says, hey, while we sort out this lawsuit, you can't bring any more in. And so this is kind of like the ban hammer that they bring down. It winds up bankrupting th…

AI assessment note: “It's 21 now. That also changed recently.”

Q internet structure or the incentive structure. Uh, but, uh, it's a reaction, but it's also an action that should reshape the, the, uh, the, the relationship between different parties on the internet. Uh, are there different moments in the history of Cloudflare or the history of the internet that you think, ah, we could draw on as historical analogies, um, to learn from to see how this might play out?

A You know, I think that, um, you know, that Google really, uh, 30 years ago defined what was the business model for the web, which was, uh, you, you create content, search drives traffic, and then you monetize that content through advertising or, or subscriptions or, or ego, um, and ego and monetization, not quite the right thing, but you drive value from it. In one of those, those three ways. Um, and that, that's held up for the last 30 years. I, I, I think that this is, um, the first time that I've, I've seen something where I thought, well, there, there are really going to be, um, very, very, very fundamental changes to the structure of how the Internet works if we move from a search-driven model to an AI-driven model. And, And I think that, um, what I hope is that we can actually make this something where, um, what gets rewarded is not what gets rewarded today, which, which largely is, you know, who can write the headline that produces the most, um, you know, uh, sort of inflammatory response, uh, in, in all of us, you know, who, who basically drives the most cortisol, uh, you know, gets the most clicks and, and therefore the most ad revenue or the most subscriptions. That's, that's not a, that's not a healthy way of looking at the world. I think, you know, the alternative is if you imagine that kind of all of the AI models together are a good kind of assemblage of an approx…

AI assessment note: “Google really, uh, 30 years ago defined what was the business model for the web”

Q about like, I want you in the government too, and I want people like you in the government. And so is the government, uh, a place for high performing technologists to flourish? Is there something that needs to change there? I'm interested in digging into that, but why don't you kick it off with kind of like how you tell the story of how you got here to this moment?

A Yeah, I, I think, well, obviously, you know, from Northeast Wisconsin, Green Bay, and I kind of grew up always interested in the world outside of Wisconsin in the United States. When I went to college, I studied Arabic in the Middle East. I became fascinated by what was happening. I was an undergrad when we invaded Iraq. I was a senior in high school during nine 11. And the more I dug into that intellectually, the more I became convinced that I would regret it if I didn't step up and serve, um, Felt very strongly about service to country, but also on a practical level, really wanted to apply the language and regional and cultural skills I was learning in a non-academic context and felt like there was no greater challenge than joining the United States Marine Corps. So even though I didn't come from a military family, didn't know anything about the military, I just felt like this was the hardest crucible that I could throw myself into. I had a, just really an awesome experience, couple of great deployments, learned a ton. Was a counter television, human television officer. And kind of similarly, when, when I got out of the Marine Corps and, you know, I worked on a presidential campaign, I worked in the Senate as a Middle East guy, never thought about running myself. I'd moved back home to work for this energy and supply chain management company when some people reached out to me…

AI assessment note: “I kind of grew up always interested in the world outside of Wisconsin”

Q them the ideas, because I feel like the story that most people tell is, is, ah, this guy, you know, wants power or wants to, you know, he's on a war path. Everyone tells him no. And then he runs and gets elected or kind of sheer force of will. Um, but, but which direction does it go more commonly? Was your path uncommon or is it more the norm?

A Well, to be clear, I, I don't want to pretend like this was some pure aw shucks Mr. Smith goes to Washington thing, right? Like, clearly I was driven, ambitious, like, had kind of always had a drive my whole life, um, and wanted to, like, test myself, but honestly, it was not, like, I didn't think of, like, the political domain at that point in my life. I think more often than not, people are drawn to politics. It used to be the case that, like, you'd go to law school, You become a lawyer, you'd run for state or local office, and then that would be a natural kind of like farm team that would lead into Congress, and then you run for Senate, and then you try and be president one day. I actually think that's changing a little bit. Like one actually healthy thing I saw during my time in Congress is we started to get more young veterans, veterans of the kind of nine 11 wars running for office, which created a cool bipartisan group of legislators that were willing to work together on stuff. I think, you know, I think we're seeing voters like wanting to have more outsiders, whether it's, you know, business people, traditional or tech bros, like, you know, uh, that you guys talk to a lot. Um, I think people are just tired of business as usual.

AI assessment note: “I think more often than not, people are drawn to politics.”

Q We have a public American stable coin, you know, issuer in circle. Now it feels like yes, regulatory. Clarity is important there, but having broader clarity around how tokens are treated by, you know, how the government actually views tokens feels like it could catalyze, you know, much more of an explosion in investment activity and new company formation and new use cases for tokens. Is that the right framework?

A I think that is the right framework because, like I said, stablecoin's a big important piece of the pie, um, and, and very low hanging fruit, by the way, to my mind. Um, but Crypto as an asset class is much broader. So when you say where do the industry players, where are the crypto investors, where do we want to see it? I don't think it's a monolith. I think crypto as it grows to a multi-trillion dollar asset class, like any multi-trillion dollar asset class, you have different factions. And I think some fairly and some really smart people think just take this, take this win and move on and don't worry about the rest. And I, I think that's a little, I see why they might think that if they think that It's this or nothing, but I don't, I think that's a false choice. I think that we can have both, and this is a really opportune moment to have both. Um, and I think the, the question that has beleaguered the industry really has been the question over securities, commodities, um, which agencies are gonna have jurisdiction. I think that's a bigger question, and I think it would be a real shame if we let this moment go to waste. You know, the irony, too, so you have some stablecoins, or only stablecoin companies, who are like, yep, genius act, and move on. They don't want to get dragged into this broader, um, kind of a more omnibus package, right, with the market structure legislation…

AI assessment note: “I think that is the right framework because, like I said, stablecoin's a big important”

Q What's your, how would you, how do you think about big companies getting excited about stable coins and thinking or, or institutions and being like, there's a lot of potential here. We should create our own stable coin versus we should just figure out how to leverage this technology. Where, where do you see the kind of line and opportunities?

A Well, look, we already have two very dominant stable coins, right? Already today. Um, Tether, USDT. And then we have USDC. Um, so clearly that's not winner take all. I mean, those two are both growing and they both have market share. So we think there will be stable coins like those that will have network effects. But we also think at the same time, there are some businesses that are just so big and just so important. And if they launched their own stable coin, um, we could also see that too. Um, I don't think we see a world where everyone, I think we get asked this question all the time. Is there going to be a world where every company has their own stable coins? Yeah. Um, we don't think so. We could be wrong, but that's not the view of how we see this evolving.

AI assessment note: “I don't think we see a world where everyone... has their own stable coins”

Q think we're going to see a lot more of this. Uh, and we, we've seen some of this to date, but do you expect that to be kind of like a new, a new, Like, how, how bullish are you on, on sort of, like, entirely AI creators getting, um, getting real adoption, following, turn, followings turning into real businesses? I'm sure you've, you've followed a bunch of them already.

A Yeah. I'm, I'm personally super excited about it because it kind of separates, um, the content from the character. Like now, you know, before AI, if you were on Instagram, like you were both the character and the person coming out with the content. And so you had to like look in a way and present yourself in a way and talk in a way that was like interesting to the Instagram algorithm. And now it's like anyone with a good idea can create a compelling character. Um, and so I think Some of those are human characters. I've already seen way too many examples in my Reels feed of OnlyFans models who promote themselves with AI avatars of themselves now, which works shockingly well. There's some photorealistic human influencers, but honestly, some of the more interesting ones are things that could never be influencers before AI. So there's one called, like, Raccoon Stole My iPhone, and it's a, it's an AI raccoon influencer. There's, like, AI capybara influencers. There's, like, mystical creatures, like, all of these things that just come out of people's imagination.

AI assessment note: “I'm, I'm personally super excited about it because it kind of separates”

Q But maybe it just happens super, super fast, 24 seven. Or, or is there like a new machine language that emerges? Yeah.

A Um, I mean, I think one thing that's really helped us so far in AI development is, uh, to come in with some priors for, um, you know, how humans do things. And that's actually, um, you know, if you bake those priors in, they typically are great starting points. So I could imagine, like, maybe you start with something that's Slack-like and give it enough flexibility that it can kind of develop beyond that and really figure out the way that's most effective For it to communicate. Um, one important thing though is, uh, you know, we want interpretability too, right? I think it's, it's very helpful for us today that what the agents do is, you know, uh, easy for us to read and interpret. And I don't think you want that to go away as well. So I think there's a lot of benefits just even from a pure, like debug the whole system perspective, or just let the models, you know, speak in a way that it's familiar with us. And, you know, you, you can also imagine like we might want to plug in To the system too, right? So, you know, um, whatever interfaces we're familiar with, we would ideally like our model to be familiar with as well.

AI assessment note: “maybe you start with something that's Slack-like and give it enough flexibility”

Q Got it. Um, and then, and then does that lead to, um, More model fragmentation models that are good at programming versus writing versus poetry versus image generation, or, or, or does this all feedback into one model? Does the idea of the consumer needing to pick a model disappear? Are we in a temporary period for that paradigm?

A I think the main reason that we've seen that so far is, uh, because people are trying to make the best of the capital. Like we are all still GPU poor in many ways. And people are focusing those GPUs on the sort of like spectrum of wars that I think is most important. Um, and I'm, I'm a bit of a big model guy. Um, I, I really do think that similar to how we saw with large pre-trained models before with small fine-tuned models made it Like, had gains over the sort of GPT-II era, but then were obsoleted by GPT-IV being generally good at everything. I think, to be honest, you're going to see this generalization and learning across all kinds of things that means you benefit from having large single models rather than specialization or area fine-tuned models.

AI assessment note: “you benefit from having large single models rather than specialization or area fine-tuned models”

Q dropping a buzzword on it is like program synthesis. Is there a world where, uh, I know that, I know the tokens, like the images, we see them as, as renderings of squares and different colors, but, uh, the, when they're fed into the LLM, they're typically, uh, just a stream of, of numbers effectively. Is there a world where actually adding a screenshot is what's important? Like visual reasoning.

A Yeah. Yeah. So I think, I think that could be important. It's just like kind of, uh, you know, whenever it comes to like textual representation of grids, um, models today just don't really do that well, right? And I think it's just kind of because humans don't really ever write down textual representations of grids, right? Yeah. You know, we have a chess board, like no one really kind of just like types it out in a grid. Um, and, um, and so the models are kind of like, Undertrained a little bit on, on what that looks like and what that means. So, um, you know, I, I think with more reasoning, we'll, we'll, we'll just bridge the gap. Um, I think with better visual perception, we'll just bridge that gap.

AI assessment note: “So I think, I think that could be important.”

Q Yeah. And I guess like when I think of agents, it's, it's deep research is like, you know, curating Information on which you can take action on, but it's like, at what point is action a part of that sort of loop, right? Where you can not only curate a list of flights that you want, but then, you know, actually go out and, and, and have agency.

A I think one of our explorations in that space is operator, right? It's where you kind of just feed in raw pixels from your, your, your laptop, um, into, or, you know, from some virtual machine into the model and it, it produces, you know, either a click or some keyboard actions, right? And Um, so there it's taking action. Um, and I think the trouble is, you know, it, you don't ever want to mess up when you're taking action, right? I think the cost of that is super high. Um, uh, you, you only have to get it wrong once to lose trust in, in a user. Um, and so we want to make sure that that feels super robust, uh, before we get to the point where we're like, Hey, look, here's a tool. Um,

AI assessment note: “we want to make sure that that feels super robust, uh, before we get”

Q Yeah. But what's the, what's the actual format that you use? Like describe the video. What is the hook? Like break it down for me. Like you're explaining the, like the art behind the viral format.

A There's a caption. It starts with the face. Usually a handsome dude or a pretty girl is saying, damn, this interview is starting with the interview is starting with the hard questions. I should have been a CS major, not a business major. That's because people are saying like, bro, like CS is way harder than business. Then it turns around, the interviewer asks like, like, Hey, how are you doing? Why should we hire you? And then this guy uses clue to generate a response, but he can't fucking read the response. So he reads it hella autistically, like, oh, I revel in detail. And, and then that's, that is like another conversation point. Like people are cooking on the guy cause he, uh, he can't read properly. The guy is like doing a really dumb interview using Cree.

AI assessment note: “There's a caption. It starts with the face. Usually a handsome dude or a pretty”

Q I want to talk about Base. Obviously, it's powering a lot of, uh, consumer crypto use cases right now. Like, what, how did that come about? What has it been like incubating something like that inside Coinbase?

A You know, I can't take too much credit for it, by the way. The big, the biggest thing that I did was I, we have, we have a very entrepreneurial culture of different people we've brought in, and, um, Jesse Pollack came to me at one point, he was like, hey, I want to work on a new chain, and I was like, cool, like, take a small team, go run at it, you know, like, two pizza team or whatever, like, five people, and I had no idea if it was going to work, but the, the biggest thing, I didn't, he deserves all the credit, to be honest, the only thing I did was I kind of shielded it from, The rest of the organization a little bit and just gave them time to cook. And he went through like five or six iterations and I, I saw a couple of these along the way and I was like, honestly, I have no idea if this is going to work. Um, and it turned out to actually really hit. And so that's, if there is like a secret to this, which I don't, I don't know if it's a secret, but it's like you have to allocate some percentage of your, of your budget to these venture bets and taller and make them small bets. Cause you know, you want to have two, two, three, four people like working on it, kind of like a YC company, not like Don't put a hundred people on it or something. It's gonna be, like, a big bureaucratic thing. Um, as long as they're small bets, you can tolerate, like, you know, 75% of them not worki…

AI assessment note: “Jesse Pollak came to me... the only thing I did was I kind of shielded it”