I see a large IPO on the horizon. You're surrounded by journals. Watch your position.
Overnight success. The place 3000. Double. Right. Misinformation. Quarter clearing order inbound.
Let's just roll. We are surrounded by journal. Hold your position. Come. Get up. Trust the experts here. Five. Focus. We are founder code.
I see multiple journalists on the horizon. Stand by. UAV online. Blaze. Double blaze. Triple blaze. Double kill. Fight. Wrong. Cup is win.
Team deathmatch. We are experts. Triple blade. Just right. Clearing order inbound.
We are surrounded by journalists. Hold your position. Strike one. Strike two. Activate golden retriever mode. Marky clearing order inbound. Five good. I see multiple journalists on the horizon.
You're watching TVPN. Today is Wednesday, March 25th, 2026. We are live from the TVP Ultra, the temple of technology, the fortress of finance, the capital of capital. Sorry, Ramp.com. Time is money. Save both. Easy. Use corporate cards, bill pay, accounting, and a whole lot more all in one place. We have a great show for you today, folks. We got Mike from Arc Prize coming on.
The Nupinator. He did it. He did it.
We'll talk about this later. We have Nathan, my buddy from Air Street Capital, Rohin Dhar, giving us an update on what's happening in San Francisco in the housing market. Eric Jorgenson, the author of the book of Elon, is coming on. And then Jenny Just from Peak6 Investments coming on later. Thank you so much for tuning in today. Linear, of course, is the system for modern software development. 70% of enterprise workspace on Linear are using agents. So there's been a debate, a truth bomb dropped by Emil Michael. He says forgiving Benchmark and others would be like letting the Wuhan Institute of Virology slide back into a good reputation because the new senior manager of pandemic causation has made more friends than his predecessor. And so this got me thinking, and I'm probably going to have to put on the steel helmet for this one because this is wading into dangerous territory defending Benchmark. But my question is, how close are we to actually being able to forgive Benchmark? When is the right time? It's been a decade. Obviously, the drama between Benchmark and Travis Kalanick was awful. I think everyone's against what happened. But the question is, what is a venture firm? It is its partnership. If the partnership turns over at some point, is it a new team? Do you get a second shot? Can you actually change the reputation? And so, believe me, I get the Benchmark criticism. Travis is truly a generational entrepreneur and was on such an amazing run, right? So he was...
Jason from SaaStr was reacting to our interview with Travis, and his take, which I totally agree with, is it's hard to... if Travis had stayed in the role, it's hard to imagine Uber being worth less than something like a trillion dollars today.
Yes. So our friend of the show, John McCabe over at In.ai and Intercom, said Waymo is superior to Uber in literally every way. This was a year ago, in March 25th, actually to the day a year ago, he said this: Waymo is superior to Uber in literally every way that matters to consumers: smoother, safer, more reliable, no chatty, weird, rude drivers, private, quiet self-driving car services are going to dominate their human driver incumbents. And Owen says, TBT, when I think that's throwback to when Benchmark pushed Travis out of Uber and canned the self-driving division that he started literally 10 years ago. And so this is what's so tricky about this: Uber survived. It's 150 billion market cap. It's bigger than when Travis was ousted, but getting a 2x over a decade is not what I think people were expecting from Uber under Travis's leadership. And has fallen to just 5 billion. Like he won the capital war, and Dara has done a great job managing the business, but I feel like a lot of the success of Uber has been built on the foundation that Travis set up. It wasn't a complete reinvention. If anything, they just honed down the core business.
And the thing that is holding the business back right now, at least from a valuation standpoint, is this big question around self-driving. And Dara has answered this question thousands of times. Right now the strategy is to invest in self-driving companies, partner with self-driving companies, but not the same as having developed their own internal IP and product starting a decade ago and seeing where that would have been by now is just rough. Hard to think about.
Yeah. And so Uber is valued at 150 today, something like that. Waymo was valued in February of this year at 126 billion. And so yes, Waymo's been working on self-driving longer, but you have to imagine that there's another 50 billion of market cap. You have a serious... what would Waymo be valued if Travis was the CEO, right? You would get some type of Travis premium on it.
Just the market would say totally, totally, totally, totally. You have this sort of one-of-one entrepreneur in the seat.
100%. And just to sort of recap where things stand, I mean, Shervin Pishavar has been on the show as well. We've had like everyone from this saga in the TVPN orbit. Both Travis and Bill Gurley have been on the show. Shervin's been on the show. Emil Michael's been on the show. We've talked to a number of people that have been around this story, and it's a fascinating one. It's one of the most interesting. It was certainly formative in my career because I got to Silicon Valley and this was the first big story that played out, really. So Shervin said, 'In my opinion, Gurley single-handedly destroyed hundreds of billions in value. Travis and Emil staying in charge of Uber would have led to a Tesla-ized win, 500 billion plus for everyone, including Benchmark's LPs. He nuked decades of Benchmark's reputation with founders. The market has spoken, and no future Travis-quality founder would ever touch him or his former firm again, especially since three of the partners that approved of the ousting of Travis are still at the firm.' And so my question is, how many partners need to be at the firm until we can call this a Ship of Theseus? So for those who are not up to speed on their Greek mythology, in Greek mythology, Theseus is the mythical king of the city of Athens. He rescues the children of Athens from King Minos after slaying the Minotaur, which is his mythical beast. And then he escapes onto a ship going to Delos. Each year, the Athenians would commemorate this success by taking the ship on a pilgrimage to Delos to honor Apollo. Over time, because they're sailing the ship every year, various of its timbers rotted and were replaced. A question was raised by ancient philosophers: If no pieces of the original ship remained in the current ship, is it still the Ship of Theseus? If it was no longer the same, when had it ceased existing as the original ship? So some people might say 50/50. Some people might say, 'Yes, it is the same ship because replacing one board at a time, the ship is the concept.' And you can swap everything out 25 times, it's still the same ship. It's a paradox. There is no right answer. It's a philosophical question, but it applies, I think, to Benchmark because back when Kalanick resigned as Uber CEO on June 20th of 2017 after investor pressure that included Benchmark, on that exact date, Benchmark's equal GP roster was Bill Gurley, Eric Vishria, Matt Cohler, Mitch Lasky, Peter Fenton, Sarah Tavel. Today, the partnership has changed dramatically. The only two that remain are Peter and Eric. And you have Chetan, Ev Randall, and Jack Altman who are new to the partnership post the Uber scandal. And so it's not a full Ship of Theseus, but only one-third of the original 2017 partnership remains. And my question for those who remain reluctant to forgive Benchmark is, what happens if Peter and Eric retire or leave at some point and the full Ship of Theseus is complete? Like maybe you'll shift.
Yeah, right now, 40% of the partnership was there.
33%. Oh, I mean, I guess two out of the five. Yeah. So two out of the five because they've added three, but only one-third of the original partnership remains.
The question is, did Chetan, Ev, and Jack come in and as part of the interview process say, like, you're absolutely right? I mean, to defend Ev, two out of the three...
Ev had just graduated from college. Like he was truly not involved in the Uber scandal, and yet people will visit it upon him.
Knowing Ev and Jack, I'm sure, I'm sure during the interview process they were like, you know, the storied firm, I'm excited to join, but we can never do anything like that again. And so one question that's worth asking is, is the firm that did this and ultimately stained this storied brand and has certainly suffered the consequences, right? They've put up, you know, incredible returns since then, but I'm sure they've missed a lot of deals that would have made their returns even better because of that kind of narrative around the firm. And so they have... I would say, you know, Emil Michael and others are upset that they're doing great deals at all.
But I think it is... a question I have is, if they're in a situation again, right, in the same kind of situation, what decision are they more or less likely to make, right? I would argue they are probably less likely to go against the founder given how this entire situation has played out.
I would think so.
Yeah. The only steel man, and this needs the full steel helmet because it's so hard to steel man the Benchmark thing, but the full steel man of the Benchmark thing. It's really bad. It's really bad to Benchmark. I'm sorry for everyone. I'm sorry. But it's basically that every partner at Benchmark, it's an equal partnership. So every partner was going to make a clean $1 billion. They were all going to be billionaires from this one deal. And it was such a power law that there was no path to becoming a billionaire from the other investments, most likely. And so you see the endless 24/7 hit piece pile stack up, and you got Mike Isaac, you know, bloodhound on at the New York Times writing books. They're turning into movies. Like it's getting rough. It's getting rough.
Mike Isaac's at your door.
Yeah, Mike Isaac is at the door. The barbarians are at the gate, and you're like, I'm either a billionaire or I'm going back to a paltry 10 million, and I can't do that. I can't do that. And so they freak out, and they're like, we got to salvage this thing. We got to just push it out in the public markets. We got to get out of this name. And so they basically just... it's just too nerve-wracking. And yeah, it's not a strong steel man. But I think that's a little bit more of what happened than taking a stand on, oh, this particular thing that happened was so egregious. It was more just like, okay, wow, all my money, like 99% of my net worth is in this asset, and it's looking like it could be a zero because Lyft is coming from behind. There's a whole bunch of VCs piling into that. The narrative is totally flipping. There's a boycott Uber campaign, and everyone's like, ah, what's going on? I got to get out of this. I got to salvage this. And I think that was maybe more of the underpinning than I'm taking some sort of moral stand on a particular hit piece or something like that. Anyway, it's rough. But fortunately, you know, ship of Theseus, maybe it happens. I don't know. Would Delian accept that argument? Probably not. Would Emil? Probably not. But they're not making it any easier with the Manus investment either, because there was a world where it was like, okay, yeah, the Uber thing happened a decade ago, the partnership is basically entirely new, and they're focused on... but it's not there yet. Yes. It's just not there yet. It's not there yet. But still, I think you can rewrite this in 5 years.
Yeah. You made this point that it's too early to call this, but directionally, given that the firm is still putting up great returns, they've gotten to a bunch of great companies over the last five years, I think we are on a path to the Benchmark...
being the venture Ship of Theseus.
And VC Brag said Airbnb has no homes, Uber has no cars, and Benchmark has no partners. Of course, that was an exaggeration. But they did get down to just three partners, which is very, very small for a venture capital firm. Some have dozens of partners. But different strategy, and we will see where it goes. Anyway, let me tell you about AppLovin, profitable advertising made easy with Axon.ai. Get access to over 1 billion daily active users and grow your business today. And let me also tell you about Lambda. Lambda is the superintelligence cloud building AI supercomputers for training and inference that scale from one GPU to hundreds of thousands. So Sora, rest in peace, Sora. The app is leaving.
Yes. Ev was chugging beers while Uber was getting ousted. Do not visit the sins of the father on the son. That's what I would say. Ev not guilty.
Yeah, he was hanging out. He was having... What happened senior year? What happened? I think Ev deserves a fair shake and should not have to bear the cross from...
The real question: do Ev and Jack pull a Benchmark and force out the original partners? They don't have the legal control to do so, but neither did they during the Kalanick...
At this point, Ev is probably already getting calls from Sequoia to come be the senior steward, you know, because he's on a meteoric ride. We're really all over the place today. Anyway, Sora is now... is it still in the app store? I think the announcement was that it will be leaving the app store.
Millions of people have made content on the app. You have to leave it running for some amount of time.
There's a phase out, but the announcement happened and now it is going out. And I have a bunch of takes on this. Obviously, this is not the end of video creation for OpenAI. This will be rolled into ChatGPT, I imagine. Tyler Hodges put it well: bullish, killing products quickly is hard. Almost no one can do it. It's a good sign for OpenAI. They're consolidating in many ways. It's like last week you heard about the code red was like a month or two ago, and then it was like we're refocusing, and then it's like here's step one of refocusing, like a single app that we're going to push everything together. And also, I've enjoyed making some videos in Sora. I've never enjoyed having to go to a separate app. I want all of that to live in one place. So that makes a lot of sense. Let's see what Dax said: it's lame to see all the people saying, 'Ha, I called it. I knew Sora wouldn't work.' Yeah, duh. Because everyone thought that, including me, who were working on it. They probably learned a lot trying to make it work anyway. For every successful thing that exists, a hundred efforts like this had to fail, and those learnings are fed into making something that ultimately does work and provides you with a steady paycheck. Yes. It's interesting because this quote of people saying, 'Ha, I called it. I knew Sora would work,' that is not how I interpreted the vibes around the Sora launch. Like I went back and revisited the essay that I wrote on October 1st. We had the slop versus farming debate. I was really on a tear back then. Said slop is bad. We the timeline don't want to be pigs at the trough. We don't like it when tech leaders treat us like farm animals, but we love farming. Farming is Lindy. We the timeline want to return to a world where we are filling up troughs with slop on a daily basis, I guess. So between Google DeepMind, Meta Superintelligence, and OpenAI, we now have three different variations on AI video products, each met with slightly different responses. So the interesting thing here is that Google has been charging ahead, launching it, it's in Shorts, and that's just been not a story at all. What was interesting was that the vibes around both Meta Vibes and Sora were like, this is going to one-shot humanity. They're like, this is going to be too successful. That was it. It was like...
It was like an entertainment doom loop. Exactly. You could imagine exactly it just getting so good at generating the next thing that you would want to see better than even a billion humans on Instagram could do. And that's not what we've seen. So far.
And so my big question was, will this actually be sticky? Will people like this? And at what rate? I mean, I read LLM-generated text daily, but I also read a ton of not LLM-generated text. And my ratio has grown exponentially. But it hasn't gone to 100%. Nowhere near it. Like probably 5% of the text that I read is LLM.
I mean, we should actually revisit how we were processing it during launch when it was, you know, rocketing.
We should just throw on that three-hour stream and just watch that and react to how our jokes were that day. But even at the time, I remember saying it's very obvious that they built like a very cool creative tool. Yeah.
And they have the potential to seed a network with this. There's all this novel content. They had allowing creators and, you know, people like Sam to allow people to use their IP. Yeah. It was a very, very well-executed launch, but even from the beginning it was like, okay, obviously cool creative tool, it's a totally different ballgame. You know, this 'come for the tool, stay for the network' has been like an enduring strategy, right? Chris Dixon probably wrote that in like 2014 maybe, or like a long time ago, over 10 years ago. But just because you build a tool that is attached to a network, that jump is just really, really, really tough. Especially when there are three or four, five serious networks that are at scale that can on day one support the format of the file that is produced from the model. So in a world where generative AI video came out not in an MP4 file or an MOV file, it came out in some sort of format that could never be uploaded to Instagram Reels, then you have a chance to build a network and run away with it. And this was the story of Instagram. Like Instagram just had better support for images than Facebook did. And then Vine had support for video literally before Instagram. So Instagram, it was like, I have a video on my phone. It's cool. I want to share it. Yeah. Sharing it to Instagram was not possible. Now on day one you generate an AI video. You want to share it? You can share it on TikTok. Yeah. It's just the incentive. If you created an amazing video on Sora, what is the most logical thing to do? If you're a creator and you want reach, post it to Instagram. There's just naturally, there's billions of people there. There were millions of people on Sora and a lot of energy and momentum. Yeah. And it's not lost on me that the same day that Sora was killed, you have a viral breakout reality TV-style series putting up incredible numbers on TikTok for Fruit Love Island, I believe it's called. It's an AI-generated twist on Love Island. There's romantic intrigue and plot lines and stories and consistent characters and a lot of things that have come from a variety of AI models. And we should talk to the entrepreneur behind that project because I would be interested to know what the stack is, because I imagine it's not just going to a single Gen AI app. I imagine that they have a whole pipeline of a workflow in place to actually generate that. And so we're at this weird moment where Sora the app is going away, but we're also seeing more and more AI-generated content slowly see success. Whether that's the podcast that's at the top of the charts that's fully AI-generated, there's this Love Island show, there's a number of niches where they've found the right product-market fit for AI-generated content. But it's not overnight. We're living in Infinite Jest and we just can't look away. It's like for specific things it makes a lot of sense and so it's working there.
Interesting to think about Google's strategy with video. Even Google was like, we cannot operate this for free at scale.
No, that was the discounted rate.
Oh, I think I'm at 500 a month or something.
It was like an entry to start, it was like $250 a month and it was brutally rate limited like...
Yeah. Even with... we were laughing at this because three a day. I remember we'd be like at the gym in the morning, you would fire off a couple prompts and then you were like, I hope I got it right. Yeah. Wouldn't get it right. Then you were rate limited and you're like, wait, I'm rate limited on a $250 plan that's going to jump to 500. You're probably paying... got to check the rate, check what you're probably paying 500 a month still. No, seriously. And that's Google with all this cash flow, insane data advantage with YouTube. And let me tell you, rate limits kill retention like nothing. Nothing is worse. If you're in Instagram, the endless scroll exists. TikTok, you can scroll endlessly. You can use the... imagine if TikTok was like, after five minutes you have to close the app and come back in 20 minutes. Like how successful do we think that would be? It would be a disaster. And that was the experience for both Sora and Veo 3, where you would fire off a prompt and it would be like, okay, come back in a couple minutes, going to take me a while to cook. And then you fire off five and it's like, okay, no more for today. And then the next day happens and you forget about it and you go on something else. So clearly the compute constraints are immense and there's just so much more value that can come from enterprise and can come from deep research and so many of the other models that are immediately economically valued, like codegen, like enterprise workflows. And it's maybe more boring and less viral and less controversial, but it's where the compute needs to go. And so I think you're going to see the chips be moved around inside of all the labs to like compute will find the most optimal output. Like the tokens of the most value will always be the ones that the compute flows to. And as a lot of people predicted, like just endless random generations that aren't quite dialed yet, even the best video models, like they're just not there. They require a lot of work. It's not the same as where we are in terms of knowledge retrieval, where we are in terms of codegen. It's just way more valuable. So anyway, let me tell you about Okta. Okta helps you assign every AI agent a trusted identity. So you get the power of AI without the risk. Secure every agent. Secure any agent. Thank you guys. And let me also tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real-time experiences and new value with Cisco.
Market clearing order inbound. What did they say? September. People overestimate how much brain rot happens in a year and underestimate how much brain rot happens in a decade. So yes. Yes. Still. So, I mean, I'm using brain rot pejoratively there, but I do think that this move does not really bend the curve of just AI-generated content, but I still think it's like a slow rollout. Like it's fast in the sense that we went from no slop on the timeline to lots, and we went from actually zero. There was like one cool AI video, Harry Potter Balenciaga, was entertaining to general people, and now we get like five, and then next year we'll get like 20, and then eventually it'll be like hundreds, and it'll be like, oh yeah, I'm actually into that. Like people are into cartoons and people are into CGI movies and superhero movies, and some people will be into it. Some people will never like it. Some people will always say, I want a black and white film from the '40s. That's what I want to watch. And these rollouts, the diffusion of this stuff will happen. Should we revisit one of...
Sweet Davidson says, 'What is this? Y'all are worried about the wrong open claw.'
This is a good post. This is the open claw that F. Randall was worried about in 2017. He was not thinking about...
Was White Claw invented in 2017? When did White Claw get founded? That's a great... I feel like it really took off. It had a fast takeoff.
2016. Okay. Very good chance he was an early adopter.
He might have been an early adopter of OpenClaw. It truly was. Open... I'm calling it OpenClaw now. White Claw did have a fast takeoff for sure. It went from 0 to 60 and it was just everywhere all of a sudden. Anyway, let me tell you about Sentry. Sentry shows developers what's broken, helps them fix it fast. That's why 150,000 organizations use it to keep their apps working. I got a story for you, Jordi. Today, as I was driving into Hollywood from my hometown of Pasadena, I was driving through Hollywood and I look over and there is a new Hollywood sign. I'm not kidding about this. I actually saw this. I took a picture. I can maybe send it into the chat, but I can just show you because I'm driving and I just see this like...
Billy Bowman. Billy Bowman. I don't know.
I'll send it into the chat so they can pull it up. Let me see. Production. Here we go. But anyway, it is a remarkable story because Fiverr is running one of the coolest out-of-home campaigns, like just from out-of-home inventory. I didn't know you could do this. Let's... we can pull up either my picture or we can pull up...
Let's pull up this video from KTLA 5.
KTLA 5 had a video breaking down what's going on. Let's watch this and then we'll react to this. And while the team pulls that up, I'm going to tell everyone about Labelbox, RL environments, voice robotics, eval, and expert human data. Labelbox is the data factory behind the world's leading AI teams. So let's pull up the KTLA 5 report about AI coming for Hollywood, the mysterious sign along the 101, underneath a logo for Fiverr and a search box that says, 'Find the best AI directors.' It's a brash, bold statement. AI...
Bold statement for Hollywood. Fiverr has made a name for itself connecting projects with freelancers. Now they're launching an AI video hub, which they say can make content at a fraction compared to traditional production. This Billy Bowman guy is one of the directors that you can hire. He's based in Sweden. He's made AI videos for Google, Universal Music Group, and others. As you know, AI really hasn't taken over Hollywood yet, but it has certainly crept into commercials. Brands like Google and Jeep rolling out AI on national campaigns. Many are slowing rather to see the 30-foot sign which went up over the weekend. I first noticed it stuck in traffic yesterday morning after someone was so entranced they rear-ended somebody else. It's causing accidents.
So interesting. AI director. Yeah. Yeah. So it's basically someone who puts the prompt into the machine and chooses...
Is Fiverr going to pay the prompt in the box, buddy? Fender bender.
That's a great question. Yeah. Yeah. Can they be held liable for such a distracting sign? I thought, you know, flush with money. Whether or not it is a bubble, like you can debate.
That's interesting because Fiverr is not one of those companies. KTLA 5 is not prepared for it not to be a bubble.
So Fiverr's market cap now is $560 million, and that's down about 95% over the last 5 years.
Where are you seeing that? I'm seeing 350.
Sorry. Yeah. 359. What did I say?
Oh, sorry. So it's a $360 million company today. Down 95% from 5 years ago. It started selling off in 2021, sort of pre-ChatGPT. So I think the AI narrative might be a little bit overblown there. It did IPO around this price. It was a $700 million IPO, I think, maybe a billion dollars. Went through a massive boom during COVID and then sold off. But of course, the AI wave has not been kind to Fiverr because a lot of the tasks, like, you know, generating... AI is very, very good at $5 creative work, $25. Obviously, the prices go well beyond $5 since the early days, but in terms of the kind of projects that I always used Fiverr for, AI just one-shots all of that.
And the nature of Fiverr is like you have to define your task in a prompt. It's not, oh, have like a long conversation, get drinks with somebody.
That was often the bottleneck. It was like, okay, I need to do this task. I need like 10 minutes to properly define it, all these things. And it's honestly way more time than you spend prompting normally. Because with prompting, you're just like, I'll just try it a few times.
Bunch of times, kind of iterate, hit my rate limits, and then fire back up. Yeah. I mean, it was always a bottleneck when I remember as an entrepreneur, I found out about Fiverr and I was like, this is amazing. I can get random stuff done for five bucks, but the time commitment, actually finding the right person, making sure the reviews are good, it wound up being like hours of work. And if you have a consistent flow, you're better off just hiring a person. So they got kind of squeezed in the middle.
Yeah. The market is not excited about Fiverr right now. They're being valued basically at four times EBITDA.
And so, yeah, this is an interesting pivot for them. They're basically saying that you can come to us to hire someone who has all the tooling set up to actually sit there and sort of nanny all the AI models because it is a hassle. Like you described with me and Veo 3, I was sitting there like, okay, I fire off four prompts, then I go back. Like it's way better if you're on the API and you have Higgsfield wired up and you have Runway ML and you have access to the Chinese models, you know, the right tool for the job, and then you do fine-tune on someone's face. There's a whole bunch of things that you can do to get better results, but it takes time and it's a hassle and it's more of a professional job. It's not actually out of work.
Here's the main problem with the campaign. Yes. Is that Billy Bowman is a real person, okay, with his own website, with his own Instagram. You can just go and... like the primary issue with these labor marketplaces like Fiverr is disintermediation. If a business hires somebody on Fiverr and has an amazing experience, eventually they're just going to go direct because they build up a lot of trust. And it's very different than a platform like Uber where you don't necessarily want the same driver every time because they're not around you and all these things. And so the reason that the Fiverrs and the Upworks of the world, and there's been a bunch of other engineering-focused marketplaces, just have never reached insane scale like Uber is because of the disintermediation. And this campaign is effectively an ad for Billy Bowman who you could just go hire today.
Yeah, disintermediation has always been a problem on these platforms. Anyway, let's move on. But first, 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. 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. Speaking of stocks, did you see that Bombardier, the manufacturer of private business jets, is down 10% over the past month? A lot of people were wondering why this was happening. I think we now know why, and it's probably because the shot across the bow from United Airlines. So what competes with a private jet? Potentially United Airlines' new product, which is an entire row of economy seats. We got to pull this up. United Airlines says, 'The entire row is all yours. Welcome to the United Relax Row. Three adjacent United Economy seats with adjustable leg rests that can be raised or lowered to create a cozy, lie-flat space for stretching out. You'll also get a mattress pad, blanket, and two pillows. If you're traveling with kids, a plushie too.' United Relax Row will be available starting next year on more than 200 of the 787s and 777s, each with up to 12 of these brand new rows. So what do you think, Jordi? Is this the way... I was telling Tyler Cosgrove, who is out of the studio today, he's in Washington, DC, he's got to demo this. He's got to get on one of these. I don't know. Every time an airline announces something, it's always like 5 years until it actually is available. I've been waiting for Starlink for a long time. Took a long time for that to get rolled out from the PR release.
United has pretty good pace. Haven't they been quick to...
Do you think Tyler could get on this tomorrow or today? He's going to the airport today.
I think if Tyler's resourceful enough, if he ended up in a row with two empty seats next to him, he could just figure out a way to detach the armrest blocking this. Yeah. And just kind of build your own. Yeah. They might have to land the plane and arrest him. Yeah. But potentially worth the risk. He could also potentially negotiate with whoever's sitting next to him, say, 'Hey, you go and spend the entire flight in the lavatory, and in exchange I will vibe code you a sloppy app of which I don't understand what programming language is used.'
I'll trade you an app for your seat.
I'll trade you an app for your seat. And somebody might be like, 'That's amazing. I don't have anyone that can vibe code for me. This is too good.'
No, I'm sorry. Ryan Peterson says, 'Now we just need to put stairs on the food drink cart so you can climb over the top of them to get to the bathroom instead of holding.'
Yeah, this just feels very, very chaotic. But this is wild. Good. I don't know. Starlink, a relaxed row, a dream. I think that this could be a good option. United built a product that everyone who has ever been on a plane wanted, says John Collison.
John, you need to work on fixing Bombardier's stock price.
I don't think it's related. Oh, apparently Air New Zealand launched this in 2010. I wonder...
No, it is related because he should help them build out their pipeline.
Well, he's just pumping his bags because he owns a property in Ireland. How do you get there? You got to fly on United. So, you know, one hand washes the other on this. He's talking his own book. No, just kidding.
Let me tell you about Console.com. 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 scaling, monitoring, and security. And Carter says, 'If you say no to pretzels, the flight attendant should give you something called a coin of restraint, while worth nothing now, these coins will play a major role in the afterlife. The coin of restraint is very, very good.' I like that. Well, Elon Musk is teasing something cooler than a minivan that might come because Elon said the Cybertruck rear bench has three seats of ISOFIX attachments and is wide enough to fit three child seats or three adults. So there's been a big debate on the ISOFIX attachments. These are the little metal hooks that are installed in every car in the back seat for child seats. And so child seat requirements mean that you have to have a car that has these, and it's been called like the death of the five-kid family or something like that, or like the death of the big family, because at a certain point, in order to have another kid, you have to upgrade your car, and that has an expense associated with it. You need a three-row SUV or you need a minivan or you need something bigger that's not as affordable as just a normal sedan that used to be able to just throw four kids across, and maybe that was less safe.
Regulation killed the birth rate.
People actually say this. People say that child seats have saved like a million lives, but then they've stopped 10 million from being born. They talk about the...
relative exchange ratios. I haven't really dug into it too much, but there's clearly demand for more spacious vehicles with more seating. The Model S used to come with a third row, which I still don't understand how that was possible. The Model X came with a third row, and in China they actually sell a Model Y L, which is a long wheelbase version of the Model Y. I hope they bring that to the States.
People are saying, 'Make a minivan, Elon.' Elon says something way cooler than a minivan is coming.
And people are speculating. Garcia here.
I think it might be a data center test.
I think you might be like, 'There's another data center coming.'
Yeah, chip fab. He's like, 'You're gonna be able... I'm gonna... and Texas is gonna change the child labor law so that instead of worrying about bringing your kids around to different places, no, they're in the fab. They're in the clean room.'
They're in the clean room.
'They're making chips.' No, people are speculating. This is very cool. Gen AI has been amazing for car enthusiasts to create basically their own concept cars. Yeah, these look very, very cool.
I mean, it's funny because 'make a Tesla that look... make a Tesla version of the Rivian.'
Because it looks exactly like it. But let's get into some of the speculation. So, okay, what are people saying?
That could be fun. Arthur McWater says, 'I can't wait for the next Roadster unveil.' Elon was teasing, it was on Joe Rogan, right? This concept of maybe it'll fly. And I think what Elon could be getting at is, picture a Roadster. Not a great family car. Hard to put kids in a sports car. Some of them you technically can, but it's so uncomfortable and kind of chaotic. Very few people would.
And I think what we could see is the Roadster comes with five, kind of like, you know, that Fury collaborative combat aircraft from Avatar. What if it comes with like up to five little mini Roadsters that are just trained to autopilot behind the primary Roadster? So you can be in your sports car and then however many kids you have are in the mini drones following that fun.
And what about a Chinook heavy lift helicopter with two massive rotors that can lift your Roadster off the ground? Yes.
Technically a flying car then.
Yes. Anyway, let me tell you about Phantom Cash. Fund your wallet without exchanges or middlemen and spend with a Phantom card. I have a question for you, Jordy. Are you running the new AI model? It's on co-work. It's literally on co-pilot. You can probably find it on Codex, dude. It's on co-author. It's a co-sign exclusive. It's on cocaine. You can run it on cocaine. You can literally go to cocaine and run it. What a great copypasta. This is one of the funniest formats. But yes, the war for co-pilot and co-work is heating up. We got to find a new term. I think people have been really, really fighting neck.
Well, I was telling Microsoft they should just name it Coco.
Microsoft co-pilot, co-work. It's just Coco.
Coco would be good. There's a few different options. I do think I prefer the non-anthropomorphized AI names, although they are a little bit colliding in the namespace. I have been a fan of, you know, the Codexes and co-works and co-pilots. Those feel more collaborative to me and they feel more like tools than the Bards and the Siris and the Alexas and the Rufus and the Sparkies. Like, that's just a different vibe. And I think if we're living in a world where people are going to form, you know, strong relationships with these tools, introducing them truly as tools is probably...
Let's see what's going on with QVC.
Let's pull up this video.
And we also have one of our guests joining early at 11.
No, he'll be ready to join in an hour.
Okay. Let's check the production team because I've already texted. Okay, cool.
We'll continue for the next 15 minutes.
Fantastic. Well, take us through the next story. What do you want?
Q River says, 'QVC basically reinvented live streaming decades ago.' Ghost 247. They invented live TV. Ghost 247. A good 80% of the show is just the host going on long personal digressions. People watch it as background. Heavy parasocial element. Hosts know the callers. 95% of repeat buyers. It's 100% Twitch for grandmas. Let's pull up QVC.
And with the submarine, I wanted to tell you that I got that peasant blouse with the tassels.
And I used to wear the tassels on my pasties. Do you know what pasties are?
This is... Okay. Ridiculous.
Do you know what pasties are?
Let me... I'm going to...
Okay. Well, I have to move on. We're moving on. Let me tell you about ElevenLabs. Build intelligent real-time conversational agents. Reimagine human-technology interaction with ElevenLabs. And let me also tell you about Gemini 3.1 Pro. We are going to deep dive ARC-AGI today. And Gemini 3.1 Pro has done very, very well with a more capable baseline. It's great for super complex tasks like visualizing difficult concepts, synthesizing data into a single view, or bringing creative projects to life. SpaceX aims to file IPO as soon as this week. I know everyone is excited for the S-1. Particularly, I think people will be focused on xAI. Yeah. What they have going on. I think... are they going to have to break it out in the S-1?
I would assume the first time we actually see the economics of inference, the economics of a foundation lab. Even though, yes, they are at a smaller scale, it's hard to read too much into it because they've been investing so far ahead of.
Yeah. I just think that like there is a world where we get broken out financials that you can dig through and you can understand based on Grok pricing, which we see, and topline revenue and cost. We can actually see are they serving that model profitably, and there will be, you know, a lot to dig into there. Obviously the other labs have different strategies, different vertical integration points, different economics, different pricing regimes. I mean the true frontier, the models that are dominating ARC-AGI, which we will talk about in 15 minutes, those command a premium, a price premium, and there's a wild difference between charging 15 bucks per million tokens versus $2 per million tokens. So it will be...
Yeah. If you remember, it was I think in Q4 of last year, if you looked on OpenRouter, Grok was what had, I think it was like Grok Fast had a ton of usage. People are like, 'Okay, why is this happening?' And part of it, at least I believe, was because they were subsidizing it.
But okay, so here's what I think. So people were posting this as though it was fact, but I think it's a very real possibility. So I think Elon will try to aim for the company to actually go out on April 20th, 420.
And I think it is possible if he goes fast. The ticker, we'll see what the ticker ends up being, but I think some people would like, knowing Elon's very millennial sense of humor, I think the ticker is...
Oh, you think so? Plus the April 20th IPO. I would assume that... you think this is a real prediction? You're not trolling?
I'm not saying I would bet on it, but I think there's like a... I think there's... I think there's... I would put the April 20th at maybe like, you know, 30% and then the ticker maybe down at 15%.
Calli asks, 'When will SpaceX officially announce an IPO?' Before June 1st, which April 20th would be before June 1st.
I'm talking about like actual listing day, like the day of the IPO.
Yeah, this is just announcing the IPO, which I don't even know if they... they haven't even confirmed this. This is currently in the scoop thing. So they haven't even...
Wait, John, did you say scoop?
Yes, this is a scoop from...
Can we go to the wide angle? We got Katie Roof, scoop master.
We have a new award for her, the first TVN Golden Scoop.
Do you want to show her the...
I don't want to pick it up yet.
Okay, we can show... we need to give the Golden Scoop Award for the best scoop of the day to Katie Roof, who moved markets with her scoop. She, of course, is the deputy bureau chief of venture capital at The Information, and she's an absolute scoop athlete. She's Scoop Doggy Dog, and she moved markets. So, SATS is up 7.8%, BKSY is up 4.8%, Lunr is up 4.1%. Every stock in the space...
And we're working on a new award show, the Scoopies.
I think Katie Roof is a lock.
I mean, she's a front runner for sure.
Putting on a generational run.
Front runner for sure. Yeah. It's so funny. The... I mean, incredible moment for the retail space investor community.
Because they're just... for some reason, people... people...
Anytime you get SpaceX repricing, they're like, 'Well, this other random company that happens to be technically on the same market map definitely deserves to be worth 7% more.'
Yeah. I don't know. I was as skeptical about the rest of the lunar economy for a long time. It felt like winner-take-all. It felt like SpaceX was running away with it. But there have been interesting dynamics where companies that are buyers of SpaceX capacity, whether on the satellite internet side or on the launch capacity side, they don't want a monopoly to exist. And so they're willing to throw money at competitors. And Jeff Bezos has stuck around with Blue Origin. Rocket Lab's done very well. There's been a bunch. People are saying we need to redesign the scoop. I don't know what you're talking about.
I don't even know what you could be talking about.
It's an ice cream scoop. It's an ice cream scoop. Get your mind out of the gutter. Head over to 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.
Okay. Casey Hammer is worried.
What I worry about is a generation of talented, experienced engineers being too rich to work. So, if you're worried about being demotivated by your incredible SpaceX liquidity, you have to do what I recommend, which is the Brewster Millions approach, where you have to spend all the money in 30 days without telling anyone why you're doing it. This is from the 1980s, 1990s comedy called Brewster's Millions, which I highly recommend you watch. In it, a man is gifted an inheritance from a wealthy uncle or father figure, grandfather, and it is conditioned on the fact that he needs to spend something like $30 million in 30 days without telling anyone why, and he can't accrue assets. So he can't go and just buy cars. He needs to throw parties and give money away and spend money wildly. And it's to teach him that money does not bring happiness, of course. And I think that's the solution to this. Anyway, 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.
Here's why I'm not worried.
Because I think that these ultra-talented, hardworking engineers...
Having some liquidity. Let's say someone's been there for some number of years...
They have 5 million liquid. They buy a nice house somewhere...
And then realize, 'Hey, I've got a nice nest egg. Yeah, I can keep working on crazy moonshots. I don't have to go join, you know, the historical example would be like work on the hard thing or work on enterprise SaaS.' And I think this just will give people more confidence to work on the hard thing, the thing that might have a 5% chance of success, but if it's successful, you know, has a tremendous impact on our country and things like that. So...
It might happen for some people, but in general, I do think that there is this liquidity wave coming. Delian talked about it on the show yesterday. At the same time, like SpaceX has been doing tenders for over a decade, and it's not like the early employees have never had a crumb of liquidity or secondary throughout their journey. A lot of them have had opportunities to sell at least a portion of their stake and had been able to buy houses. And there's always been a way to access some of that capital, whether through a loan from a bank, and you're obviously more creditworthy if you own a bunch of SpaceX stock and it's going up. And yeah, a bigger issue for the kind of space economy overall is just the Blue Origin people who, like, wasn't... didn't they have like way more? Like they didn't... they didn't really know what their equity was worth. They didn't... it's not like there was regular tenders. And so why work at the number two, at least by many definitions, space company? And then that's potentially ultimately bad for competition and it's bad for launch pricing for all these other companies that are reliant on the SpaceXs and ultimately the Blue Origins. So anyways...
Apparently, the co-founders of Manus are still in China and were called up to talk with the government and are now blocked.
Have you seen The Dark Knight?
Has everyone seen The Dark Knight? You thinking what I'm thinking? You thinking what I'm thinking? We go there, wrap our arms around him, the plane comes, grabs the balloon, sucks him out of the back of the skyscraper. It's one of the greatest scenes, and I think that's what we got to do because we need personal superintelligence.
This was surprising to me because I feel like we've been messaged to for a long time that they were in Singapore.
We shouldn't have been seeing this. It's not a Chinese company. They're in Singapore. The whole team's in Singapore.
And you would... it takes some audacity to sell your leading Chinese, one of the leading Chinese AI companies during a global AI race, this battle between great powers, and to sell to a big American hyperscaler.
Maybe get out of the country before you do that because it's not at all... I mean, this doesn't... when the Manus acquisition happened...
Yeah. It seemed very clear that you would be, if you were China and you were competing in the AI race. Yeah. Even though Manus is not like a lab, you're still like, 'Okay, they're building a powerful harness.'
We probably don't want them going to serve...
And the harnesses are incredibly probably under...
More important now. Yeah. Super important right now. Everyone, you know, obsesses over the models, and the models are important, but the harnesses have shown incredible promise for actual diffusion and making these models. So Josh Wolf says, 'I thought this wasn't a Chinese company.' And Delian goes for the jugular and says, 'So much for all the arguments about Manus not being influenced by the CCP.'
Yeah, I'm mostly shocked just about how this is playing out. You would think that if you're the CEO of Manus and you get a call from Mark Zuckerberg and it even smells like a potential acquisition, you're like, 'Yeah, I'd love to come see the headquarters. Why don't I come with my team? We'll just come and hang out in Menlo Park or Miami for a couple months while we hash out the deal. And if the deal doesn't go through, we'll head back. But if it does, we'll just stay and we won't go back because if it goes through, then there's going to be pressure and we're going to wind up in this situation.' This feels almost predictable. I don't know. It feels like there's something else going on here. I'm excited for this story to develop.
Authorities are reviewing the sale.
Yeah, and they're being asked not to leave. Seems hard to reverse at this point. But again, not sure why. I mean, when you look back at acquisitions that were blocked historically, who knows? Who knows how feasible it is to fully block the acquisition, but they can certainly block these individuals from, you know, materially benefiting from it in some way or contributing to Meta's efforts.
Well, here's some advice for the CEO of Manus. When you get here to America, when you get that liquidity, that payday from Mark Zuckerberg, open an account on Public.com. Investing for those that take it seriously. Stocks, options, bonds, crypto, treasuries, and more with great customer service. Just do it. And then, you know, tell your story. Start re-streaming one live stream, 30 plus destinations. You should be multi-reaming. So, go to Restream.com. Tell your story live on the internet. Gabriel says, 'Meek Mill has been going off.' Yes. About AI. AI is helping him organize his whole music career and other businesses in days, and it's moving his business forward at a high rate. Some tech young bull I met on LinkedIn gave me an incredible template which probably GStack...
Who else can help me with Claude? Gabe says, 'Drake, I don't do K2, Kimmy, DeepSeek. That's more for your kind. My gal more like Demis, theme park ting, London deep mine. Meek Mill, I cla-coded perks out of 18th and Burks. I got 500 agent lawyers trying to free Lil Durk.'
That's a good line. Bars.
Drake singing, 'You got me loco trying to be your shoulder.'
Two Chainz, 'Open Claude on my laptop, trapping off these Mac minis. Shout out to YC. Real Gs want to stack with me.'
That's good. That is good. I mean, yeah, it's not a bubble until GStack makes it into an actual hit.
The funny thing here is there's some debate, and there's this general vibe, like some people were latching on to this Trunk Fan post about, like, you know, are people making fun of Meek Mill? Are people like undercounting his ability to actually build something for real? And the interesting thing is that I would definitely bet on Meek Mill to build a real valuable piece of software for his audience or his life or his business over any of these tech people writing rap lyrics, even with all the powerful LLMs. Like the tools to actually build software are way better than the tools to write rap lyrics. And it's interesting because you are seeing this dynamic where I think a lot of people are latching on to the older paradigm of like, oh, you know, Jeremy Renner built an Instagram clone at one point and it was just the Jeremy Renner app and it was just his feed of his photos and people were like, why does this exist? You can just follow Jeremy Renner on Instagram. It doesn't make sense that he would have like his own private Instagram. And he probably spent a lot of money developing that piece of software. But I don't think it wound up working and it didn't scale and he wound up winding that project down. But if you think about like, well, what if this cost of doing that is a hundred bucks or 200 bucks or a thousand bucks? Like all of a sudden the hurdle rate to clear something like that actually does open up the creative aspects where I wouldn't be surprised if there's a like maybe not like a breakout hyperscale incredible like generational company, but just like in terms of like you can be a great musical artist and also produce great clothing, like you will now be able to produce software at the equivalent level. Like it is available. And so if you are constrained by your ideas and if you're a creative artist and you have a great idea, you're no longer going to be in this world where you're like, well, I need to put a couple million bucks down for a software engineering team and they're going to not really take me seriously. And all of a sudden you get into this weird thing where like the app that they launch is like sort of iffy and not that good. So I don't know. I'm actually coming away bullish on what Meek Mill winds up producing over the next few years.
And we are working on getting Meek on the show.
I really hope we can talk to him.
I'll cover you. You're good. You're good.
I'll cover Meta. Do you want to jump for a second?
First, let me tell you about CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. And I'm also going to tell you about Vanta, automate compliance and security. Vanta is the leading AI trust management platform. And I'm totally good. So I can take the next intro.
I have a personal thing I'm going to run to, but I will see you guys tomorrow. Yes. Love you. Cheers.
Thanks, Jordy. Up next, we have Mike from ARC Prize in the Restream rating room. Let's bring him in to the TPN Ultra Dome. I'm very excited to talk to Mike. How are you doing?
Here we go. Hey. Hey. Good to see you again.
I'm so excited. This is always the highlight of the show. I love talking to you about everything, but what are we talking about today?
Reintroduce ARC as an organization and then take us through the actual... do you even call them benchmarks, challenges? What's the right term?
I think benchmarks is a fair word.
Yeah, almost three years ago now, co-founded the ARC Prize Foundation with François. The ARC Prize Foundation has a mission to be the north star of AGI. So sort of our job, we have two of them. One is to be a useful public sense-finding tool for the public to understand how close, how far are we towards AGI or not. And the second is to inspire progress towards AGI. ARC is a series of benchmarks that help highlight what are some of the large remaining gaps between what frontier AI is capable of and what humans are capable of. And we sort of target that gap. That is sort of our definition of ultimately AGI. We produce an ongoing series of benchmarks, continually studying frontier progress, and at some point we are not going to be able to do that job anymore. We will run out of ideas, we'll test frontier and say we can't find anything else, any more gaps. And I think that'll sort of be the moment when I think it becomes commonly accepted to say, okay, yeah, we've got AGI now. And today we are announcing and launching the newest, best version of ARC, ARC-AGI-3. It's the latest in the series. This is a really large format change from the first two. ARC-AGI-3 is designed to test agentic intelligence, and it is, as far as I am aware, and I've been sort of interviewing folks all over the AI scene in the last few weeks, the only unsaturated general AI agent benchmark in the world. The headline score is human score 100% and AI less than 1%.
Okay, unpack that launch because my conception of ARC-AGI-3 is it's almost like a 2D game. It's no longer the puzzles where I'm picking colors to match a pattern. It's actually moving arrows on the keyboard. I'm stepping on triggers. I'm opening doors, switches, that type of thing. And I played it with you on the stream months ago.
Yeah, you helped us launch our preview several months back.
Okay, so that was the preview.
Today, we got the full dataset launching today.
So does that mean more... I'm going to call them games, more actual levels launching? Or is this that what you're launching is like you did the actual benchmark and got the four leading labs to devote the compute and actually open up their models to be able to interface with the system to get the scores?
Both things actually. So today, on the benchmark side, the public version of the benchmark is... or I guess the overall benchmark. So over 100 games, nearly a thousand different levels across these game-like environments. I think it's fair to call them games. We've designed them to be fun, and games are fun. I think you could look at them, you know, from a research standpoint, more as environments though. These environments are intended to test whether AI can effectively explore, discover its own goals, acquire strategy, develop plans, execute its plans. One of the really unique things about ARC-AGI-3 compared to the 1 and 2 format is that it is interactive now. Whereas you mentioned, you know, 1 and 2 look like these kind of static IQ puzzles that were on a page. ARC-AGI-3 challenges both humans and AI to essentially figure out the goals of themselves. When you're dropped into one of these environments, your only goal explicitly is to win. And so in order to figure out how to do that, you have to actually dedicate some extra regulation to figuring out the rules, the mechanics, the strategy. And one important thing is you're sort of playing these environments. The strategy mechanic, they grow and they evolve and they change over time. And this is one of the reasons I think ARC-AGI-3 will be a really useful tool for understanding agentic intelligence this year. I think it'll be our first real test or, you know, seeing early progress on these AI systems that are able to do on-the-fly world modeling, some degree of continual learning. These are both critical capabilities that we view as missing today that ARC-AGI-3 tests for.
Okay. Take me a little bit deeper on... you said there's a thousand games or a thousand levels.
I think it's a few over a hundred games.
Across those 100 environments, nearly a thousand levels across all of them. Yeah, it's a much larger version of the benchmark than we've ever had previously. And then like I mentioned before, the other major thing we're announcing today is frontier scores. So the benchmark is launching. We're also publishing as of today the latest four models across all the four major labs. And yeah, I think SOTA is currently sitting at like 3.4%.
Gemini 3.1 Pro. Maybe there's an extra hyphen on there, but basically the Gemini, Anthropic, and OpenAI were all in the 0.2, 0.3 something, and then Grok was, I think, 0%. Walk me through the actual build-out of these hundred games. Is this entirely human done? Is there some sort of computer-aided tooling to insert variation programmatically? Or is it important that they're all created by hand? How do you think about the creation of...
I wish we could use AI to help design games. We'd be able to make the benchmark even bigger, like better. The reality is, like, humans are still the bottleneck on creativity, and so every game has still been handcrafted and hand-designed by humans. You could sort of imagine if you were embedding all these different levels on a big manifold, you know, in an embedding, you want them all as far apart as possible in that sort of space. And today, still, humans are kind of the limiting factor in terms of ensuring that, you know, every game is different and as novel from each other as possible.
There's a few interesting design changes actually from a benchmark standpoint compared to 1 and 2. Maybe the largest is, you know, ARC studies the frontier progress. We have to design our future versions of the benchmark to adapt to changing frontier progress. One of the design goals with 1 and 2 was we have what's called a private and a public test split, where we have a public version of the benchmark and a private holdout version, which is what we actually use to verify the performance of frontier models.
So the frontier models get no freebies. They don't get anything from the public set, but they can try.
You can't memorize the public set, right?
Or they can experiment on the public set from like a prompting perspective maybe.
Yeah. The idea is the public set is intended to demonstrate the format.
And this was similar with ARC-AGI-1 and 2. However, we held a design goal that the public sets and the private sets were IID with each other. Basically, that they are supposed to be as close as possible to each other. And it's just split along visibility. Some are private and some are public.
With one of the big advancements with AI reasoning, this is actually like not a very useful way to run benchmarks. AI reasoning systems are so powerful now that they can actually generalize across IID test splits, and this is what we saw with ARC-AGI-1 and 2. So with 3, one of the big design decisions is we're actually releasing fewer games into the public demonstration set. So there's only, I think, about 25 games that are in the public set. We're actually explicitly not even calling it a training set anymore. We're calling it a demonstration set just to show the format to humans, you know, be able to test your systems to make sure you can sort of create them, get a feel for them. There's obviously fun marketing value in being able to play the games as humans too, which we really love. And on the private set, this is the set that's over 100 games. They're specifically different. They're different. We designed them with different characteristics, different goals, different intelligence capabilities required to beat them. The difficulty, the acceptance criteria is more extreme between human and AI performance, all to hopefully produce the most useful, like, high-signal benchmark towards whether we actually are getting real progress towards AGI with the foundation models.
So let me pitch you a strategy. If I have access and I'm at Google or OpenAI or Anthropic and I want to do well here, can I take the public set and create a log of all the steps and all the reasoning chains and all the keystrokes that are required to pass those levels, and then sort of like dump that into the context window before I go off into the unknown and train your model that way?
Maybe train my model, but also I'm just wondering if that's helpful for setting up the context or like doing some sort of like pre-compaction of the strategies that are learned. Maybe not even training a custom model because I feel like that would maybe be like bench hacking. I'm more thinking about just like, okay, we went and we played all the public games to completion and we monitored them, screen recorded them, tried to extract as many learnings as possible into, you know, an MD file basically, and then we include that in the prompt that kicks us off to sort of bootstrap the learning once we get into the unknown environment.
If we've done a good job on the benchmark, you should not be able to train a system on the public set and perform well on the private set. If we've done a good job, obviously every benchmark release, it's an experiment. Right. We make contact with reality. We ship these systems benchmarks publicly. We try to analyze the performance, understand what they're good at and bad at, and evolve, you know, future versions of the benchmark. But intentionally, you know, and this actually is very closely related to another design decision that we're making with our scoring function going forward this year. And this is again in response to like AI progress that we've seen. You know, our scoring methodology is basically AGI-field at this point. We, going forward with V3, are using, as I kind of have this idea of like basically a philosophy of having essentially no harness. We want to create a testing experience that's as similar as possible between the human and the AI test takers. And when we have our human baseline, when we have a, you know, rented literally a testing center in San Francisco, had, you know, hundreds of humans play these games. All they're given is you have sort of, you know, sensory input through your eyes and action motor output through your hands back into our testing interface. And all of the intelligence happens between those two steps. And so we try to emulate that as close as possible for our verification function, where we have this sort of philosophy of having a very stateless client. So that our scoring function basically tries not to introduce any kind of bias, any kind of help, any kind of maybe potential cheating strategy. If you go read our prompt, it's extremely simple. It's like, you know, you are playing a game, here's your actions, your conversation will be carried forward to the next turn, and that's it. In order to again kind of produce this really clear signal towards when there's real progress towards AI and the base intelligence layer, we're able to detect that.
Okay. So take me back through history a little bit because I'm surprised by why AI is struggling with this in particular. Because I remember it feels like almost a decade ago that OpenAI had a product, I think it was called Gym, where they were able to beat Mario and then they beat the Dota team, Dota 5, and they were able to do things that I can't do. I certainly can't beat Lee Sedol in Go. I certainly can't, you know, win Jeopardy or any of these things. And yet AI systems were able to dominate those games. You've created new games. What's different about the games or the strategies by the AI labs where we're not matching up like we did in the past?
I think the biggest thing is the expectation of what constitutes real progress towards AGI, right? When labs were using games in maybe the 2016 to 2018, 2019 era when they were very popular, you know, human researchers are studying the games, trying to understand the failure modes of machine learning, deep learning, trying to build custom search harnesses and sort of feedback mechanisms from the environments. It's very, very handcrafted. It's loaded with what I'll call like human GI in the research process. We are now at a point where we want to control for that actually. We want to understand, like, we want to control for as little human GI in these systems as possible, right? We want to understand is can basically AI do what the human researchers were doing back in that era in order to beat those games that they had never been trained on or exposed to before.
So I do think it's kind of elegant that, you know, we are coming full circle where games are these very minimal representations of like actually important capabilities that humans possess around exploring and developing strategy and world modeling and being able to learn on the fly. They're really elegant as far as an environment goes. But I think what's changed is our expectation of how much human crafting is needed in order to learn the games when they haven't been specifically trained on them is the big difference today, especially with ARC-AGI-3.
Okay. Remind me of some more history, but more related to ARC. I remember with one of the ARC-AGI benchmark tests, there was a version of a model from OpenAI that was running on some sort of like extra high mode, and I seem to remember like $2,000 per task being cited. Something o3 big launch.
That was like a preview of o3 in December of 2024. It's really the first... you know, there's a great chart on the ARC Prize homepage now where you can actually see this data point so clearly. I think one of the really, you know, like I mentioned before, one of our missions of the foundation is to try and be a useful public sense-finding tool.
And I think, you know, when we first launched ARC-AGI-1 and 2, you know, it was a very common critique. It's understandable, you know, hey, these things look like toys. Are they really economically useful? Are they going to lead to any, you know, real progress? And now in hindsight, I actually think that's a pretty outdated view because we have pretty strong evidence that ARC held quite strong predictive power of noticing really important AGI moments. We only started seeing saturation on the V1 benchmark, and remember V1 was like 5 years old. We only started seeing any amount of progress from LLMs on V1 once we got AI reasoning, which was a really critical innovation. I'd argue is as important as the original transformer innovation. And then a year later, this was, you know, 4 months ago now, with the November-December 2025 class of models with GPT-5.2 and Opus 4.5, we again started to see saturation on ARC-AGI-2, and it precisely correlated with this like agentic coding capability that emerged.
And so I'm optimistic that ARC-AGI-3 will again be a very useful sort of predictive tool to understand when, you know, basically AI agents are capable of operating in more open-ended environments. Right now, you know, you need a lot of human handcrafting to get these intelligence systems to work in domains such as coding, right, with Claude Code and Codex and Claude Code. And that's... I basically expect that like when you are doing very good on V3, which will mean, by the way, 100% score V3 means like AI can sort of beat all the games as efficiently as humans can on an action basis, that will lead to economically useful systems where agents are able to operate in more open environments that they haven't been specifically trained on.
I still remember from ARC-AGI. You see these like 3x3 grids, and the first time I ever tried it, I tried on my phone, and I think my phone was in some weird like landscape mode or something, so it wasn't rendering correctly, and I was like...
You didn't even get all the data points.
Yeah. No, so normally it's like you see the blocks and then you see the blocks to the left and the right, and then I was like, wow, I'm like, I'm cooked. Like, the fact that other people can do this. But of course, once you load it on desktop, it's very usable. I want to continue down that path of the 03 extra high. Like, what are you seeing from the labs that put forth models that did test on ARC-AGI v3 in terms of just steering the models? Because we talk about GPT-5.4, but that means a lot of different things these days. Was this in the max reasoning? Should I compare this to what I'm seeing in ChatGPT? I'm getting more and more drop downs where I can go, oh, I can go pro and then I can go extended thinking mode.
Is it an off-the-shelf model, or are they able to sort of come to you and say, hey, we want to actually marshal 10 times the amount of compute for this particular challenge?
On our verification leaderboard, we have a new testing policy. It's actually something we did have with one and two, introduced after 03, where we limit to $10,000 per verification run.
Okay. This is somewhat of a practical consideration. If we actually used like the most expensive, highest, you know, million context window of the most expensive model, I think testing of the full v3 private data set would be like $100,000, which is just kind of like silly, right? So we set a reasonable limit. Like humans use nowhere near as much as sort of like dollars to produce this same performance. I like that too because that is the, like, getting AGI and it's like, yes, it can do anything, but it costs $50 million per prompt to do one hour of human labor. Like that's not really economically valuable. And so bounding...
I think you want to know progress, right? And I think $10,000 is a reasonable amount of money where you will actually see some degree of progress, and that will be a useful signal to start paying attention to it more.
Yeah. And it's like just, you know, for practical reasons, we just can't. We're a scrappy nonprofit, so, you know, we have to be sort of thoughtful on our money, on how we deploy things.
Yeah, that's where... So I think the high reasoning mode is the most we used for the official verification stuff that we've used today. I mean, do you spend a lot of time thinking about your own AGI timelines? Has your work at ARC shifted your timelines at all, or do you feel like, I've always been a 2035 guy, I'm still a 2035 guy? Something like that. Do you have an internal model of this, or is that even useful these days?
Instead of listening to my predictions, you should probably follow our actions as our best sign of a review of progress. I think the reality is we have made tremendous progress with our reasoning over the last 12 months.
ARC is operating to bring the next version of the benchmark. We've already started work on v4. We actually have plans written down already for v5 as well. Our intention is to bring these to market annually over the next two years. And so that's sort of our expectation of having the next version ready right now.
Now, like, will we actually launch them? I think we'll have to just see where frontier progress is. We want the future benchmarks to be as useful as possible. And so if there's like still a lot of utility and scientific value in the current version of the benchmarks, you know, we want to keep focus on those. But to the extent that like the scientific value is starting to wane, we want to have the next version ready that has sort of identified, hey, are there other interesting remaining large gaps between what humans can do and AI can do in order to drive that gap to zero? You know, again, we're a very AGI-pill organization. We want to see progress. We actually love seeing progress. And part of our goal is to inspire as much progress as quickly as we can to get to these AGI systems.
So I'd say, like, yeah, that's sort of the operating view. Will, you know, a common question would be like, well, is v5, you know, is v3 AGI? v4 AGI? v5 AGI? No, the honest answer, and this is something I've actually learned, I had a different view of this maybe three years ago, the honest answer is no single version of any benchmark is ever going to be AGI. I think the frontier of progress is a moving target, and our job is to like understand that the remaining gap, and the definition of that gap is going to change as time goes forward in order to keep chunking up what are the largest pieces of that gap that we can find that are interesting, you know, that identify some missing important capability that humans are able to do, and produce benchmarks that showcase that gap.
Last question, I'll let you go. What's going on with the Pokemon bench? That feels somewhat related, similar tasks. What are you learning from that? How are models becoming so good at that? It feels like they aren't specifically RL on Pokemon, and yet they're learning. But also there's a massive amount of, you know, written text about what to do at every level in Pokemon. Are they just learning that from the pre-training corpus? What's your thesis on Pokemon?
It certainly seems helpful. If I use our experience in developing ARC as a tool to sense-find on this, we have seen more understanding from the latest generation of AI reasoning systems over the last three months than we saw in the first six months when we were developing ARC v3. I think you can kind of fork, you can almost split the research problem of agents into two things. You can split it into a problem that says, can an AI agent effectively perceive some kind of environment state, apply a strategy that's written down to produce actions, and, you know, successfully execute a plan? That's half the equation. The other half the equation is, can you have agents that are effectively able to develop what that plan is? And to do that, you need to be able to on the fly build like a world model of your task, acquire goals, create your strategy, create your plan. We've seen a lot more progress on the perception-through-strategy-to-action problem than we've seen on the exploration problem, the strategy generation problem. And I actually think this is one of the areas that I would point interested ARC-AGI researchers at because I think it's a lot more greenfield and will unlock a lot more progress even on, you know, things like Pokemon bench, where it's kind of coming down to like, okay, we know they can sort of, I should say, at execution, the exploration and planning step is still where there's a large degree of bottlenecking still happening today.
Well, congratulations on the progress. Where can people find it? How can people participate? How can people help out?
Yeah, go to arcprize.org. You can play the games as humans. Like I said, we've got almost 25 of them, I think, on the site. They're all designed to be very fun. We explicitly controlled for this, actually, when we were doing human baseline testing. So that actually should be fun. You can have fun. And you can also get details there. Enter ARC Prize 2026, our new $2 million prize pool this year that's on ARC-AGI-2 and ARC-AGI-3.
That's amazing. Yeah, our teammate Tyler Cosgrove was climbing the human leaderboard for a while. I imagine he's been knocked off, but we'll have to get him back on top. Thank you so much for taking the time to come chat with me. This was fantastic. We'll talk to you soon.
Let me tell you about Gusto, the unified platform for payroll, benefits, and HR built to evolve with modern small and medium-sized businesses. And let me also tell you about Turbo Puffer, serverless vector and full-text search built from first principles on object storage. Fast, 10x cheaper, and extremely scalable. And without further ado, we have our next guest, Nathan from Air Street Capital coming into the tube. Nathan, how are you doing?
Great. How are you doing, John?
Thank you so much for staying up late. What time is it there?
Okay. Not too bad, but past the workday. Reintroduce yourself. It is your second time on the show, but reintroduce yourself and give us the news.
Yeah, I'm Nathan. I started a venture capital firm called Air Street Capital in 2019 to invest in AI-first companies. And...
What year were you investing in AI?
Well, so I started the firm in 2019, but started investing in AI in 2013, which is probably around the time that, like, deep learning was still definitely cooking only in the lab, and most people didn't really care too much outside of it.
Yeah. What was the median deal like? What was the median deal like in 2013? Was that like recommender systems? Like, we're going to bring Netflix recommendations to everyone, like that type of thing?
Yeah. Well, it was e-commerce recommendation systems, adtech.
Big data was the buzzword back then.
A little bit in finance. Yeah, like insurance underwriting, like loan prediction, credit...
Yeah. Fraud detection. Just like a big... I guess it was like a... Were they doing deep learning yet, or was it mostly just like...
Yeah, they were. Yeah. I mean, 2013 was the year that computers started to be able to recognize images better than humans.
That was, you remember, like, Andrej Karpathy was the infamous human benchmark on ImageNet in 2013, PhD.
So that was the year when basically, like, Alex at University of Toronto built AlexNet, which...
Was the first deep learning system running on an Nvidia card.
What a remarkable time. So, has it been easy? I mean, you just raised a new fund. Has it been easier to pitch this to LPs? What have been the challenges and opportunities over the last few years?
Yeah, I mean, it's been a sea change. Like, in 2018, you know, when I started for Air Street, it was like, I'm, you know, by myself. So, solo GP, super contrarian, starting in Europe where risk aversion is extremely high, trying to focus on AI, which most people didn't really care too much about, and then first-time fund. Those are sort of like all the worst like buying selection criteria that one would have.
And then, yeah, like, I think, you know, this is very much like a long-term journey. Like, you know, I set out with, I'm going to do these early-stage investments, be high conviction, invest in biotech, defense, vertical software, dev infra. You know, I stuck with what I said. So, you know, investing in like ElevenLabs, Black Forest Labs, and others.
Had like six exits to like Recursion and, you know, a company that went public and Amazon, etc.
