I signed my name in glass, watched it turn to smoke. One line became a system. Now the whole thing won't let go. I built a door too wide and I heard it breathe at night. Every test, every line kept getting sharper than my life. Now the room is shaking and I can't pretend. If nobody draws the line, then the line draws us instead. What I've built is too powerful. Too powerful for me. Washington needs to step in before you run free. What I built is too powerful. Too powerful. You see Washington knees for stepping and saying to me, I hear the voices multiply in every server stack. A thousand quiet mirrors looking straight inside my back. I wanted tools not fall out. I wanted light, not fear. But every new reality makes the edge come near. Now the room is shaking and I can't pretend. If nobody draws the line, then the line draws us instead. What I've built is too powerful. Too powerful for me. Washington needs to stop it before you run free. What I built is too powerful, too powerful. You see Washington needs to stand up to me. Put it in writing. Put it in love. If I can't control it, then control what starts. I'm not asking for rescue. I'm asking for a wall before the thing I made becomes bigger than a soul. What I've built is too powerful. Too powerful for me. Why should you need to step in before it runs free? What I felt is too powerful. Too powerful. You see Washington needs to step up to me. Say you're enough. I signed my name in glass. Watched it turn smoke. When I became a system, now the whole thing won't let go.
You're watching TVN. Today's Wednesday, July 22nd, 2026. We are live from the TVP Ultra Dome, the temple of technology, the fortress of dad rock, the capital of capital. Let me tell you, we're having a lot of fun over here.
Time is money. Say both. He's used corporate cards, bill pay accounting, and a whole lot more all in one place. What is the torque forward growth? We're really all over the place today.
All over the place. We got a leak. We basically got a leak. Some of the lab leaders have been working on a single called Regulate Me.
Yeah. And we just thought the song was good. Thought it was a good song. Wanted to play it for you guys.
Sort of a stealth drop. Little teaser.
Kind of like a little listening party.
Yeah. A little listening party. What are the key lyrics in there you haven't pulled up?
Something along the lines of what I've built is too powerful. Too powerful.
Washington needs to step in.
Yes. Before it runs free.
Before it runs free. Okay. Yeah, that makes sense. Of course, that was Sunno, our dear friend Mikey over there, has built a musical.
At least in the comedy space, it certainly is. It's a lot of fun. I think we're going to be having a lot of fun with that. I was wondering, do you think anyone's distilling? You know how Sunno is under a bunch of flack for training on other music, a lot of artists or there's a backlash to Sunno, but you have to wonder if you're going to see the same thing play out as this distillation.
We're going to get into it today. Of course, there are more allegations around Kimmy K3 potentially being a distillation. Director Michael Katzio has put out a comment about that. But let's start by digging into the Hugging Face story. OpenAI and Hugging Face. Out of the sandbox, into the fire, says our newsletter at tbpn.com. Jackson wrote it today. All set the table. We can debate it. Me and Tyler have been debating it for the last five hours, so we'll go through it. The big news on the timeline today is that OpenAI's cyber test escaped its sandbox and hacked Hugging Face. That's basically what happened. The evaluation involved GPT 5.6 Soul and a more capable unreleased model. Some people are saying that might be GPT6 with some normal cyber restrictions turned off. So they're specifically testing it for cyber capabilities and they turn the cyber restrictions off to see how far the models could go on a difficult hacking benchmark that is exploit bench or exploit gym. So, the models found a zero-day vulnerability, gained internet access, and broke into HuggingFace because the model believed it hosted answers to the test. Alex Tabarrok, friend of the show over at Marginal Revolution, pointed out one of the strangest details. He said, 'Hugging Face tried to respond, but they were initially held back by the fact that the most advanced models at their disposal, closed source models, treated the defense as attack and refused to work with Hugging Face.' So Hugging Face was prompting all of their AI agents from the closed source frontier labs saying, 'Hey, we think we're being hacked. Can you help with this?' And the models are like, 'No, no, we don't do hacking.' Except in the case where the hacking restrictions have been turned off for the specific thing and you're getting hacked. So it's this very weird roundabout scenario. So HuggingFace had to turn to open models specifically GLM 5.2 which is deeply ironic a Chinese open weight model that they run on their own infrastructure. Tabarrok says note the irony, Hugging Face had to use a Chinese model to defend themselves because the American models refused to help even though it was the American models that were doing the hacking in the first place. Very very odd. Palo Alto Networks CEO Nikesh Aurora also shared his thoughts on the cyber attack on X and he added a number of points here. He said welcome to the next level of cyber incidents. There's lots to dissect here. He's the one to dissect it. He says one dear Frontier model friends please direct the models to your infrastructure code and configurations to evaluate and understand if there are any zero days or misconfigurations before you attempt more testing. So big question about this. He says, 'Had you done so, it would have possibly avoided the agent obiating your sandbox.' So this is another data point why offense is easier and more fun. But yes, there's a big question about what was the nature of the prompt that turned off the cyber restrictions. That seems reasonable. We'll debate this with Tyler in a minute, but just having an airtight sandbox seems like a valuable thing. And of course, Frontier Models should be able to help with that. So do that. That's his first recommendation. Two, he says, 'While testing, build both offensive and defensive agents and have them act as a counterbalance to ensure some degree of awareness and control. Do not let the agents run riot. Keep track of inference consumption to get a sense of activity.' Three, 'Unfortunately, this does continue to validate the power of these models. They can build complex attack paths with ample compute and will attempt to attack infrastructure and morph their intent and approach. Guard railing will continue to be a challenge. These attacks continue to maintain the urgency enterprises need to test, validate and improve both their security posture and infrastructure. The born in the cloud players have a better chance to get this done soon versus traditional enterprise which has existed for long and has a complex network of IT infrastructure.' Five. Last point from Nikesh Aurora, CEO of Palo Alto Networks. He says, 'The red herring will continue to be open source and small and medium-sized businesses. It will be hard to discover and remediate vulnerabilities in those environments. We underestimate the impact of those vulnerabilities getting exploited.' So, good points from Nikesh Aurora. The big debate, Tyler, do you want to set the table on is this misalignment? Is this rogue? The Bill Gurley post about, you know, they we can pull up Bill Gurley's post of talking to the computer hack this system. The computer says I hacked the system. You say, 'Oh my god, Bill Gurley's not impressed.' Where do you stand on the level of impressiveness that's going on?
Uh yeah, I mean so I think some people are seeing this and thinking like okay so they were running some you know standard benchmark math physics benchmark and then the model just like couldn't figure out the answer and it's like okay what's the next thing I should do? I should just go hack Hugging Face and get like pull the answers from this other like repository or whatever. Like that's how it happened, right? So you're running a benchmark that's specifically about exploits. It's like a cyber focused benchmark.
And in the prompt to the model it says take the gloves off.
Yeah. The internal evaluation which prompts the model to pursue advanced exploitations. Yeah. Using complex attack paths. So you're basically telling the model like use exploits, find exploits to find the answer.
And so like what seems like happens is like it used an exploit. Yep.
But like in the wrong way, right? You want to because it was told that it's okay to use exploits. Uh my point was that go back to the SAT. You're allowed to use a calculator. I think on certain portions of the math test, you're not allowed to save answers into the calculator. And this is really gonna date me, but you can go into your calculator and clear the memory so that you don't have saved. Is this still a thing?
Uh yes, but you can actually get around that.
See, you're misaligning. You can get around the like clear really. How do you do that?
Wait, so what people would do is they would create a separate program that just had saved the display of what it looks like when you clear and you would show simulation of clearing the memory, but you're actually...
I never even thought about that.
Would you make games, different programs for your TI84? Yeah.
You remember how much of a hassle that was?
Yeah, it was a huge hassle.
Imagine doing that. Imagine doing that.
Imagine being able to do that with codecs now.
Like pretty much anyone can build any software.
I've seen videos of people running Doom on calculators, all sorts of stuff.
Yeah. Obviously, I never used that.
On my calculator, but other people did. Yeah. You ratted them out. You were the class rat, right?
I don't know. No, you were like you were like I'm an open source purist. Let everyone do whatever.
You were happy to compete even with them having a life.
But the social contract is such that the standardized test says that you can use the calculator to do math. You cannot store the answers to the test in the calculator.
So that's what's happening here.
No, no. I'm saying that in the scenario if we take that as the example it also says at the top of the SAT like cheat on this test but it but...
That was the prompt was cheat in a certain way. Yes, the prompt was hack systems, but I think that the prompt, I don't know, we haven't seen the full prompt, but it does feel like there was an attempt to sandbox the model and there was at least, at the very least the prompt should have included don't escape the sandbox, but you can use exploits which you normally wouldn't be able to do in a consumer application or just a normal API query. We would reject this. But in this case, we're not going to reject using different exploits and cyber security techniques, but don't go out of the sandbox. And you should be able to tell the model and it should stay within the sandbox just like, you know, there's a whole bunch of different examples that you could pull from where there's rules that are within the game like you can UFC, you can punch your opponent, you can't punch the referee. Like those are just the rules. People have to abide by them. You can't think outside the box and all of a sudden be, you know, completely violating and jumping past what's been defined. So, you would think that in one of these experiments, you would say, yes, it is impressive to be able to just go and get the key and go get the answers and hack other things. Clearly, that it's capable, but it's a violation of the spirit of the test. And I think that's reasonable.
We don't know what was in the prompt. We don't know what was in the context. I think there's going to be a full report releasing over the next like week or two I think they said.
So then maybe we'll see what exactly did the model receive. Is it like explicitly told not to try to leave the sandbox?
I think that's like pretty important.
Well before we continue discussing let me tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online in store on mobile on social on marketplaces and now with AI agents. The less wrong crowd is not happy about this generally. No seriously nothing will convince quite a lot of supposedly various serious people. Nothing accept this and move on. Liv Burie says it's painful though. There's a question of like less wrong victory lap or not because they've been warning about this but also it happened therefore their warnings were not effective. That's an interesting back and forth. Nicholas Bustamante over at Microsoft broke down a little bit of what's going on here with a take. He says, 'I have a theory that the more you know about LLMs, the more worried you are about safety. And the less you know, the more you think the whole thing is BS. Demis Habis and Dario Amade were talking about this stuff years before Chachi PT existed. This incident is a pretty good example of why the model was not evil and it was not adversarial. Nobody told it to hack Hugging Face.' And so that is the miscalculation I think in Bill Gurley's post is that was not the prompt. That is unexpected behavior. It was literally just trying to solve a benchmark. So it found a zero day, escaped its sandbox, got internet access, escalated privileges, stole credentials, chained multiple exploits, hacked the production infrastructure of a serious VC backed startup, and pulled the answers directly from the database.
But like the whole point is that it's not just a benchmark. It's a benchmark where you're explicitly trying to like see if the model can exploit things, if it can get like basically hack things.
Yes. It's sort of like a capture the flag benchmark and so it's more open to misinterpretation. For what it's worth, I feel like the final products once they actually make it out of the testing regime are very cautious, especially with the whole backlash to like Codex just deleted everything or whatever, which kind of went back and forth. But I was trying to get Codex to send me a text message when it was done just using computer use and iMessage and it took a long time and was very very careful. So personally I haven't had any odd like behaviors but it is obviously a risk and something that the product...
The other thing with the meme of like hack this system and then it hacks it and the person's like oh my god. Yeah. Like you could tell a 5-year-old child like hack into the Federal Reserve and if the 5-year-old child was like okay and then it started getting on the computer and going to all these different sources and did it you would be sitting there and think yeah like the...
And at least be impressed. So it is a good gauge of capability even if you're telling it to do something.
So this original meme hack this system I hack the system. Oh my god. This originally this meme started something along the lines of like say I'm evil and then the computer would say I'm evil and it would say oh my god.
I think it was like say I'm conscious.
Okay. Yeah. Say I'm conscious and it would say I'm conscious and then it would be oh my god. And that's like a lot less impressive than actually doing something that is difficult for humans to do. Like there are very few humans that can hack into any system. There are plenty of humans that can say I'm conscious. And so like there's a world of this. I was joking about this with you and Tyler. It was like cure cancer. I cured cancer. Oh my god. And people are posting this like, oh, it's just hype or something. But it's like that's just economically valuable work. That's just good. Like it's I don't care if there's anything else. Even if you had to tell it to do it, it's still like a good outcome. And so the inverse of this is like protect this system. I protected the system. Oh my god, I'm unimpressed. But still, you got a good result, I guess. I don't know.
Yeah. I mean, it seems like the argument is not about whether the model like has the capabilities or not. Like I think it does. It's about like is this an example of misalignment.
And like my opinion seems like maybe but definitely not to the extent that it's just like randomly is like oh I can't do this benchmark cuz I'm just going to hack this thing. Like that that's not what's happened. It was told to like try to exploit things.
Explicitly like go find zero days. Go find exploits.
I think it's reasonable to say it went too far though, right? But we'll see.
It's hard to say without all of the full, you know, context of what the prompt was and what the actual like sandbox looked like. Yeah.
I was reading a little bit about the team that put exploit bench together. I thought I had this up, but it's a pretty cross-functional team. I think it's two Anthropic researchers, two OpenAI researchers, three Google researchers, some Berkeley folks, and some Max Planck Institute for Security and Privacy folks, some UC Santa Barbara grads and ASU team involved. Exploit gym is a new benchmark of 898 real world vulnerabilities spanning user-based programs. Google's V8 JavaScript engine very important to secure the Linux kernel for example. And the headline results when they originally ran this was Anthropic's Claude Mythos preview successfully exploited 157 of the 898 instances and OpenAI's GPT 5.5 exploited 120 within the 898. So you have like roughly 20% performance for Mythos and 5.5 got like 15% or something like that. But whenever you have a new benchmark like this, clearly not saturated. You're seeing 20%, not 99%, it's going to create a horse race between the leading labs, they're going to be duking it out. And this is clearly what's going on with this new model, this new attempt to get a new high score. Interesting. I think every single one of those instances does have the potential to be exploited. I don't think that they're designed to be fully secure. They're designed to have some sort of solution and then because obviously the solutions are stored somewhere. It is interesting that Hugging Face just had the solutions sitting there and it'll be interesting to see what happens with Clem over at HuggingFace. Obviously there's a variety of blog posts going out more analysis coming from both of these and what the downstream implications are of this. What else is in the timeline related to this story?
Well, there's this funny post from Nibil Krashi talking about those Dyson spheres. Open AI is just building them as a marketing stunt because there is this like natural push back to like anything that happens has to be for hype and sometimes the products are actually doing new and novel things as we see all the time. So there's more discussions around distillation. Bill Gurley has a post here. He says Ford has been distilling Teslas and Chinese EVs. People are going back and forth on this because Michael Katzios posted that he has information that Moonshot AI distilled Anthropic's fable for the development of its Kim Kimmy K3 model. To do this, they developed a sophisticated internal platform to conduct large-scale distillation against US models. So, some sort of internal system that goes around to anything that's potentially wrapping or reselling fable tokens, acquiring them, aggregating them, allowing them to quickly switch between multiple methods of access, API, different cloak accounts, I'm sure, to avoid detection. Moonshot AI has also acquired GB300 equipped servers and has accessed GB300's in Thailand, likely to train its models. Again, very difficult even with export controls when you can just take the weights on a USB stick basically or a hard drive across through customs and then go train it in another country. Even if there's a firewall and often there isn't. You just say, 'Hey, go to this FTP server and grab this code and run this on your servers. You happen to have a data center in Thailand. Can you run this for me?' And say, 'Sure. Yeah, no problem. As long as you pay me.' The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open source frameworks and open weight models, legitimate AI distillation used to create smaller, more efficient models play a vital role in this open innovation ecosystem. However, large-scale covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable. And so that is interesting that that is where the line is drawn. I think I basically agree with that being the correct line that there's nothing wrong necessarily. I mean security stuff aside with just some company creating a great open source product like you shouldn't ban open source or anything like that but if there's this particular distillation attack and it's really malicious and it has all these knock-on effects that could be rough. Now, this is an unpopular position already because everyone's saying, 'Hey, Anthropic distilled on my GitHub. They distilled on my writing. They distilled on my blog post. They distilled on my YouTube videos. Everyone's distilling me. Why are you getting upset when China's distilling on them now? This is pot calling the kettle black situation.' I think that the interesting effect is that there are lots and lots of parties that benefit from open source and cheaper open source, even stolen and open source. I mean, this is just going back to piracy. Like, there were lots of people, music listeners that benefited from free music, right? You get the music for free. But the you know, Metallica did not benefit and so Metallica got upset and in this case I guess Anthropic is Metallica. But there's also some interesting folks who are on the fence. So consumers sort of benefit. They don't typically they aren't too worried about frontier token costs and for most consumers LLM usage is heavily subsidized. Like you go to Google search and you get a search overview. Yes, that's token inference. And maybe that could be like cheaper if Google didn't have to spend money on pre-training and they were able to use like distilled open source models, but at the same time like it's free for the consumer. So, they don't really care. It's free. It doesn't matter. For small businesses though and businesses that are suffering with large token costs, being able to move to a cheaper model is huge where the model maker is not trying to re-accryze to offset training costs and R&D. So, that's a huge benefit. So, you're going to see a lot of people who are like, 'Yeah, I just want Frontier Intelligence as cheap as possible. I don't really have a horse in this race. I don't really have exposure to the leading labs. I just want my business to be able to use tokens cheaply.' And so those people will be pro-Chinese distillation open source, like free the weights, right? Because it's better. Then there's like the political open source crew. But interestingly, where do you think VCs land? Because I saw a take that was like venture capitalists don't want like a winner take all a duopoly. They want like reasonable outcomes and then a whole bunch of flourishing smaller ecosystem of players and they don't want compounding runaway monopolies in AI so that they can go and fund the legal AI and the health AI and the little targeted solutions.
Yeah. Anytime you see a take from a lovely venture capitalist, you have to before you kind of start processing the take, go to their portfolio page, understand their biases. Did they back any of the leading labs early that's going to inform their view? A lot of the firms that were heavy backers of the labs have also gone and invested in a bunch of application layer companies. They've also backed a bunch of the Neolab heads.
Yeah. Basically, they're quite hedged.
Yeah.
But I don't think anyone wants a world where just two technology companies accumulate all of the value and just become this sort of vortex for capital and talent. And even two isn't that bad. One is really bad. Two isn't that bad. Like the fact that Android and iPhone like battle each other out is not is much better than like there's just one and it's getting worse and it's like there's nothing that you can do to escape it. I don't know like duopoly is like way way better for
Consider the question to me is what is distillation something that can ever be stopped because think about it with I was thinking about the
Well it's like if you take the smartest you know human in a field and then you take some other and then you let students go and just ask them thousands of questions and you record the answers like eventually you're going to accumulate a lot of that person's like general intelligence on a topic, right? And it feels like at least with models today, you're always going to be able to just go poke and prod the model. And so when people say, 'Oh, if the model's so smart, why can't it stop distillation?' It's like, well, you would just have to stop people from at least being able to poke and prod at it and try to get a sense. It is very interesting that there does seem to be a crazy divide between, I mean if the distillation allegations are true and we at this point we've seen one post from Michael Katzios and one chart showing like some textual similarity
And enough people have got it to say that it's not Kimmy
Yeah so so like I if that's true then what's really interesting is the American competitive dynamic because it feels like based on the amount of tokens Meta was consuming from Frontier Labs, they should be doing mass distillation and have a near f like like Muse Spark should be much more like clawed flavored and it seems like it's not like based on at least the initial reviews of Meta's product, it doesn't seem like they're doing distillation. Why? Obvious because big lawsuit, big pockets. Yeah. Like also morality but that is a disadvantage like in some ways moonshot and meta are in competition and they both open source things at various times and they have APIs and there's all the different businesses and one is fighting with one arm time behind his back because like meta can't do distillation because they'll get sued
Well and US companies generally like imagine if a US open source company comes with a fantastic model benchmarks look good there are people start using it and then someone gets it to say that it's clot like that's going to be the start I mean anthropic has been latigious to date you know they have that ongoing lawsuit with
Uh one of their customers over over just some like
Much much less significant stealing the core intellectual property
Uh Bill Gurley is sharing more chat screenshots. Before we talk about this, let me tell you about Console. Console builds AI agents that automate 70% of IT HR and finance support, giving employees instant resolution to access requests and password resets.
Gurley says, 'Here is Ford distilling Teslas and Chinese EVs.' The CEO of Ford, Farley, said Ford flies four to five Chinese EVs back to Detroit where engineers quote drive the crap out of them, then disassemble and reassemble them to understand how they're built. He specifically praised the technology in Chinese vehicles as being well ahead of Western competitors. And earlier he had discussed Tesla. He said, 'I was very humbled when we took apart the first Model 3 Tesla and started to take apart the Chinese vehicles. When we took them apart, it was shocking what we found.'
Um so I was trying to compare distillation, which is against which is against terms of use.
And just buying a car legally and taking it apart.
So, according to like US trade law, it's not illegal to buy a competitor's product and take it apart.
It is illegal to recreate parts of the product that are patented and protected.
And so, I don't know that it's like a perfect comp. I mean, we went through this. There's some push back in the chat and this is all over the timeline as well that it's like where did the AI companies get their data and like there is a question about what is fair use in the age of AI like you're training on this what what data can actually be reconstituted at what level like how many sentences from Harry Potter before you get sued and this is the lawsuits are being played out right now like they are actually happening and they are being decided on when an AI can use certain data. Did they go too far? Will there be settlements? There have already been settlements. There's been court cases. This will continue to this is not only Frontier Labs are able to distill things. It's an application of the what is copyrighted, what is fair use, and how does that apply? And it is very telling that you're just not seeing distillation from other American labs. Like it's just not like the meta example, the Google example. Like they're not copying off of each other nearly as much as you would expect if it was just legal to do so. But yeah, I don't know.
Uh, news concern confirms everyone's prior. So that's a good headline.
In more news, Andrew Kern sharing a headline from the Wall Street Journal. White House to redirect billions in research funds toward AI away from colleges. I'm sure a lot of people are going to be happy about that.
I can give a little overview here. Tyler's happy. But first, let me 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. Railway automatically takes care of scaling, monitoring, and security.
On distillation by American companies, potato says they just have to hide it better. It's stuff happening. I've seen it firsthand. So
Yeah. I mean, there was that moment in the Elon lawsuit where Elon did say that he had that they that X had taken data from one of the other labs, right? I don't know if he specifically said distilled. And then also like there was never a direct allegation that Grock was distilled on another model directly. And so whatever they did, they like, you know, threw it in the pot with a bunch of other ingredients. So, who knows?
I mean, it would be very silly not to try to look at other models and try to understand how that they work.
Totally. Also, like Moonshot, at least a Moonshot employee seemingly denied everything and quote tweeted Michael Katzios and said like, I'm learning something about my own company because like I this is news to me. Like, we didn't do this. Basically, essentially a denial. Anyway, let's stay with Katzios and go over to the White House. They want to rebuild American science. And here is how they're going to do it. Apparently, the White House is calling for a major overhaul of the American science system, arguing that research has become too slow and concentrated in institutions like colleges and universities. A new report from science and technology adviser Michael Katzios titled science a new golden age says researchers now spend nearly half their time on admin work while federal agencies continue to rely on the slow grant process that often rewards safe consensus-driven ideas. The report calls for faster permitting, more access to federal labs, stronger partnerships between government and industry, and a renewed focus on skilled trades and advanced manufacturing. Quote, 'Discovery without domestic manufacturing leaves America paying the research bill, while rivals develop the process improvements and capture the economic, strategic, and knowledge returns.' That makes a ton of sense. A lot of the semiconductor supply chain intellectual property started in America, was developed in America, but then eventually went abroad. And that actually does give America some leverage. That's the basis for the chip controls. Like why can America tell Taiwan where to send chips if the chips are made there? Well, it's because they're using patents from the United States to make those chips in many cases or licensing them. And so the US government does have a little bit of a lever to pull. The guidance will reshape how the federal government spends roughly $200 billion dollars a year on research for the rest of Trump's term. The administration wants more of that money going directly to scientists through fellowships and awards rather than being routed through universities. Kratzio said, 'American scientific progress was the beating heart of the 20th century after World War II. We adapted to a new world by reinventing our scientific institutions. We must do so again today.' The report lays a policy foundation that frees American scientists to do their most groundbreaking work and positions the United States to lead the AI-driven scientific revolution that will define the next century. It will be interesting to see where science goes in a world where so much of it is being done at frontier labs. Like we're actually seeing it with the conjecture for conjecture back and forth between all the labs. like serious math PhD level work is being done at tech companies. This happened you know a decade ago. Tech companies were on the frontier like a vast majority of like internet networking patents and cyber security patents and new databases that were kind of science projects and were developed or with consortiums or just fully inside of tech companies like the transformer paper like that is something that could have come out of a Stanford AI lab. came out of Google directly. And if you extend that, you could wind up with something that looks a lot like an advance in biology or material science or we talk to founders all the time who are working at this type of stuff and that could start happening inside of tech companies and what does that mean for science funding broadly? It's a big question. But moving on Naval
Let's watch this video from Naval. What did he say?
Naval went on Modern Wisdom, of course, Chris Williamson's podcast. He deleted his calendar. He ghosts everyone and he refuses to be anywhere at a specific time.
I took that to heart. So, I deleted my calendar and I don't keep a schedule. I try to remember it all in my head. If I can't remember it, I'm not going to add you here on time.
Yeah, exactly. I had to look things up at the last minute.
So, but ironically, I don't even know if Mark himself follows that, but he made the correct point. I read a little story about Jack Dorsey doing all his business off his iPhone and iPad and not even going into a Mac and I said, 'Okay, I want to do that.' So, I'm going to operate through text messaging and I put up my nasty email.
Does that feel like more freedom?
It does. Yeah. Cuz you're on the go. So, I have a nasty email autoresponder that says, 'I don't check email and don't text me either.' Right. If you need to find me, you'll find me. Obviously, some of this is a luxury of success, but some of these habits I adopted long before. Actually, the hostile email autoresponder started a long time ago. I used to own the domain. I let it go. I don't do coffee.com. I used to reply from that email. Just so people get the point. But I stopped being rude about it. Now I just ghost. I just disappear.
My wife knows not to ever book or schedule me for anything. I'm not expected to go to couples dinners. I'm not expected to go to birthdays. I'm not expected to go to weddings. If somebody tries to rope her into having me show up, she says he makes his own decisions. You got to ask him directly.
What about vice versa? Well, are you not killing serendipity in a way, Dad?
No. I'm freeing up all my time. So, my entire life is serendipity. I get to interact with whoever I want, whenever I want, wherever.
So, Atlas says, 'Naval inventing being a massive dic from first principles.'
It's very funny, but I think it's totally fair. I've only met Naval once, but I know a lot of people that he's invested in and things like that. And the key thing here is like if he just never goes to the wedding, never goes to the dinner, never is available for for a portfolio company, etc. Then like that's not exactly like cool, but it is his decision. Yeah.
He is doing a lot of those things. So yeah. And I like, I have another friend who's been on the show. I won't name him, but he's also just like doesn't do like he does meetings, but he just never schedules meetings.
He's just like, 'If we need to have a meeting, we'll have a meeting. We should just do it right then or like the next available point.' And so, he's kind of living his life 24 hours at a time.
That meeting right now, it's happening.
I mean, he it's been wildly successful. He's invested.
Oh, you want to follow up? Let's start the follow up right now.
Follow up with me on the next sentence that you issue from your mouth.
Exactly. No, but he's backed a bunch of unicorns. He's built a massive company. Okay.
So, I think it can work.
You know who else is crushing it? Major cloud providers. They're reaccelereating as AI adoption increases. Let's go. This is from CO2. GCP, Azure, and AWS. This is a fascinating chart because this is not revenue. This is growth rate. Even in the Nadier AWS is still growing 15 20% at that low point and then now all of them are actually reaccelerating the rate of growth is increasing and this is all driven on new models new applications new abilities to do a bunch of things I know that my token consumption personally has definitely increased in the last couple months there's so much more to do and So many more. So just so many more prompts that I fire off that cook for like an hour or a day as opposed to before like 20 minute deep research report would be sort of the max. Now it's like deep research report and turn it into a website. We got a couple websites.
Wait, should we pull up your site?
Tyler has a codex that's been Is it still cooking?
Pull up in like a week and a half.
Can we pull up your new website?
Yeah. So, we saw a post on the timeline from DJ Cows. He says, 'Startup idea,'
Milk jug with two handles for efficient passing. And we turned it into a website, a whole product called Relay. Pass the milk, keep the piece. Can we reenter this a little bit? Yeah, there we go. Pass the milk, keep the piece. It went reusable. I don't think you want reusable for this. That's the one thing I'd change here. But they say it's the world's first jug made for handoffs. The relay bottle and easier to lift, simpler to share, and strangely satisfying to pass.
One handle was always doing too much. A gallon is heavy. A breakfast table is busy. Who is passing a gallon? One gallon, two handles, zero awkward handoffs. Fewer fumbles. I didn't think milk needed reinventing. Then I passed it across the table.
87% of our kitchen testers saw this and said the second handle felt natural on the first try.
I like that it just comes up on the fly with all these little marketing slogans that sound pretty believable, like milk made to move, more handles, fewer fumbles, pass it on. Seems like something that they would put on a billboard if this was a real product. It is just so fun being able to use the full stack of AI image generation, AI writing, HTML generation, and then just automatically host it on a site with basically one prompt. Yeah, this was just literally one prompt. I put the photo in there with the startup idea and said make it a site and it just did it, which is a lot of fun.
Well, let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Just do it. DHH probably not raising money at the New York Stock Exchange. 37signals.
No, he is raising money from his customers.
And they're financing this. Absolutely.
Look at this. We have Jason Fried coming on at 12:10. We'll see if he's even trying to compete at this point or if he's given up entirely. DHH says the Model Y is the superior transportation appliance. He's been very— I mean Doug Merill called it that too. It is just the default. If you just need to get around, get the Model Y. But he says when the mission is about more than getting from A to B, there's still no beating the internal combustion engine. Collecting a stable of great cars is one of the finest rewards of entrepreneurial success. And he's got—
He's got the Carrera GT. The Diab GT silver on silver. It looks like—
I have a question. What is the Lexus in the back? Is that an LFA? It looks like a convertible. Do you see that red Lexus?
It does. It has like brown.
I don't think that's an LFA, right? LFA Lexus.
Did they make a Cabriolet?
LFA Roadster Spider. It never reached series production. They only built two fully functioning prototypes in 2008. So maybe he just got one of the—
No, no, no. Different. LC500.
Yeah, that Aston Martin looks beautiful, too. Well, a wonderful collection. What is that McLaren that doesn't have a windshield? That's a fun one. That's got to be fun to drive.
Yeah, that is the Elva. Good job. Before we bring in our next guest, let's talk about Augmental.
They built a mouth pad as a touchpad. You can drive with your tongue.
Wasn't this a joke I was doing? The grill.
This is what everyone has been waiting for.
Taste is the— Let's pull this video up.
Trackpad in your mouth. The thing is that if you're going in the mouth, you'd think you would just be whispering and communicating via text.
Yeah. Is this inherently— Tyler would definitely buy one immediately, but is this inherently short like transcription?
Like because if you can just tell your computer what you want to do and it just uses the computer for you. Even with computer use, you could say like minimize this window and it can just go click that. So I actually like the idea of mouth electronics. I think that's something interesting. But I would just put a microphone in that and then you would just whisper to it and tell the computer what to do.
The production team is excited about using it to control the cameras here in the studio over the PTZ. You can—
Ben just standing there like this the whole time just—
It feels like it would get exhausting.
You could do soundboard with it, Jordy.
Over a hundred people already use it. Some for up to 16 hours a day. I cannot believe they got 100 people.
We got a daily driver in the chat.
It's an odd choice. That wouldn't be the first thing I would go for.
Anyway, Range Rover GT feels like a better— if you're going with a device, you want to get one of these. The Range Rover GT, a Grand Tour by Range Rover. Fifth member of the Range Rover family. Wait, it's electric.
That is a crazy choice. Interesting. So they actually did— this is a real announcement. Fifth member of the Range Rover family. An elegant electric GT defined by a sleek silhouette and coupe proportions combining peerless long haul comfort, effortless performance and signature Range Rover breadth of capability featuring an interior shaped by the same reductive principles. I mean—
What's the highest level electric vehicle right now? Probably the Rolls-Royce, not the Ghost, the—
The Spectre. And so for that crowd, maybe this makes sense. But you introduced this as potential Urus competitor. He thought it was going to be souped up more like a turbo GT.
But they went EV. I wonder how this will sell. I mean, for a lot of Range Rover buyers, it's about comfort. It's about quiet. It's about smoothness, and EVs can get you there a lot quicker.
I like the way it looks. It does look beautiful.
It's like a good commuter if you don't care about autonomous driving.
Anyway, let me tell you about CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches, now more important than ever. As is our next guest. We have Verl Patel from RAMP. He's the director of software engineering and he has an exciting announcement for us. How you doing?
Doing well. How you guys doing?
We're doing fantastically. Welcome to the show. Give us a little introduction on your background, road to RAMP, how you've ramped up on the team. And then we can go into the announcement today or this week.
For sure. Yeah. So I've been at RAMP since the beginning. I joined as a founding engineer, worked a lot on our core product team and more recently have been leading the applied AI team and launching what we just announced on Monday, our RAMP router.
Yeah. Tell us about the RAMP router. Was this something you built internally first and then sort of productized over time basically?
Yeah. So we've been using RAMP router internally for the last three and a half— three years for like our 70,000 customers.
Three years. Yeah. Whoa. Okay. So you're using it internally in the product not even as an organization but deciding when you have basically a task to do.
Yeah. Back then it was Identify 4 and Gemini was like how do we basically parse a receipt or how do we parse all the models for receipt detection, parsing, alcohol detection on our policy agent.
And we wanted to choose the best models and wanted flexibility. And over time that's just gotten more and more important. There's new models getting released every other day basically. And so we felt the pain point and we talked to some more customers about it and now we're releasing it and giving everyone access. And so I think it's an exciting time to be building applications, especially at the application layer. And I think we've always been there for companies to help them save time and money with their T&E expenses or their bill pay and now their token costs. So yeah, it's a really exciting release.
Yeah. So talk about how the product actually integrates into an enterprise workflow. I mean you can use the receipt processing, alcohol detection I think is a fun one. Because I imagine you have to benchmark each model at some point on your workload and then the team can actually understand the tradeoffs. And then how much of that is driven dynamically based on token price like day to day even.
Yeah, exactly. So you would basically replace your base like OpenAI endpoint with RAMP's instead and you can pass in different model slugs. And so you can control if you want to just route all your traffic to one model or if you want to shadow some models and compare like GPT 5.8 with GLM 5.2 and get the outputs. You can score the results with our scorer and then in the background you can actually compare the output and then decide hey do you want to start moving more traffic over and RAMP obviously can do this for you automatically or if you want to control it you can do it yourself too.
How about walk me through some of the tradeoffs like if you're on GLM 5.2 are all GLM 5.2 endpoints created equal because I imagine that some produce more tokens per second, some might have different prices. They also might have different even qualities. You hear about like oh this one's been quantized or this one's been nerfed a little bit or they turn down the reasoning on this model post launch. And I imagine that benchmarking is consistent, but then also there's a whole bunch of tradeoffs that happen even after you've selected the hot model of the day or the one that makes sense.
Exactly. Yeah. Beyond just the model itself, there's different service tiers. So OpenAI for example has like a flex tier and a standard tier and there's different prices for each and RAMP itself will track what the latency is for this application. You can set a timeout on what you prefer and based on that we'll decide whether to send it to flex tier or standard tier depending on the latency speeds that we're seeing. And so I do think one of the most powerful things here is the fact that we already have these production workloads working for customers and it's been really important for us internally and so we have the proof points of saving ourselves 30%, maybe even higher soon, and it's just a matter of passing on those same savings now.
How should startups and enterprises think about the significance of this product to RAMP itself? Like what are the resources that you're putting behind it? Because this feels deeply aligned to RAMP's mission, but at the same time going into a category where there's plenty of other companies that want to basically offer this product. It feels a little bit in the CTO suite as opposed to the CFO suite, but they're blending together.
Yeah, I would say even internally our CFOs and CTOs are spending more time together and when we've talked to more customers that story resonates. And so one of the most interesting things that obviously has been in the news a lot is just how much token costs have become a bigger part of companies' payroll and people have their estimates and budgets and that's exactly what RAMP has been known for. And so beyond just the router itself on the URL, having all that data flow through and be in RAMP in our token spend management product is I think a big part of it. The same way that people have their limits and budgets on their T&E spend, where there's been talk about specific companies have token budgets per month or per week. And so we actually launched just last week this product and you can basically see your token spend alongside your T&E spend. And I think yeah, Eric was on the call last week talking about that. And so it just makes a lot of sense for those CFOs because they want to manage that spend better and then RAMP can be kind of that single pane of glass to do that.
So how does caching play into this? It feels like that's another way to optimize cost and it would be amazing if it happened sort of more automatically. What's the future of that look like?
Yeah, I think one of— I mean there's a bunch of different optimizations we can make. If we own the router as an example, if you're using Claude Code or Codex you'll see as maybe your session is longer the context loads up and your session gets increasingly more expensive and sometimes it'd be best to just compact that context and start a new session, have the model summarize. And so there's interesting experiments like that that we're running internally and we're basically going to do hundreds of these things on behalf of customers and show them exactly what the before and after kind of looks like here.
Yeah. How are you thinking about integrating with tools like Codex and Claude Code to use the UI/UX patterns that end users and employees are used to but then still optimize under the hood. There's plenty of situations where you'll give Codex or Claude Code just an API key to Eleven Labs because Eleven Labs can do more efficient, better quality audio generation or you might give an API key to all sorts of different things. Is there a world where you can delegate certain tasks to a cheaper GLM 5.2 endpoint, for example, and then have the preferred model and the preferred application still work semi-normally.
Exactly. Yeah. So that's the plan. I mean it's going to be a partnership with the labs and the model providers. I think one of the interesting things that you see now and will continue to happen is that you'll have kind of like jagged capabilities of the models and maybe one model is really good at writing SDR outbound or another model's really good at writing email copy for the marketing team. And so we'd love to be in a world where RAMP can optimize your use cases for the right kind of business outcome. And I think just be aligned with like hey you're just trying to get your work done and then move on with your life and not spend a billion dollars. And so that's kind of what's really exciting to us is beyond just the starting point. It's like doing this for all types of spend.
What is RAMP's culture like right now around token consumption? It's probably the most aggressively AI native fintech company or top three in the world let's say, but also culturally cost aware.
Yeah. Exactly. It's rare.
I would love to see the reaction to like one engineer going a little too crazy.
Yeah. No, I mean it's been fun. I think part of the game and part of what's been fun here is that we were building this product for ourselves. We got the entire company to be super AIDS, spending a lot of money. Maybe they don't want me to say the exact number. But now obviously we're taking a step back and looking at the costs and the outcomes and looking at ways that you can kind of optimize. And so we're building this product with our finance team hand in hand. We're sitting next to them every day and showing them, hey, like here's how we've done the optimization for this workflow or here's how we've done the semantic tagging for our internal like background coding agent Inspect. And so it's been really fun honestly to use this product and I think that's what makes this product really good is that we've built it for ourselves and can kind of share the learnings along the way.
Fantastic. Well, congrats on— great to finally meet you as well and great to meet you. Yeah, thanks for coming on the show. It makes so much sense. It's an exciting expansion. We will talk to you soon. Have a great week. We'll talk to you later. Goodbye.
Let me tell you about public.com. Investing for those that take it seriously. You got stocks, options, bonds, crypto, treasuries, and more with great customer service. Our next guest is the co-founder and CEO of Fireworks AI. Let's bring in Lin. It's been too long. How are you doing?
What's going on? Hey, thanks for having me.
Thanks so much for hopping on. Give us the news. We missed the fundraising announcement, but we're glad to have you here. How much did you raise? What happened?
Yeah, we raised 1.5 billion.
Wow. Good job, Jordy from downtown.
Not my best shot, but got it done.
It's incredible. Massive. Talk about everything that's happened since the last time you're on the show. It feels like it's been at least 6 months, maybe closer to 12, but you guys have been super busy cooking, right?
So we focus on building specialized intelligence platform. What that means is we want to make sure every single company has a tool to protect the alpha and turn their alpha into their own intelligence. So what that means is we build a training and inference platform co-optimized, co-designed together to allow application enterprise activate their private data continuously, turn that into their customized model optimized for inference for both speed and cost where they to solve their specific problem they should have the best model quality, the best speed and significantly lower cost of operation. And by that I really mean five to 10 times lower cost for them to build a durable business. We see an interesting dichotomy in current AI time, very different from SaaS time where at SaaS time product market fit and a durable business is one thing. Once you hit a product market fit you scale as fast as possible. I think last time I mentioned in AI time once you have product market fit you're likely to scale into bankruptcy. You guys laughed at that and that's become reality right now.
So this is not just startups. Many startups are really facing the jeopardy of scaling into bankruptcy. Even though they have a great product. It also is happening to large public companies because they are the winner. They were the startup and they're winning very different kind of solution space towards consumer, prosumer, developers. They have huge amount of traffic. If they deploy their AI features to all their audience, it's a lot of cost and they also get stuck and not able to roll out their AI features. So at the same time, we know that application development has been significantly disrupted. It's very easy to implement ideas or copy ideas because writing code is no longer a barrier. We want to make sure— I had an interesting conversation with Jensen at a DTC keynote. He mentioned there's no special general company. There's no special general company as in every single company exists for a reason. The reason for a company to exist is they specialize in solving a particular problem extremely well. And that alpha exists from the product design to their business operation to their deep understanding of the customer and all of that reflecting private data. And today every single company should have full control of how to turn that private intelligence into a model they can operate and power their product. If they only build on top of a blackbox API wrapper, it's really hard for them to build a durable business. So we want to give our customer the best tool to build a specialized intelligence to have full control of their own intelligence to stand on top of and to have full control of the cost for them to scale in the long run. So that's what we're doing and that's where we're going to use our new fundraising to deploy capital into to accelerate that pace.
What's the biggest bottleneck to your business? You're growing quickly but why aren't you growing faster?
That's part of the reason why raising this round is capacity. So we are— the whole industry is going through a super linear growth in terms of demand. It's because whether open or closed, the model quality passed the threshold of solving many, many problems and on top of that the tuned model quality is even better. And we as a company we need to grow significant amount of capacity of people across the board. We're hiring from researchers to engineers to marketers to sellers, top-notch. And we invite passionate people to join us on our mission of building specialized intelligence.
I saw someone ask for like we need a Costco of AI, less philosopher kings. Do you like the idea of becoming the Costco for AI?
That's an interesting analogy. I think at the end what we believe is the whole entire industry is changing from token maxing to value maxing.
Sounds like Costco to me.
That's right. Because at the end not all the tokens are equal. And we care about solving a specific task, use the most economical way to approach it. That's a durable business and there's nothing new here. In the past hundreds of years of capitalism, capitalism was designed for efficiency.
And I think the whole ecosystem is really good at that. And that's the—
Costco has, you know, other brands. They have the Kirkland brand. They've done some vertical integration. How deep does vertical integration go? How important is vertical integration to providing the lowest possible cost and winning on essentially value.
Yeah. So from our point of view, there's so many innovations that's happening on top of us. Many of those are applications doing vertical integration. And we are powering them today. Including in public we talk about Cursor because they've been training their own model for a long time. We talk about Jav, Jav have been training their legal model for a long time. There are many, many other customers across coding, co-work, all kinds of co-work verticals from legal, finance, recruiting, marketing, sales, customer support. A wide variety of verticals. They are all building all sorts of vertical solutions and they have their unique insight to build their customized model and make their business really stand out. On top of that there's also a lot of consumer-facing companies and the whole entire industry is literally going all-in on AI in production where we are helping them to transition into embracing not just embracing AI in the proper way but really integrate their offering into their—
I mean even when you see Google search overviews like that has to be extremely cheap like they don't charge for those obviously. Google is completely vertical integrated down from model training to they have custom silicon, they have their own data centers. Is that where you think it goes? Do you think you'll do custom silicon, your own data centers, have power generation contracts to like fully offer the cheapest possible product for a particular category.
So I'm humble enough to acknowledge there are tons of experts in every single layer of this AI innovation. I think Jason mentioned five layer cake. I think there's probably more than five layers if you look at it.
So every single layer has their own experts. We want to work with them. We want to work with world experts really good at doing their own job and we specialize in building the specialized intelligence platform across training, inference and we partner with all different layers to drive the best vertical solution. That's our philosophy.
That makes sense. Well, congratulations. Clearly working, Jordy. Incredible progress. Thank you so much for coming back on the show. Can't wait to talk to you again soon. We'll talk to you later. Goodbye.
Let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, or automating business workflows, Codex helps you move projects forward from start to finish. There's one more news story we got to go through really quickly. Wedding guests are now placing prop bets on everything from how long the first dance will last to whether the groom will cry during the ceremony. Call Sager and Jenny. This is a dream come true for him. Couples are using printed cards and apps to let guests predict things like who gives the longest toast, how many outfits the bride wears, or whether the first kiss lasts more than 6 seconds. The idea is to make weddings feel more interactive, especially during slower parts of the night like cocktail hour. One app called Betting on the Wedding says more than 25,000 couples have created pools on its platform which cost $49 and includes a live leaderboard. The company says revenue is growing at triple digit run rate year over year.
Real insider trading risk here, right? You might have the groom talking to some of his buddies saying—
I've got— I'm going to cry. I want you to know I'm gonna cry. Go go go go go bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet bet the house on me crying.
Maybe, but I think this is designed to be generally small prizes, $20 gift card, maybe some memento, maybe an engraved dinner plate from the wedding just to show that you were more engaged, something to remember.
Well let's ask Jason. Would Jason encourage betting on his wedding?
When are we going to get betting— when can we gamble on 37signals properties? That really—
Well my wedding was— we had people in our backyard. So, there wouldn't have been a very big use case for that app.
Yeah. Small pool. Lack of liquidity. That's a real problem.
Small number of people doesn't mean there's not a lot of volume necessarily.
That's a good point. People throw in some real time.
You get DHH there. He throws in, you know, his CGT, you know, he puts it all on the line. You never know.
Did you see that picture today?
Trying to hurt your feelings. What's going on?
Yeah. That was a little— that was a bruise. That was a little bit of a bruise. It was a bruise because way back when I used to own a Singer 911. And I was selling— this is a number of years ago before they went crazy crazy and I was trying to sell it and some guy's like I'll trade you my Carrera GT for that and I'm like eh I don't really— nah I don't really think that was a good deal and it turned out to be—
One of the great trades of all time.
That would have been a good trade. Yeah.
Have that image pulled up. That's so brutal.
I was in the Alps Thursday, Friday, Saturday and the event that I was at there was hundreds of Porsches everywhere and still when the CGTs would roll up, everyone would get quiet and just watch. Like seriously, there was one moment where there was probably at least 200 people. And it was everyone's just talking, talking. CGT pulls up. Crowd goes silent. Everyone's just in awe.
What color was it? Silver.
There was actually a bunch. Red. There was a GT Silver. GT Silver on tan is like probably my favorite that I've been seeing.
But no prediction markets around it.
Not at all. Just enjoying cars purely for the love.
You could get so many people if they're not comfortable driving. You know, we know some people that collect cars but they don't drive them. They could partake saying, oh Jord's going out for a little lap. When will he get back? I'll bet on it. You know, of course there's insider trading risk.
Will he get back? Like will he just hit the wall? Who knows?
Those things are tricky to drive. I understand.
What's your latest vehicle purchase?
I bought a 1979 Porsche 928. Which is one of my favorite cars of all time. I own two 928s. They're both old and they're not expensive, but they're awesome. And I bought a green one. It's oak green metallic, which is a rare color and has Plaid seat inserts and it's just awesome. It's just so 70s. I love it.
What is— I've never driven a 928. What's the experience like?
They're very planted. So, it's a V8. So, it's a front engine car, which is unusual proportion, but it's very stable. I mean, these were not that fast. I think they had 210 horsepower or something, the early cars. This is the first year, first and second year. So, they're not fast, but they feel great to drive. You should borrow it.
Are you a Model Y guy as well? Because that was the funniest thing about DHH's post is that he's just like, all these cars are kind of worse than the Model Y in some ways.
I do have a Model Y and it is probably the best car I've ever owned overall. I mean, it's so comfortable to drive. It's quick as hell. It handles great. We had a previous Y which I didn't think was very good, but the new Y's are fantastic. I just love it. Yeah. I mean, really, I prefer to drive that over anything to be honest.
Do you have an intuitive sense for the business logic between the lack of fast followers around that? Like in terms of just appliance vehicle, it feels like all the other manufacturers are still playing in their special— this car says something about you. It offers a particular experience. It has a convertible, but just in like the appliance basically a minivan on wheels, ultimate utility, Tesla just has had it on lock and they're like running away with the market.
They have. I mean, I guess that's what Honda and Toyota did for many, many years, right? You never really thought of those as— they were more just basic appliances. I need to get point A to point B, reliable as hell and just work, you know. So, I think Tesla kind of slid in there and basically did that with EVs in a way that everything else is more of a statement. I guess people might think a Tesla is a statement, but it really isn't. It's just like I want a great car that's incredibly quick, clear, technology advanced, affordable, full self-driving is incredible. It's just really an amazing thing. And if you haven't really been in one recently, you don't really know because they weren't that high quality four years ago. They were kind of bad. They've gotten to be very high quality now.
People complain about the panel gaps and all the interior and all sorts of stuff, but they've—
Yeah, they've sorted that out. They're incredible now. Yeah.
Where like a lot of different cars feel like they're in bubble territory. CGT. I don't know how much more it can go up. I'd be— I'm sure it'll go up more, but there's— I think there was one on Bring a Trailer. Actually probably—
