Robot Bans, Zuck Op-Ed, eBay Settles | Joe Weisenthal, Jameson Detweiler, Stanley Tang, Amjad Masad, Adam Cohen

You're watching TVPN.

Today is Wednesday, July 29th, 2026. We're live from the TVPN UltraDome, the temple of technology, the fortress of finance, the capital of capital. Let me tell you about Rank.com. Time is money. Save both. Easy to use corporate cards, bill pay, accounting, and a whole lot more. The US government has banned Chinese humanoid robots. The US is banning new imports of foreign-made humanoid robots, quadruped robots, robot dogs, you might know them by. Very controversial based on We got to get general catalysts to weigh in on this because aren't they anti-robot dog? Wasn't that the whole thing of the video?

Maybe I'm misremembering it.

That's right.

So they must be celebrating.

Some platform VCs are very pro-robot dogs.

Everyone should be pro-robot dog. Anyway, there's nothing wrong with a robot dog unless it's a geopolitical discussion, and then there's another layer to the debate. So the Federal Communications Commission, FCC, today announced some restrictions, and these apply to new versions of those products. FCC Chair Brendan Carr said the move is intended to, quote, secure America's critical supply chains. China, meanwhile, has condemned the decision as protectionism. Is that because they're protecting the small children from getting smacked in the face by the humanoid robots?

It is protectionism. I know, I know that video. I think that you're referring to.

Yeah, brutal.

Absolutely brutal. We should try to pull it up. We should try to find the humanoid, uh, kicking a child, please. Sounds wrong, but I just want to get at the truth. Yes, I just want the truth.

The move comes as Chinese companies have established a commanding lead in humanoid robot robotics. Of the roughly 15,000 humanoid robots shipped worldwide in 2025, China's Unitree and, uh, AGI Bot. AGI Bot. I'm not sure how you pronounce that. Each delivery—

okay, we have the video of the robot kicking the child.

Oh, brutal. There's another one we're spinning around and gives him sort of like a spinning back fist. But that is crazy. And that is— seeing that, you're kind of like, yeah, I'm up for some protection.

Gross negligence.

If it's going to, uh, if it's going to protect children, I think it's maybe in the best interest. No, that's not why this is happening, of course. So, uh, 5,000 units were delivered from from each Unitree and Ajiabot. By comparison, US companies including Tesla and Figure shipped only a few hundred robots each, if that. China is also estimated to control about 85% of the global humanoid robotics market and also a ton of the deep supply chain, obviously. According to the Associated Press, the restrictions are expected to add new tension to US-China relations ahead of a planned September meeting between Chinese President Xi Jinping and US President Donald Trump. The FCC argues the measures are necessary to reduce potential risks posed by connected foreign-made devices that could become part of critical American infrastructure. What's interesting here is that we're seeing a ban on a Chinese technology that is so nascent that you don't have a coalition of companies that are saying, whoa, whoa, whoa, I've deployed— my business is built on top of Unitree. I know there's geopolitical considerations, but I can't make the cars that I make without them. I can't run my business without them. And so this is going to hurt me. No one's really stepping up in the same way that they are with Chinese LLMs, which a lot of companies are like, no, I'm actually going to cut my inference costs by 80% if I can.

And you have all of the, the various cloud businesses that are already making money, making money from them.

Exactly.

Yeah, I'm very different. I saw a bunch of people that were sort of saying that they didn't like this, that it was, you know, protectionist. And, and that there must have been some American humanoid or robotics companies that were pushing, lobbying for this in DC. Maybe there were, but I personally don't want 100 million humanoid robots in this country running around that are not made by a US company or at least a Western ally.

Yeah. Where— since there's been the DJI ban, there's been various surveillance tools that are banned, but there's still so many different pieces of technology that come from all over the world. I wonder where people would like to draw the line generally on technology, like is it okay if the Wi-Fi router is made We drew the line on Huawei. Yeah, on Huawei on the 5G towers specifically, but most— you can still get a router that, you know, a Wi-Fi router for your house that comes from China. And you have to imagine that many of the actuators, the motors, the batteries, CATL, CATL, that company ships batteries for all sorts of stuff. The Google Waymo approach I think is probably the correct, like, finding the line. So Waymo is sourcing, like, almost all the parts for the new Waymos from China, but they are building them, I believe, in the United States. And all of the software is from Google or Alphabet or Waymo. And so you have to imagine that they're able to tear all these pieces down and sort of maintain their own cybersecurity posture, which is really what this is all about. But it is interesting. Do you think this actually accelerates the American humanoid market? Does this lead to up rounds from 1X and Figure because the playing field is more open? Does Tesla meaningfully pull forward their roadmap? I was—

Yeah.

I mean, these companies have already raised a lot of money. Elon's putting billions into it. Figure's managed to raise billions. I think that the first thing that we need is profitable use cases for humanoids. And so I think there's a lot of work that the industry needs to do to make these products genuinely valuable in the home, in workplaces, in factory settings, all these things. I think that's a bigger problem. I think if you— it does feel inevitable at this point just because there's so much money going into it, so many smart people dedicating all their time to it. It feels inevitable. Maybe it speeds it up a little bit. I would, it just felt like we were going to end up in a situation where they would just continue to dump humanoids onto our, you know, you could buy, you probably still can buy Unitree robot on Walmart.

I was asking Tyler about this because you used to be able to buy Unitree G1, I believe, one of their robots on Walmart, which was sort of crazy. And the price was accessible. I mean, it was still tens of thousands of dollars, but it was something that, a research lab could buy pretty easily. I wonder if there will be other ways for companies to get access to these products, at least for benchmarking, because they will be competing internationally. Like, even in a world where Tesla gets to the frontier of humanoid robots with Optimus, and then they're going up against a Unitree, for a deployment in Germany like BYD, right? BYD and Tesla compete internationally even though Tesla's Teslas can still be sold in China, but not vice versa. But internationally there needs to be a lot of like, what are the capabilities so you can actually sell against your competitor? Yeah.

Even if you're— so far everything Unitree I'm seeing on Walmart is out of stock. So I don't know if they just marked it out of stock when this went live or they had just sold out.

These would probably spike if they hit Cars and Bids or Bring a Trailer. I think you throw a Unitree G1 last edition, final edition, banned, can't buy anymore.

Who's making Bots and Bids?

Bots and Bids. That might be it. Let me tell you about Cisco. Critical infrastructure for the AI era unlocks seamless real-time experiences and new value. Cisco. Really, really raising the bar for podcast ads. Podcast ads.

Thank you, Cisco.

I was listening to Elon Musk on The Economist, he sat down with The Economist for a long-form interview. There was a clip that went sort of viral for some sharp jabs back and forth about the media and how the public perceives the media. But the much more interesting substance was about his vision—

You're describing the crash-out?

Yes, it was described as a crash-out, but that seemed like a very small section of what was actually interesting about the more than an hour they talked. One was— I mean, he talks about humanoid robots and he paints a very clear vision for the Optimus program. And it made me really hope that there's a lot of focus there because that feels like an industry where like, he obviously has a very strong philosophy around xAI and Grok. And you saw it in the SpaceX S-1 where the market for enterprise software was like $22 trillion or something like that. And so, there is some value there. It might make sense from a business perspective, but it was less inspiring than rockets, less inspiring than self-driving electric cars. Because as you go back, you think about SpaceX and you think, if he hadn't worked on that, probably wouldn't have a rocket company in America. And same thing with the electric cars. Like, yeah, GM had the EV1. There were a couple other companies that were sort of like, testing things, but we'd probably be 10 years behind if he hadn't sort of like sheer force of willed that into existence. And it feels the same for humanoid robots where if he can really go and push that extremely hard and marshal all the capital and do all the difficult things to build the supply chain, make it happen, make it work, you can get to a place where there is a national champion in America. But if Grok stopped training and there were no more xAI LLMs, like there would still be a massive competition between OpenAI, Anthropic, Meta, Google. Like there's plenty of really well-funded hyperscalers and leading labs that are going to provide that service of like ask a question and get an answer or fire off a prompt and get some code back. That the American humanoid robotics industry feels much more fragile in the sense that if you don't have this massive hammer of capital and will coming to bear, you could just wind up with it never happening, which I worry about.

Yeah, and America needs to figure out how to make billions of robots.

Mm-hmm.

Itself.

I do wonder how this is going to play into all the other discussions that are going to happen at that US visit by Chinese leader Xi Jinping. He's scheduled to meet with US President Donald Trump in September. I'm interested to see, are they going to be—

Can you see if we can possibly get him on the show?

I wonder if he—

What if Xi was just like, yeah, I'm meeting with the president, but I'm also going on a 50-podcast tour while I'm here.

Get him on. Yeah, there's a bunch of good shows he could do.

He should do— Modern Wisdom. I want to understand his mindset.

Hot Ones would be great. I'd love to see him on Hot Ones.

Joe Rogan, of course.

Rogan is a no-brainer.

That should be kind of the last show.

But Nelk would be like the obvious choice.

Yeah.

A lot of world leaders have been on there. That's true.

Controversial world leaders.

That's true. Why not Xi Jinping? Get him on there. Let me tell you about Consul. Consul builds AI agents that automate 70% of IT, HR, and finance Finance support, giving employees instant resolution for access requests and password resets. Let's go to some of the reactions. "Today we're launching Tao's humanoid cleaning service in San Francisco at $30 per hour. Access is initially invite only as we scale operations." So Alexander Koch is talking about his company Tao Robotics. He's the founder and CEO, Z Fellow, Emergent Ventures Fellow. Love that.

Fellow maxing. This fellow.

And Sean Perry from My First Million says, why did putting clothes on the robot make me like it 5x more? Yeah, just way— looks way more human if you don't see the metal. That makes a lot of sense. But are the clothes particularly for just the vibes of deploying this product into someone's home and it just makes people feel like, oh, that was a good experience? It does look like it has a Wi-Fi router on its head. Head. What are all those bars coming off the top? That's a little disturbing. It looks like Edward Scissorhands or something.

Interesting. So they are doing teleop. He says each humanoid is jointly controlled by a human operator and AI. Interesting they come out and charge $30 per hour. That feels about what you'd pay for a human. And so it can either do human-level task. And clearly, I would have expected them to come in and say it's like $5 an hour or something like that, right?

Probably not unlimited venture funding. And it's probably better to test the market with something that's actually—

Yeah, something on the higher end.

Yeah, to see if people will pay.

Anyway—

If you can scroll down, you can see it in action. It does look so funny with clothes on. Because it kind of looks like it walks like a toddler. Yeah, but it's like full size.

Yeah. Well, that does look very funny. I like this.

All right, we got one more video. It's doing some dishes.

Okay.

It gets the soap onto it.

It's washing its own hands.

Why does it need to do that? Did it get dirty? I guess, I guess if it gets dirty, it's got to wash its hands. That's pretty cool.

You don't want a robot with dirty hands.

The purple eyes do also do make it look more friendly as opposed to those two. That is way creepier. That looks like, yeah, that's, that's full Terminator mode. Anyway, Dean Ball's going back and forth with Will Manidas. Dean says, do we think the Silicon Valley leaders who were so up in arms about using Chinese LLMs last week will be similarly outraged about the US government's decision to ban import of all advanced robotics from all foreign countries. Is Manitas going to write an essay about how the US robot companies need to be nationalized in exchange for this policy favor? Will says, "You're welcome to tag me if you want a take." What's your take? He does ask him for it. And Will says, "Dean seems to not understand that the FCC is issuing a published rule under statutory authority with a waiver process and judicial review is different than a thinly veiled suggestion that regulators carry out an extralegal whisper campaign to protect the interests of his employer. The robotics companies need not be nationalized until they also ask for public infrastructure, preemption, federal backstop of their debt, and a swept market just so they can call themselves national institutions. When they do, send them my essay. And Dean Ball says, so wait, if the US straight up bans Chinese LLMs using explicit statutory authorities, e.g., supply chain risk, that wouldn't necessitate lab nationalization, but doing it through soft law would.

What?

Your instance that my analytic prediction was a thinly veiled suggestion continues to be bad faith. I do wonder if we'll just see more, you know, Figure shows up in those secondary demand tables. I wonder if this is bullish for that. I don't know. 1X just seems to be on a good path. Sunday Robotics just seems to be on a good path. We've talked to a lot of other companies where the regulatory and competitive landscape has always— it hasn't seemed like it's a driver of financings, right? It's not like, oh, you're waiting for approval or you're waiting for some rule to break. It's like, yeah, at some point in the future, America is going to want an indigenous robotic supply chain. It'll probably be a company like—

Compelling teams have raised anywhere from $100 million to billions of dollars. And now they need to effectively deploy that and get commercial use cases starting to ramp. And then I expect them to be able to raise an order of magnitude more capital each, right? So I don't, again, I don't think capital is really a constraint here, but I think they've raised their sort of like, they've raised their like 5 rounds, let's say.

Yeah, yeah, yeah. Well, let me tell you about Railway. Railway is the all-in-one intelligent cloud provider. Use your favorite agent to deploy wireless servers, database, And more while Railway automatically takes care of scaling, monitoring, and security. Mark Zuckerberg is in the Wall Street Journal paywall. You got to pay to read what Mark has to say. But he penned an op-ed and he says the US should accelerate AI development, not restrict it. He says the AI future is for everyone. There's some interesting quotes in here. He's taking shots. He's taking shots at doom, fear-based marketing. He says, it is surprising that the discourse from many of those who are developing artificial intelligence is so filled with doom. I don't understand why anyone who believes that AI will eliminate most jobs and much of humanity's relevance would rush to build that future. The only— the notion that AI is so dangerous that only— that the only safe path is an extreme 'Absolute concentration of power seems dangerous. Historically, hoping that an absolute power will benevolently provide for humanity, if sufficiently enlightened—' Is he talking about machines of love and grace? I think so, something like that. 'Hasn't led to safer positive outcomes. Humanity has witnessed many transformative advances. Each time there is fear that people will be left behind, but each time humanity has come out with more people sharing greater prosperity, health, and freedom. We believe this will be true with AI as well, and the abundance of the future can be shared by everyone. He's abundance-pilled. And so I think it's good to sort of reset, you know, where he is. He's been, he's been, he's talked about AI a bunch, but it's never been like right on the cusp, but he's just like throwing his hat firmly in this camp of, He's obviously playing in AI but has a particularly positive vision, and he's sort of throwing down the gauntlet. If you think AI is so bad, then stop working at it and let me work on it. Let me cook. I don't know. But not too much like chatter around this. It did stand in opposition. He's saying the United States should accelerate AI development. At the same time, there's this open letter from a bunch of lab employees saying like, hey, maybe we should do a mutual pause. Maybe we should all slow down. And I think even some Meta researchers signed that as well. But there's like a wide group from OpenAI, Anthropic, a bunch of other firms, Google, that all signed. John Palmer, of course, ever the TVPN promoter, says TVPN must publish an open letter against distillation after their model was distilled by every competitor in the industry. And I said, I'm advocating for a "Complete and total podcast ban. Multi-year federal approval process for new RSS feeds to get issued." Uh, he says, "Better hurry up and write that open letter before Chinese technology network launches." I think a Chinese TBP would actually be electric. I think that'd be amazing.

I would love to watch.

I would love that. Uh, you know, just wildly different. They just don't do the going direct thing over there the same way. You know, there was like the Deepseek investor meeting 4 hours that was like leaked. That's not the same as actually sitting down with someone that has a particular point of view.

Sitting down with a stack of 200 posts. Yeah. And just chit-chatting.

I don't know. Maybe it does exist. But, uh, when Dwarkesh went to China, it was such a funny observation. He had a bunch of observations, but the one that really stuck out to me was he was like, "I asked everyone, and it is clear China does not have a Joe Rogan." Like they don't have someone who is like a larger than life yapper political commentator, just like a guy who talks about everything, including politics. And there's political reasons for that, but it's also maybe cultural. And it's just interesting because the demand for that would be electric. I feel like you have to imagine that like Joe Rogan clips make their way around the world and are sort of like a barometer for where the American psyche is at. In a certain way, like a certain group of Americans feel that way. There's this other post that's sort of interesting to see where the Anthropic back and forth on the open letter goes. Obviously Anthropic clarified their stance on open source, not advocating for a complete and total ban, but they're very worried about distillation. They're very worried about cybersecurity, and they're very worried about authoritarian uses of AI. All good points. Chetan over at Benchmark says Anthropic very publicly claiming there needs to be a crackdown on distillation is puzzling. It's a $1 trillion company with incredible resources, and the distillation attacks at the scale they describe would seem easy to detect. I don't know how easy they are to detect. I think the people, you know, harness up a whole bunch of different VPNs and stuff, but at the same time, it's like it's an AI company. Put AI on the job. Maybe you should be able to detect it. So we'll take his point.

He's asking AI to be able to find a needle in the biggest haystack of all time.

Yeah, the world's largest haystack. The final benchmark. Uh, yeah, maybe DistillBench is, uh, Distillation Attack Bench is the one that needs to be hillclimbed next. Uh, but he says the only cost would be giving up the associated API revenue. So you should have a very aggressive a very aggressive internal policy around if you're suspected of a distillation attack, you're kicked off the API. Of course, in a very short amount of time, you can fire off a lot of prompts and get a lot of tokens. And you can do that a lot around across a lot of different accounts. What do you think about this?

Yeah, I mean, detecting distillation seems like a very hard problem. I think if it was trivial, they would have just done it. Yeah. And also seems like it's not that big of a like, percentage of their revenue. Like, I'd be surprised if, like, distillation attacks accounted for a large amount of, like—

Yeah. I don't think it's meaningful.

Also, so it does feel like it's hard because a lot of it doesn't actually go through the direct API necessarily. Like, it goes through a company that is vending a service, and then someone is prompting that service. And so you're further obfuscating it. Like, there are a number of different sort of wrappers that effectively wrap various Frontier APIs. And if you're getting hit in a bunch of different— you're blending all of that together, you're looking at open source stuff. There's also just people that will— like, this just happened with— was it Matt Schumer? He basically uploaded to GitHub a game that he made with Opus 5 along with the prompt. And so, you can sort of distill on that. Like, it's only one data point or, you know, a small bag of tokens. But there's a lot of people that will share a prompt and the actual output. And so you can train on that and there's no way to detect that you found that. What's interesting here is that the knock-on effect of Chetan's point is that you, like, if Anthropic takes Chetan's advice, you wind up giving out the model less frequently and being more closed about it, which is something 3 months ago, everyone in the ecosystem was like, this is annoying. Like, I'm a small startup. I can't get into the Mythos tier. I can't get access to Project Glasswing. I'm not an enterprise-level customer. But I still— so I just don't get frontier intelligence, and I just get steamrolled by Microsoft, who has access to it, or Apple, who has access to it, or whatever. And so this— The whole idea of like, if an amazing intelligence, an amazing model is made, should it be widely available? That's actually at odds with distillation because you have to hunt those bad actors down and you play this whack-a-mole. So I don't know, it is tricky. We'll see where it goes. We'll see where it goes. Anyway, let me tell you about public.com, investing for those that take it seriously. They got stocks, options, bonds, crypto, treasuries, and more with great customer service.

Apparently over on Facebook Marketplace, people are adding a sort of gold aura to their listings. What is this? Would this make you more likely to buy, John?

Oh, it does look like a rare drop.

Yes, for only $25, picking up this table with that level of aura, certainly tempting.

It does make me like it more. I think it actually works. I've seen people do this where they add they'll add like a woman to it as like the salesperson basically with AI. And that draws more attention on Facebook Marketplace. I don't know if I've ever bought anything on Facebook Marketplace, but maybe, maybe we should try and have Tyler do like a trade, trade a paperclip into a car challenge. Have you ever seen these?

Ryan Trahan and Tyler would be exceptional at that kind of thing.

You think you could trade up from something on your desk?

I think this would definitely help it. Yeah.

So basically you have to look for the most unoptimized listings that are undervalued, and then once you buy them, give them the ultimate glow up with AI images, put them back up for a higher price, increase your margin, start printing. This is a very Gary V-esque hustle for sure. Asking a guy, okay, where do you wanna go next? Let me tell you.

Wise and Tall joining in just one, just one minute. If we have time, we should cover this story coming out of eBay. Let's do it. Pretty incredible story. eBay pays couple $56 million for its role in cockroach harassment campaign.

That is a crazy headline.

The couple received threatening deliveries, including funeral wreaths, after criticizing the e-commerce retailer in a blog. —so this couple, David and Ina Steiner, were harassed, it sounds like, for years for criticizing eBay. Who knew that criticizing eBay could get you in this much trouble? Of course, it's eBay that ended up getting in trouble. But—

So they had an online e-commerce blog called E-Commerce Bytes, and they criticized eBay., on that. And then back in 2019, so this has been going on for like 7 years. There's a whole new executive team in, in, in place now, but there were actually criminal charges in 2020. 7 employees of the company were sentenced to prison terms ranging between 1 and 5 years.

And so what were they doing? Uh, eBay leadership apparently mailed a series of threatening deliveries to the Steiners' home in 2019. That's crazy. Including a box of live cockroaches, a bloody pig's head mask. I don't even know what that is. Head mask? Is this a real pig head that's been turned into a mask, or is it, uh, hopefully just a, uh, sort of a Hollywood remake of that? Uh, and a funeral wreath. The Steiners also accused executives of online harassment, including online posts inviting strangers to their house for Sachs. Also absolutely insane. Who, like, in what, in what world is, is a company, is a public company like eBay just getting some criticism, which all companies do? Yeah, yeah, like, I'm sorry, it takes it this personally that you're gonna try to ruin the lives of two people that are already in your, in your industry?

The impact of e-commerce bites blogs showing up in your quarterly reports? Like, was it actually—

eBay pleaded guilty in 2023 in a criminal case over harassment and paid a $3 million penalty following criminal charges in 2020. 7 employees of the companies were sentenced to prison terms ranging from 1 to 5 years. It's absolutely brutal. According to court documents, eBay executives grew increasingly upset by the Steiners' critical coverage of the company in early 2019, leading then-executive Devon Wenig to test— to text another executive that it was time to crush the Steiners and take her down, referring to, uh, Ina. That summer, harassment against the Steiners escalated from threatening messages to delivering items to their home, which also included a book about surviving spousal bereavement.

It's like sort of a death threat, I guess. Yeah, right.

What's, uh, also sending, uh adult magazines to neighbors' homes in the husband's name. So just like completely trying to destroy this couple's life and marriage. eBay employees also traveled to Massachusetts to vandalize the Steiners' home and stalk them around their hometown. All the employees involved in the harassment were fired in September 2019, eBay said. Uh, the lawsuit was actually started in 2021. Under the settlement— so we have a, we have a good ending here, I think it's a fair ending— under the settlement, the couple will receive $48 million in compensation from eBay and former employees. eBay will also pay $6 million to charitable organizations, and Wenig will pay $1 million to a free speech charity in Ina Steiner's name.

This is like the ultimate Streisand effect. I never would have heard about eCommerceBytes or their criticism of eBay if they hadn't done all this.

What a mess. eCommerceBytes.

I'm, I'm looking up exactly what they said because eBay is, in my opinion, just the most set-it-and-forget-it company that's just had the same service for 20 years, basically. Well, but now we're going to get to the bottom of it.

Ines Steiner published, the news that this settled on her blog. So EcommerceBytes has been giving sellers a voice since 1999. So I do think that it, it appears just from looking at it very briefly that EcommerceBytes clearly is an important voice in the, like, eBay seller world, but still absolutely insane way to go about doing business. And I'm glad that eBay cleaned up their act in 2019.

In 2019, eCommerceBytes posted an article, said, did you know that eBay built a lavish New York City pub-style lounge? Because they had constructed an elaborate replica of a Manhattan pub at an eBay campus. And the framing was basically, this appears to be a no-expense-spared executive amenity. Maybe they were going to do a podcast. You never know. Anyway, we have Joe Weisenthal from Bloomberg Odd Lot. You know him, you love him. He's back on the show. Welcome to the TV Game Ultra Gun Battle.

How you doing? Joe.

It's been too long. I miss you guys.

I miss you too. Okay, so I have a question for you. You've had a podcast for a long time. Have you ever gotten like a negative review or a negative comment? Ever? No, never. Never. Now be honest, be honest. Maybe you've gotten one. Have you ever thought about sending them a box of live cockroaches, maybe vandalizing their car?

There are actively people mad at Joe right now on Twitter about his latest take.

Hopefully she's not watching right now, but my co-host Tracy would if one of us would do it. First of all, she is the one who reads the comments. I don't think she would, but I think if like you're to float this idea, like she might at least like spare a few mental cycles.

Like, I've thought about doing that. I ran it down, but I know the whole, the backstory with eBay. I try, I try to be the voice.

It's a cautionary tale.

You guys don't want to have to, you know, down the line pay a $50, $60 million settlement. Yeah, definitely not. No. Sending a horse head to a former listener for leaving a bad review. We have—

that is not on the roadmap.

No, not at all.

I'm trying to think if I can— I don't know if you caught the story right before this, but basically eBay had to pay a couple $56 million because like a bunch of executives were harassing this couple for just like criticizing the company in 2019. So like, no one is— no one is even really thinking about eBay. In 2019, like, it's just not like— it's a big company. It's a company that everyone, a lot of people have used and a lot of people love. But no one's like really— it's not really that top of mind. And 7 executives are like becoming criminals.

I will die for this company.

You know, in 2018, they might have said, oh, that would never be me. You know, maybe who among us, if you're late at night, you start going down a path, you have to exhibit really good cognitive hygiene, because who among us knows the path to psychosis online? Many such cases. And maybe we all need a little bit of humility about what the set of conditions that could put us in that path.

Yeah. Sometimes one prompt is all it takes to lose someone.

Sometimes one prompt. I took a week of vacation last week and I was like, I'm not going to post anything for a week. And I did. I actually did. But I felt it all like bubbling out like Friday evening when I sort of mentally was like, oh, vacation's over, it's the weekend now. French toast. And I was like, it just like all, I was like incontrollably. So it's like we all have to be mindful and self-aware about what we're capable of. 'Cause some, maybe you don't really know like when it comes to internet vendettas, what could happen.

Do you think social media is addictive? 'Cause I can't, it's so hard to tell when your job is social media. Like I am the screen time, I have 3 hours of livestream time every day. Yeah, but it's like how I buy food. So it's like, am I addicted to my job? Like, I don't know.

Well, I've thought about the funny— you mentioned food because this is how I justify it, right? I've probably said it before. This is how I justify always being online. It's like, oh, it's very important for me to know what people are talking about. It's how I hone my skills. But it is like, you know, it's funny, I was thinking like food specifically because with other addictions you just go cold turkey. Right? You're just like, I'm not, you know. And with food you can't. So you have to like, it's really hard to like, so even with me and social media, it's like, it's hard for me. Like cold turkey is not that hard. I'll just take a week off. Posting all the time is not that hard. The middle ground is when it gets really hard because there's just so many things to respond to. So many things waiting for my take. So many questions that only I need to respond to.

Also, yeah, there's takes that will never exist. There's takes that will never exist, right? Yeah.

The takes that have been lost to history when you were on vacation.

Yeah.

We'll never get those things. That's right.

I do feel like— That's right, and you think of— Yeah, I do feel like one benefit that you have and I have felt is when you have multiple, maybe dopamine reward paths, but like outputs, it is nice to be able to retreat because you can do a podcast, you can also write an essay, you can also rip off a post. And so if you only have one of those, then you can sort of get lost in the sauce of like, oh, it's been rough since I had a great guest, or it's been rough since I had a great viral banger post. But if you can sort of move around and be like, okay, Okay, I'm going to take a break from the knife fight at the dive bar and put my thoughts together and do an essay. No, I agree.

It is very important to have multiple outlets for man's desire to post and put things out into the world. Yes. The deepest desire, it's like in the hierarchy, is like to be someone we are listened to, right? You could see it in all of like the billionaires who like cannot resist the impulse to get online and start posting. It's clear that money, yes, it's all great. Power is great. But what they really crave is to just be heard by someone out in the world.

And so they start posting. The death of Twitter acts fully over now, given that— oh yeah, Jensen Huang joined. Mark Zuckerberg had to join. Like, it truly pulled everyone over. Yeah. Kicking and screaming. Because I imagine Mark Zuckerberg would prefer to promote his Wall Street Journal op-eds just on his own platforms. But that's not where that community is. You got to go.

No, it's true. Like with AI especially specifically, which is that it's X, right? Like every time someone leaves one of the big companies, they have to post a big thread or, you know, a really long tweet or whatever. And it's only X. I mean, they could post it on Instagram, they could post it on LinkedIn, they could post it. They don't, though. I mean, they might. I don't know. But they posted. That is where the news happens. And you can feel— like, I think the thing that was really striking last week with the discourse, as one says, about the open letter, et cetera, it germinated on Twitter specifically. Why did the White House— why was this suddenly in that week a thing that the White House in some way needed to talk about Chinese open models? It was because of how much it was bubbling on social media, on Twitter specifically. I actually think the entire thing more or less originated from one tweet that said that Kimi3 was really good at front-end design work. Yeah, that was it. It was that one post. Because if you look around, like, that actually, that crystallized it in everyone's mind. That's like, oh, here it is, maybe as good as the most, uh, advanced frontier models. No, there was one category of one overall thing, and that was the thing everyone locked into. And from there, and I believe that that was the moment straight through to, okay, now the White House and all these companies need to make comments about open models, can be traced to that specific all online, all on Twitter from that.

America, front end built this country. If we don't have front end, we don't have anything. This country needs front end development specifically. No, no, I mean, I agree with you. When I was looking at the benchmarks, I was like, the one that I'm actually interested in is the cybersecurity benchmarks. Because that's where we could actually have this meaningful discussion of, okay, you just actually democratized hacking into things. This is going to get kind of crazy. And you're going to have like script kiddies out there that are, you know, just hacking into the random flower shop on the corner and just making it a nightmare for them.

And we couldn't really get there. I don't know.

It hasn't been fully—

I got to post this, but someone was sharing this morning their trying to figure out how much you need to spend to have your own setup at home to be able to run Kimi efficiently. It was like $100,000, right? Just to get to the point where you'd be able to— Yeah.

And it's a lot of energy. —in a sovereign way.

A sovereign hacker would need to have $100 grand, which hopefully prevents the script kiddies.

It might be the new Lambo, though. You get the influencer, the LARPer. They got the NVL72 in the corner. You're like, yeah, what's up? I got the— I got the sovereign AI going. I don't know. How else did you process the, the open letter? Were you surprised by the order that companies signed? Ben Thompson had this interesting take that it's two of the last companies to sign. I don't think Apple's signed yet. You'd think that they benefit from open source since they haven't been playing that game. Amazon was also a little bit late, but Amazon and Apple both have physical products like warehouses and they sell physical things. And so they're maybe a little bit less nervous about AI disruption because they have a physical aspect. I don't know what you thought about the sequence of events.

Yeah, I mean, I think there was like two things that really struck out to me, which is that the first wave of signatories were either companies that are behind the frontier, Or Nvidia, which does not want to be left with only one or two clients. Like, this is the thing. So they're like, oh, like, this is— let's start with Nvidia. I think the nightmare scenario for Nvidia, or at least a nightmare scenario for Nvidia, is that it basically has the two— one or two labs break so far from the pack that it's just, you know, it's almost monopsony conditions. Sure. And then those clients can either— and to be—

and just to add to that, Yeah, like both the, the two front runners are already making their own chips. They're already working. Yeah. Yeah. You're the chip companies, right? So they're already like totally actively trying to reduce their dependence on— Yeah, absolutely.

So, so first they might have better pricing. They're like, no, we're just not going to pay that margin that you expect. Or they're already both diversifying, etc. So that is That is the first thing. And then the other thing that seems very obvious to me is that all the first round of signatories was objectively falling behind the frontier. And so when you're in software categories, you have, you know, either companies that are trying to build a model and they're behind. So Microsoft, at least on just, we're talking LLMs, clearly behind. Meta, clearly behind, right? And then you get the, and then you have the company, I don't know, the ServiceNow of the world who look like they could be in a position where they might have to pay a tax to the most advanced models.

Yeah.

And so like, and then you get the late signatories like an OpenAI or an Alphabet, etc., that are close but probably not at Fable 5, like max, like look at the whole thing. And then they're the last ones to join. And so like, I guess I had, I was like, Look, I love all that stuff about— I love freedom. Like, I love freedom and I love tinkering. I love tinkering and I want to, like, do all that stuff, etc. But when I think, like, what really, like, motivates these companies is, like, you know, I don't put a ton of stock on, oh, it's really important to diffuse AI into our Main Street businesses and hospitals and hackers. Sure, great. I love that stuff. Lovely. But, like, you know, I think there is, like, You know, I think the thing that would really— like, we're really not psychologically or sort of as an economy really equipped to think about what if one or two of these companies just runs away with it? What if Anthropic just runs away with it? What if, like, we sort of like, oh yeah, everyone's sort of in the game. And I think for most of us, I'm not 100% sure of this because I'm not the most sophisticated, like, person, technical person at all. But I have a feeling like most of us can't really get a good intuitive feel for like, what is the difference between a GPT-5.6, a Fable-5, or even the most advanced Gemini model, or a Kyma, or a DeepSeq-4, or whatever. Like, we're probably like not really able to personally grasp what the difference is at the extremes. But like, what if we, what if in a year from now there's one company one or two that are just so like clearly like there is no longer any debate. These other ones just are not in the game. That is a very weird, I would say, very deeply uncomfortable situation we would find ourselves. I think none of us would like that. I think most of us feel comfortable in a world in which you have a handful of companies roughly jostling around the, quote, frontier. It's like, okay, this is like comfortable for us. But if we weren't in that situation, like, I think I think we would be rightfully, all of us would be quite anxious.

So was Lina Khan right?

Well, so this is the thing. I think that like, and I think like the Mark Zuckerberg op-ed in the Wall Street Journal, I think further hammers this point. And I actually think there's a sense in which I find the Mark Zuckerberg op-ed to be a little bit maybe more interesting than the open letter because the open letter is sort of this plausible deniability. We're just, we're advocating for a principle here. That principle is open rights.

The Zuckerberg op-ed— and it's also like, and it's more their, uh, it's, it's just, again, I think part of what you appreciated is like using the open letter format to just advocate for your own commercial interests. It's like almost like stolen valor, you know? Sure.

But like, so like the, the open letter, like the Zuckerberg op-ed in the Wall Street Journal, I thought actually took it bit further. He's like, we cannot have superintelligence controlled by one company. So that was not merely saying, let's let the market decide, let's keep like light-touch regulation. He's like, no, this would be existentially bad. I think most people would actually agree with if one company were to be the monopoly provider of, quote, superintelligence, unquote.

So then you think, okay, which is funny too, because he's like, it's very funny. Well, he's probably top 3 largest customers of Anthropic. Oh, right, right. So, so he's the one fuel, you know, fueling all of this. And then, and then, yeah, right.

2 months later, he's like, okay, actually, like, to me, that like, if that is a very telling fact when you think about like who are the largest customers, the competitors. So it's like That marginal difference, like I said, like most of us probably can't really tell a big difference between like advanced models. But when you have like a really big, very high-skilled, very like sophisticated model maker and they're a huge customer of a major competitor that's better than them, like people are going to the frontier. And this is something that like surprised me. Like I thought maybe like there would be a bit more of a market for like sub-frontier intelligence or like, quote, cheaper intelligence. But look how fast Anthropic turned out Opus 5. And I, you know, I haven't really used it that much, but it could be like that in a market where there's— if there is one winner, that winner could emerge all up and down the price curve. And so you have like your Gemini Flash level frontier.

Yeah. So, yeah, I mean, that's what— that's what Sam was on Invest Like The Best yesterday, and he's talking about how you, as a company that sells intelligence, you want to have the best model at every point of the cost curve. Like, it's actually going to be important to hit all of those different moments.

Yeah. And it seems like— and look, things change by the week. So I'll come back out in a week and I'll say something else. But it seems like a while— like, I started you know, like I was just sort of poking around. It's like, oh, Gemini Flash, this is really fun for like high-batch, low-intelligence things, right? Like you could do 30,000 queries in 2 hours and it cost you $20 or so. I was like, oh, that's amazing. Could be very good for my life as a golden retriever.

High-batch, low-intelligence.

Yeah, that's right. Doing a lot of stuff. But probably, or it seems at least plausible that the company that has the most advanced model will probably also win for the sort of like low or could at least be in a good position to also win the high batch, low intelligence.

Yeah, but I would say, but I would say there has been a lot of, there has been a lot of pushback around the RSI narrative in the last 24, 48 hours because Opus 5 is actually not getting great reviews from the community. I think initially they felt like people thought it was like a great in-between between like a Fable and a 5.6 SOL. And then now I've seen a lot more negativity towards it.

This to me, again, goes back to our original point of sort of like the incredible importance of Twitter in all of these discussions. Because if you think about like each time a model comes out, okay, there's the benchmarks and it's like they, but then it's like, You'll just see like someone will be like, there was some tweet when GLM 5.2 came out that is like, some guy was like, I've been using GLM 5.2 for the last 8 hours. I've never seen an open model like this. Got like 2,000 retweets. It's like 8 hours? What are you doing? You're like, what are we doing here? Like trying to, it's like, oh, like I do, I got a sort of weird vibe from Opus 5 or all these things. It's like, I don't know. I actually think this is probably a good reason for all of us to maybe log off a little bit more and to actually just read through benchmark reports rather than like one tweet.

But also, on the flip side of like the benchmarks, like it actually makes me more strong on Mark Zuckerberg's ability to produce a thoughtful opinion because he has— he spent $1 billion on Frontier Intelligence in a month deployed across 10,000 people and can look at like the needle, how the needle moved across different models and, and actually look at the, like the efficiency, like the ROI is at the end of the day what matters in a business context. Yes, the like the vibe check matters. Yes, the benchmarks matter. But at the end of the day, in terms of moving the needle from $10 billion to $100 billion in revenue to $1 trillion in revenue, a lot of that's going to be driven by our companies actually like getting the bill and being like, yep, I'm happy to pay because it was worth it.

This is something I noticed a while ago, which is that you read through like a lot of these like popular benchmarks that come out. And the thing that has surprised me is that up until like recently, most of them actually don't talk about the dollar cost of completing a task. So it'll be like, can the model solve these 100 problems in engineering or whatever. It's like, oh, this one got 85 and this one cost— this one got 75, right? But what we're really going to need is to see like, what was the— okay, how many dollars do I have to spend to get 90% of these questions solved? And we're just starting to see that. I have a theory, but I don't really know if it's right, that probably a lot of these benchmark organizations started from like a safetyist mindset, right? Like they're worried about like, Oh, like, um, bio, bioengineering risk or how can you do it binary? Can you do it at any cost? Right? Because you don't really care about your token cost if you're like trying to exert a lot of damage, you just do it. So I think that for like the bigger questions, like hopefully we'll see more like what was the dollar cost of achieving X at this model? And then maybe we'll like really, those are the, for, I think for our conversations, Those would be a lot more interesting than some of the sort of legacy benchmarks.

How much do you think the open source— it feels like we're in this moment of like this catch-up. There was the last 6 months pre-GLM 5.2, it felt like open source was sort of falling behind as the Frontier Labs released more models and more crazy demos. And there were more stories about Mythos and solving math problems and all these crazy conjectures and stuff. Like the labs were sort of training trading back and forth with like new capability, new capability, new benchmark, crazy thing. And open source is kind of quiet. And then we got GLM 5.2, KIMI K3, and all of a sudden it was like, oh, okay, like the open source market is catching up for sure. Still on the same trajectory, maybe still behind, but like, you know, like on track. How much is that moving the market, do you think? How much are we in a replay of the DeepSeq moment, the Jevons paradox debate, because it feels like there's this sell-off. But I don't know, like, it feels weird to me that people would be processing it the exact same and making the same mistake that they made. So is this time different? Or how are people talking about it in New York?

Yeah, as people who just care about the lines going up. It is, I think, like— no, for real though, I mean, I do think, like, it's plausible that this on some level, this is just another, quote, deep-seek moment in the sense like you get a little bit freaked out. I agree with you. Like, it's weird to think that like you would do that same playbook twice. I think like another factor is just generally setting aside like maybe the Kimmy and GLM stuff is a little bit— maybe it's a little overrated in terms of a story. I think people are looking at a situation in which they just do not foresee the end of these free cash flow losses, right? There is just like— we're nowhere close. It's like, okay, when can you— when is the off-ramp for this spending? Yeah, that story isn't totally satisfying because if that were— if it were just about, oh, they're spending a gazillion dollars on capital expenditures, you would expect the Korean chip stocks to do well. I mean, the Korean Chip stocks have been absolutely getting crushed over the last 3 weeks. Yeah. So like, there's something more to the story here. Like, I don't— like, it could just be that a lot of manic— it could be two separate things. It could be there were a bunch of manic traders in Korea who were buying these double and triple X leveraged ETFs that just trade SK Hynix. And then that had to eventually flush out because there's, you know, there's so much volatility baked into that that you get a little selling. And it ripples. And then in the US, it could be that investors are looking at it like, I do not have any confidence that I could name the date at which the Alphabets of this world will return to positive free cash flow. There is no reason to think that that off-ramp is coming when they could say like, oh, we're good, we have enough compute. So maybe there's a couple of things going on that I think are perhaps a little bit separate from the pure like Oh, we're freaked out about margin compression from Chinese open models.

Yeah, it's very odd because the industry revenues are going up like, and the revenues are very, very diffuse. Everything from an ad-supported use to a $10 subscription to a $200 subscription to a hyperscaler spending $1 billion in a month on tokens. Like the demand is there in that it is this economy. It's not this total circular. There's just Yeah, like it's, it's like NFTs or something like that. It's not some dislocation. But I agree with you that as you play out the CapEx requirements, we've now drawn down all the hyperscaler cash flow. We're issuing more debt and equity. And at a certain point you're like, okay, if there's another 10x increase, if there's another 100x increase, like there's sort of not enough money in the world for this. Like, and so, no, you almost need a pause on CapEx more than you need a pause on like the capability development for— because I do think that if you just, if you just spent the CapEx that you're spending now, paused and then just did the diffusion work to get these models into all the different businesses and use cases, like the economic upside and the ROI would be there. But yeah, not when you're going to spend 10 times next on CapEx.

Yeah, I think my framework is like you have the whole world betting on a meaningful capability, like continued capability increases, which I think is like okay to lean on because we've just been seeing that quarter over quarter. It is real. And then you also have some like hoarding, which is like, not only are the capabilities going to get much more, much better, I'm also going to basically figure out a model that's going to be so good, it's going to just like hoover up a bunch of like new demand and I'll be able to ramp revenue and justify all of this CapEx. And if I— but if I can't do that, someone else will. I'll be able to offload it, right? But the concerning thing, right? The concerning thing right now is like the incremental big buyers of compute, you effectively have two players there and And then at the same time, you have the whole industry being like, well, like praying for like an oligopoly as opposed to a duopoly or a monopoly, right? Yeah. Like, very interesting scenario to have the biggest companies in the world so scared of a monopoly that they're effectively praying for an oligopoly. It's like, is my—

it's my view on it. Yeah, it's really wild. And, you know, I don't know if you guys saw today the recent blog post from Dwarkesh. Dwarkesh had a great thing about expectations that compute could get a lot more expensive, which, you know, one of the sort of stylized facts of the last few years has, of course, been the extraordinary drop in token prices, right? Like if you go back to like GPT-3 token prices, aren't they down like 99.9%? But like, you know, so we think, okay, there is some iron law that token prices computers will get cheaper over time, but maybe not. I mean, we thought there was an iron law that memory prices would drop over time, right? Or that any sort of hardware would drop over time. And what the story of the last 2 years is, that turns out that is capable of reversing. So, you know, it's not impossible to imagine that compute prices— I think it's not impossible to imagine that compute prices get more expensive and get significantly more expensive than Of course, he posited the possibility maybe there is no, almost no thing that we do economically currently with AI is economical where compute prices are going, which is hard to imagine. And then the thing that I take away from it is, but that doesn't mean you can stop spending on it because you have to worry about hardening yourself for cybersecurity. Yeah. Because it's like, all right, you know what? There's like, this is not economical for me anymore. I want to tap out of the whole AI risk. Well, then your competitor develops a stronger model and then you're done. You're out. You're, you're really done. So like, this is the scary thing, which is— and I think even just in the sort of like hyperscalers is like, what if like free— the situation is you're bleeding free cash flow, but you also don't have the option of saying we want to opt out of the game. Yeah. Which I think is a source of anxiety for investors.

It is. It is like it's so weird that we went back and back and forth on like the doom narrative and no matter where you sat on that, it was either like perfect utopia, everything's amazing, or like Terminators kill everyone. And in fact, like we are very much in this messy middle of like a lot of annoying things. Like the cybersecurity equilibrium is, I mean, maybe there's like this doom scenario, but most likely it's just like costs go up for everyone loosely because you have to harden yourself more against AI. Yeah, it's great to fix bugs, but there's going to be more bugs introduced. And it's just annoying. It's annoying. And there's a lot of scenarios where it's like, OK, there's more cost on AI slop. And then you have to detect the slop. And then you're annoyed. And we have a friend on our team who always laments when AI does some amazing math thing. He's like, no one cares. This is why you need to return to the curing cancer, the doing something useful actually helping you with some problem that people do experience. But a lot of times it's just these abstract things.

Going back to the Zuck op-ed, he's in the position where he doesn't have a, he didn't do a bunch of podcasts about AI and AI doom in like 2019 when no one cared about AI. Whereas a lot of the other key players were doing podcasts and they were talking about the risks and all these things. So he can kind of come out and just say like, like it's going to be awesome.

We just got to make sure there's fair competition.

But the pushback there is that like the average— he's not the best messenger for this because the average person believes that social media, while it is great in some ways, has damaged birth, you know, hurt birth rates, right? Like damaged teen mental health, right? Is just like draining the attention, right? All these things. So like he's maybe the best messenger that we got, but he's not the best.

Yeah. So this is the issue that arises when I think about this is that like, okay, you have two executives. One is like, it's all going to be okay. And one of them, it's we're going to have doom, right? So and then it's like, well, which one of them has been paying attention to taking seriously AI for the last 10 years? And which one of them 3 years ago was talking about the metaverse? Well, it's like, I'm gonna trust the one who has been taking AI seriously for the last 10 years. And that unfortunately that happens to be the Doom guy, or that seems how it is. That unfortunately, that like, if you just go by these heuristics we have of like, well, who was like, there have been people who are like very right about AI who are very nervous. And this is like what makes me, I don't have any ability to adjudicate these questions. Like, I'm just like, I'm just reading stuff on the internet. I'm like, you know, I'm not, so, but unfortunately, A lot of the people, a lot of the doom people have had much better intuitions about just the broader trajectory of the technology than the people who are very optimistic, who 3 years ago or 4 years ago were talking about NFTs.

Well, yeah, and there's some line, I think it's from Demis, where he's saying like he met with Zuck when they were selling DeepMind. Yeah, it's a great, it's a great part. You know, viewed like the metaverse as like, or VR as like equal, equally interesting to AI. And that was part of what made him decide to not go to Meta.

And yeah, has Dennis never watched a movie in VR? It's electric. It's the best experience. No, I just— because you got to— it's amazing. Apple Vision Pro. It's fantastic. But there was a doom narrative around crypto and there was a doom narrative around metaverse. The crypto one was that every criminal is going to be able transact anonymously and there's going to be a huge increase in drug use.

Well, the doom scenario was actually that the dollar would just completely collapse. Yes, that too. But that was something that people were cheering for.

They were excited about that. Yeah. And then with the metaverse, it was like wireheading. So everyone's going to be in the metaverse and we're going to lose any and all interaction in the real world. And of course, Mark Zuckerberg never addressed that. He was never on podcast being like, yes, it is. There is a risk that it goes into the matrix and we're all in these virtual worlds and like, I'm not gonna do that. Apple sort of like forced his hand with that, with the crazy AR and like the screen on the outside. And Apple's messaging was like, we don't wanna be an isolating device. We wanna be a mixed reality device. And they were sort of being like the anti-Doom, but Zuck never really went full like Doomer or even addressed it. And then of course we just like sort of got the good ending because it just like kind of flatlined and never really went anywhere. I don't know, I don't even know if that's a good thing.

It's like, no, no, totally. And it's interesting because when you read like that Zuckerberg op-ed in the Wall Street Journal, it reads like, you know, so you mentioned like everyone is super cynical about social media, right? You mentioned the birth rate. So everyone's got so like, so just feels so like kind of gross about the whole thing now. But, you know, in like 2010, 2011, people were like really excited. It's like, we're going to connect This is going to be great for democracy. Citizens are going to be able to collaborate and stuff like that. And so when I read the Zuckerberg op-ed, I was like, this is a kind of optimism that I haven't read from anyone in over 10 years about where digital technology is going. Unfortunately, people like— very few people are as optimistic about social media as they were in say, the beginning of the 2010s. But it did hearken back to that this is going to be very democratizing, etc. And I hope it is.

Well, I, I expect, like, personally, social media, I do think, has made my life, my life better. Like, net— like, if you net it out, it's like, yeah, how to, you know, I work— me too, I work in media. But even if I didn't, you know, pre being in, in this kind of business met a lot of people, I learned a bunch of stuff that I otherwise wouldn't. I think it's had a net positive impact on my life. I think that AI, hopefully, one of the good scenarios is that it is the same way, right? Where like net-net makes everyone's lives a bit better, even though it has these things that are— Yeah. That are quite bad, right?

I think we should be open-minded to that possibility. I mean, you know, look, We get things— everyone's dooming so hard these days. And you just go down the list. I could— we could easily rattle off 10 very— 3 or 4 big things right now. But, you know, usually the math— we're wrong. We're pretty bad about predicting the future. And so, like, we should be open-minded, I think, towards, like, you know, right tail scenarios where things go well, because I don't think, like, I put a lot of stock or I don't think there's any good reason to put a lot of stock on what the Twitter conventional wisdom is or the editorial board of any publication, what they're all anxious about right now. Like, you know, like, I don't know what that great outcome looks like per se. You know, I do— look, we could get great technological breakthroughs. I'll say one thing, which is that, like, in the circles of people I talk to, everyone is like reading more philosophy. These days, probably because of AI. It's pretty good. Like, I'm pretty happy. Like, that's a pretty cool thing that like people are sort of like a reignition of a certain curiosity about things that I think was sort of, uh, dormant for a while. So like, that seems like a pretty good—

music for the first time in years.

Making music. No, like, like we should pay attention to like the good— I actually do think we should pay attention to the good things that are, you know, because it's so easy to get caught caught, um, focused on a negative narrative. In the meanwhile, like, various, like, positive things are happening. I, uh, you know, changed a flight recently, and it was much easier because I used ChatGPT to, like, walk me through the process of, like, turning, you know, to, to, uh, a round-trip flight to a three-leg flight, which I was very anxious about It's easy. So like, we should pay attention, I think, to things in our like day-to-day life that we can say, you know what, AI is sort of making this better, even if like it doesn't feel like the whole— even if we're like anxious about where it's all going.

Last question. You have a podcast. We have a podcast. That's right. There's a lot of new podcasts. Would you support an open letter that we're writing? We would like you to sign it to ban new podcasts. It would require—

actually, it wouldn't ban, but it would— there would be a multi-year sort of approval process, bunch of safety checks. Yeah.

And, uh, so we're getting a new government agency to issue RSS feeds. Yeah, I— this is really important.

It is. I'm so glad you brought this up because someone's such an important—

someone could just launch an RSS feed that, you know, ends the entire world. Who knows? We don't know.

It can't happen. Or, or removes viewers and listeners from Modlods TBPN.

We cannot let that happen. That's ridiculous. Are you saying that I would do this for my own financial gain? This is crazy. You're saying— I can't believe you would even go there. That's wild. I would never think that way. I'm purely thinking altruistically. Okay, Joe, let's leave.

America needs to win on podcasts. We do.

And this is very important. We actually are objectively winning. We are winning. I just talked about—

we have been winning on entertainment and have no— no, there's no competitors yet. Yes. Well, thank you so much.

I'm signing the letter. I love it. Anytime.

We'll circulate. Great to see you, Joe. Have a great rest of your week. We'll talk to you soon. Take care. Goodbye. Let me tell you about Figma. Agents, meet the canvas. Your AI agents can now create and modify your Figma files with design system context. And we are very fortunate that we have our next guest Jameson Dettweiler here with us in the TVP at Ultradome. Welcome to the show. Thanks for having me. Please introduce yourself and the company.

Yeah, I am Jameson Dettweiler. I am the co-founder and CEO of Chip, and we are building what we like to think of as the ideal second car today. But as everything becomes autonomous, the ideal primarily owned vehicle of the future. Okay.

Can we get the chip that I specced out on the big board? Yeah, because it is a special, special car and I think it would that would provide some important context.

So quick facts, uh, it's an electric low-speed life utility vehicle.

My spec, I love the, I love the, the spaceship there and everything.

And the spaceship, obviously, the surf rack. Yes. And the TV on the front. Yep. This is pretty good. You could get TVPN on there, it would, uh, it would kind of complete it. But good start, good start for me. Before we get too much into Chip, what were you doing before this? What led you to creating the vehicles of the future?

Yeah, I— before this, I was at Tier Mobility, which was Bird and Lime's largest competitor in Europe, backed by SoftBank and Mubadala. They own Spin scooters in the US. So wound up there by selling my HD mapping business, Fantasmo, to them. So basically creating a digital twin of the world. And we built a computer-powered GPS. And turns out scooters don't know where they are very well because of GPS. And so Bird, Jumpit, Uber, and TIER were our biggest customers and scaled up and then also started building kind of driver assistance systems for scooters as well. So yeah, I've been in the mobility space for a while.

Interesting. I can imagine driver assistance would be great for scooters.

I don't know how—

I don't know how far along it's gotten just because driving around LA, some people on scooters do things that, that are, you know, maybe not illegal, but definitely wrong.

Yeah. Yeah. I mean, we say driver assistance is like ADAS in cars, but really it was about restricting their behavior, right? So it's like, are you drunk? Are you going— are you on a sidewalk? Are you going the wrong way down a bike lane? And so you're putting that on there and it just, you know, camera and compute at the edge running there. So it's building a lot of that kind of— yeah.

Why is autonomy important? Important for this vehicle? This feels like the last place I would want self-driving. I need self-driving in the morning commute. I'm tired. I'm going 40 miles or 20 miles and I just want to like zone out and just get to where I'm going. But this looks fun to drive. This looks like you're driving around a small community. It's not that much of a hassle to actually drive. Why is autonomy important?

Yeah, I mean, for us, like where we're starting with the autonomy stack is actually kind of like the inverse of of what you'd expect when you get from a Tesla or anywhere else, right? So like, yeah, we like to say Chip is fun to drive, but parking never is. So this is not like FSD in a Tesla. In fact, like when we do bring autonomy eventually to the vehicle, we'll be, you know, unsupervised in the way that we do it. Got it. But what we're doing today is we can take a Chip, you can pull up to, you know, a busy place in West Hollywood and skip the valet line. Hop right out. Chip will go off, take care of himself, okay? Take care of the parking, show up when you tell him to come back, either from the calendar or as you speak to him or text you, let you know when he's on the way back there. Yeah. And pick you up exactly where you want to be picked up. Unlike a Waymo, right? Which, like, will make you walk a long distance. And the way that we're able to achieve this, by the way, today is with direct remote control of humans because there's actually a data gap there.

Okay. Yeah. Yeah. Is there also an easier regulatory approval path? Because if the— if it's only autonomous or remotely controlled while there's no humans in the car, it's safer. Does that make it easier to get to market faster?

Yeah. So it's definitely an easier path. And for us, like, we think a very responsible path because like when we run our, you know, we're going to be running our pilot in Miami this winter, a great market for us where we're based and where, you know, our early customers are going to have effectively, you know, virtually unlimited access to it. And they'll hop out and we'll take that vehicle, we'll drive it. In fact, we'll drive it at lower speeds on less busy roads, etc. And so it's just a way for us to validate and bring kind of that safety level up. And then eventually we'll drive with people and we'll move from direct remote control to, you know, full autonomy here over time.

Yeah, that makes sense. Jordi Hays, sorry.

Yeah, that use case is cool. I think like, think about like You have a driver. Drop the surfboard off at the beach. Yeah, that, that. But like, yeah, there's something— one of the biggest benefits of like having a driver is just being able to be like, oh, I need to run to the grocery store, pull up outside, go, go straight in. Even if the parking lot is like crazy or chaotic, you're just going in, coming out, you're not thinking about it. To me, that's the most underrated thing.

The number of people that have drivers but not— they don't order delivery groceries, it's like very small.

But yes, no, no, I'm just saying to me that that's been a good example in my experience, one of the most underrated parts. And that's something that you expect the Tesla to be doing soon, but can even make sense in this form factor. When originally we saw the launch video and we were like, why would we need— why would I want autonomous driving for something that I'm just driving to the beach or whatever?

Yeah, well, I mean, it's so you can go to the beach for like if you're living in Manhattan Beach, right, where golf carts are really popular. Yeah. You know, it's small, it's easy to park, but there's still no parking. So in that context, you can take your friends there, get dropped off, and you just go back to home and come back. And, you know, also like really what we're trying to do here is bring this like fun robotic, you know, private driver to everybody. It's why it talks and it can go pick up the groceries, right? And go there and you can do curbside pickup and, you know, the frunk opens.. And because the confused target employee comes out, the vehicle can talk and say, hey, put it in there, we're good to go. Oh yeah, do that.

Yeah, exactly. What do you think? What do you think? Like, it feels like there's been an explosion of like just new vehicle startups. You have Slate, you have you guys, you have Ample. Like, clearly a bunch of smart people at the same time were like, why is there not this super super specialized vehicle, right? Slate was like, it's going after everyday pickup truck. Ambles, you know, focus on resorts, right? You guys are focused on everyday consumers and like in these sort of like, my view, like coastal hubs, at least like, at least that's where I see you guys getting a lot of traction. But like, what, what is it snowballing now? Should we expect a lot more vehicles Is there some like, is it just the advancements of batteries and declining costs? Like, what do you think is driving it?

Yeah, I mean, I think it's really interesting, like when consumer behavior changes in such a major way, right? Especially around cars. And like, you know, in a lot of ways, I think like the American love affair with the car is dying. Everything's too expensive. And increasingly, too, like we're re-urbanizing. And even in places like—

sorry, I was I was gonna hit the boo, but— There we go. No, no, no, I think you're right. And the one— some— even you can see manufacturers now, all the cars started looking the same 10 years ago. Less and less was like a— like nobody goes and they're like, oh, I need a new daily, I'm gonna get a yellow car, right? Whereas historically, like somebody would be like, I need a new car, I'm gonna get like a yellow Volkswagen Beetle or something like that. And now it's all like converging around like the Model 3 format, right? Or even John was complaining the other day because like he's in the market for a Tesla and he's like, I wish that they still made the S. Like, that's the one I want.

Or convertible or a full-size SUV. It feels like there's a— it's almost like a barbell effect. Like they're, they're in the heyday of the love affair with the American car. It was you got one car that was an expression of everything you needed. So maybe 4 doors, but also some power and a supercharged V8. But now, the standard, like, pod that just takes you to work is so bland and standard that maybe there's room for a second car that is just completely on the other side of the barbell. Yeah.

And I think for us, I mean, there's a lot of variables here to like how we think about this. And a large part of this is like, how do you almost like ball out on a budget? Like our car, like our cars are like our fashion and they reflect who we are. But today, if you want to buy a new car that's cool, that's out of reach for the majority of people. And so that's part of what we're trying to provide here and make it, you know, from design, but also like add a lot of function that supports the lifestyles what we live today, because like we are living closer to home, you know, today. And so it's— I think that's a large impact as well.

Do you think this can slot in? It says $15,000 estimated MSRP for 4 seats, top speed 25 miles an hour, 100 miles of range per charge.

I'm going to test the top speed. I'll be the one to decide what the top speed of the chip is.

I mean, that does feel like it would be if you are truly a budget-conscious buyer who's looking for a commuter. Yeah. $15,000, you'd go with like used Honda Civic instead of this. But is there something where the math actually does work out with more charging places where this does become a commuter? Because 25 miles an hour is just like too slow for anyone who's hopping on the freeway for a daily commute, right?

Sure. But we're in LA, right? You know, we've driven this thing. Like, you know, like we're so sad.

But that is like the future. Like the traffic's so bad. Yeah, you really do go 25 miles an hour.

I mean, I think he's just saying like you can have like very easily in L.A. you can just stay off of freeways. Sure. And it's not even necessarily slower. Yeah. Oh, no.

Well, yeah. And that's the thing. It's like if you're, you know, it's 5:30 on a Wednesday, right? And you're coming from downtown back to Venice, you know, like you can be in a car that can do 0 to 60 in 2.5 seconds. Sitting on the 10, but you're still going to get there in the same amount of time.

It's going to be a lot of fun. Are you being on the— you're like betting that there's going to be wall-to-wall traffic and then you get your chip out on the freeway and you're like, whoa, like, yeah, a lot less traffic than I was expecting.

Forgot it was the World Cup and there's no one on the road.

But yeah, for us, like, it's like the ICP that we're really focused on is like, I built this for myself. Like, I like to say that like golf carts like escaped golf communities during the pandemic and the growth area has been like young families and Sunbelt regions. And so for me, I live in Miami, I live in Edgewater, basically downtown area. I'm not going to put my kids into a golf cart, it's not safe. But I like the idea. Sure. This is built to facilitate that. And so that's kind of where we're starting in terms of the function, the price point.

Yeah. And the golf cart companies have been so asleep at the wheel. Like, they— like, when it comes to customization, it's like, oh, you want to customize your cart? You want to order this cart? It's like you have one thing that you can decide and it's the color and there's like 6 options, right? And you can't really change much of anything. And I went down, I was excited to see you guys come up and Ample and some of these other companies because I was like, I've just talked to all these legacy manufacturers and they clearly don't care to try to innovate or try to allow this like specialization, customisation, optimization. And for a car like this, which is like this, like, weekend, you know, this like weekend vehicle, you want it to be like really expressive and fun because, you know, you're not thinking, I'm gonna just be like driving to the office and parking and coming back. It's like, it is the place that you want to express yourself.

Paul Graham had this back and forth with the creator of Friend, Avi Schiffman, who's working on an AI device.. And the question was, how do you counterposition against Apple? If you come up with a new electronic device, a consumer electronic device, Apple's gonna do what they did to Pebble with the Apple Watch where they— Sure. —launched something with the Apple design language. And Paul Graham's advice was, do something so bold with your design and your aesthetic that Apple can't copy you. And this feels like a little bit of this. This feels like very antithetical to what other car companies are doing, such that you could not be fast followed even if the technology is on the same supply chain loosely. This would be such a brand change for most car companies they couldn't jump in. Is that conscious? Is that accidental? What's the thought between— behind all of the customization?

I don't know. I think Ferrari might have knocked off with the Luce a little bit. They might have gotten a little inspired with the Luce, been like, hey, chip looks pretty good, we should make it. No, it's— it's—

you're going to chop the top on that thing anyway.

Yeah. I mean, look, I think Apple tried. They missed. Oh, yeah. The car, right? They missed. No, they're coming back.

Aren't they announcing a— they're playing Apple Maps into the 40s.

Yeah.

And then they're also launching a livery for like a hypercar or like a Le Mans car. Doesn't count.

Yeah.

Yeah. And you had that and like, you know, Ford had a product in this category like 25 years ago. Really? Ford Think. There's one that like actually like is around Venice all the time still. And like, but, you know, these guys are like, you look at all of these companies and like they're all tied to products, I think, that are huge moneymakers but like create problems for innovation there, right, for them. And so this is an area where it'll be like very hard for them to kind of come in and innovate and think about it. And you have to approach engineering differently. You have to approach sales differently and marketing. Like, we're like, frankly, like in a lot of ways, I think of us more like a sneaker and culture company than I do about like a car company in the way that we are selling, marketing, engaging our community, etc. And I think that's going to be really hard for anybody, certainly any of the incumbents, to do today for sure.

Yeah, I should go check it out.

Yeah, we got to go check it out. What— Tyler, what was that car that I was showing you earlier today? Uh, the 1991. We got to pull that up at some point, but I want your reaction to that car because it's a fantastic one. But let's go take a look at the actual vehicle. You brought it here today. Let's go see. Let's do it. Do we have everything set up? Are we good to go? You got the camera? Okay, we're gonna, we're gonna switch over so we can go take a, a tour of a real chip motor. And you got me on the laugh. Thank you for coming along with us. Jordi, you want to hop up? Uh, take us through it. Where, where are we going? Right out here. Let's see, outside the UltraDome. Second car review we've done. We are— whoa, here we go! Wait, it's, it's driving itself. There we go. What's up? Look at this thing. Hey Jordi! Hey John! This is amazing. It's a lot taller than I thought it would be. Wow.

Yeah, it's like Golf carts are actually really tall. So like the typical golf cart that people have is actually like taller than every single trim of F-150. Sure. And so what happens here is there's a couple of things. So you actually want the deck here and your safety structure to be tall because you need to interact with trucks on the road. Yeah. But also, while this looks big, it's a very efficiently packaged size. It's 4 feet 11, 4 feet 11 wide. It's about 13.5 feet long. Yeah. And we've designed this. You'll notice like you're sitting like a dining room table, basically, right? You're not going to be in this thing typically for more than 45 minutes. Sure. So we want to create something that's really comfortable and that's going to actually have a ton of space. And so actually in production here, we're moving the second row back a bit and everything is going to enable, you know, for me, I got— I still have two kids in rear-facing car seats that fit in here well. And so the design exercise like started from like inside out. There's a boxiness to it that's necessary because for us, like function does like drive form. Yeah. And we also like leaned into it because I don't want people to look at this and think that it's a car that can go 60 miles an hour, right? And so the funkiness is something that we've embraced.

Yeah, it feels to me out of the Pixar universe. Yeah, yeah, that was intentional.

Yeah, for me, I mean, I grew up going to Disney World and like I really like— I grew up at like the World of Tomorrow and I love kind of the future that the Jetsons created and we've leaned into into like a little bit of this kind of retro, kind of futuristic aesthetic, because I want to see this positive vision of the future. Yeah, true. It feels increasingly like we're maybe headed toward a Blade Runner future, and yeah, that's not what I want to see.

Yeah, no, I like that. I like that. Uh, what was the process for actually building this one? Because obviously this is sort of hand-built at this point. How much of the supply chain— are you building on something as a demo unit? Are you using something else as a platform. Well, like I imagine that like, you know, you don't have your own leather tannery yet, but— Yes. What, how did this come together? How long did it take to build? Who built it? I'm curious about this process.

This prototype has been built over the last 18 months and tinkered with and optimized a lot. And so we actually imported a small like Chinese EV that we wanted to learn from and tore it down. —like, end up rebuilding.

Distilled it. Yeah, distilled it.

Exactly. Tore it down to the chassis, ended up replacing like the entire drivetrain, the battery, the whole thing. And so like, it's kind of over time, most of that like disappeared. Yeah. And they allowed us to kind of learn a lot from— I mean, frankly, they've been doing these small EVs for a long time. Yeah. And that was really, really good for us. And so yeah, that kind of exploration has gone on for 18 months and now we should have production intent units in hand in the next like 4 months or so.

And then there are manufacturers for like these types of bench seats that you can go to at this point and pull those together?

So actually our VP of Vehicle Engineering came to us from Canoo. He was the vehicle lines director there, but he thankfully started his career in seating and restraints. And so he's like, and by the way, in a lot of ways seating is one of the harder kind of industrial design problems for this. Like there's so much nuance to doing it well and there's a lot that gets improved from here. And so we're having actually like optional heating and cooling in here too. There's also like optional HVAC. And so that's a huge, a huge part of our focus. Yeah.

But wait, so you'll have a covered version of it as well? Yep. Yeah.

We have a soft top that it will come with. We have hard doors, we have an enclosure here. We have HVAC as all as an accessory. But like, it's, you know, even if you're gonna be operating this in a cold place, you know, like we took this through Aspen and people absolutely loved it. Or if you're in Seattle where it's raining, like if it's a nice day, you should be able to top off.

So our designs, we like to say— Driving this home from dinner in a winter storm would be electric. Well, if you're going to do— if you're going to make a mountain version, you got to get chains, I guess.

Yep. Well, it's all-wheel drive. We use 4 in-wheel motors as well. So it has a lot of torque. So like it has no issue in the snow.

Crazy. Yeah. Interesting. Very cool. Interesting. Very, very cool. I want one. Yeah, I can make that happen.

What is this track back here for?

This is L-track. So this is an option here. This is a standard kind of accessory you'll see in truck beds. Okay. This is actually also the same that they use to like hold down airplane seats there. So this is an accessory. You can do tie-downs and things like that. We have— so these seats actually fold down into a truck bed. Sure. So you can use it, you can put plywood. And so we designed it that way. So like the thing is, people like end up wanting to buy this kind of vehicle because it's like the ideal vehicle to go to the beach, but what ends up happening, and you've seen kind of this behavior, is like, they're like, oh, this is functional. They'll go pick up the groceries, they'll go to Costco, they'll take the kids to school. And we want to take that kind of like feeling of a vehicle that really empowers you and feels like very light and turn it up to 11. And so it's a Swiss Army knife that you can do, and you can pick and choose what you want. You know, if you're a surfer, you can have your surfboard rack here. We'll have like a hand shower for like getting sand off and just go heavily into that accessory.

Yeah, yeah, yeah, yeah, that's cool. It's a great way to differentiate. What does it take to keep this, uh, thing street legal? It feels like right on the fringe, right on the edge. You have mirrors but no doors. Like, what, uh, what else is like the bare minimum to make sure that you're always street legal?

Yeah, so the category that we're currently operating in is called a low-speed vehicle. That's what a street legal golf cart is. It's FMVSS 500. Okay.

Uh, you have to, uh, that's because of the top— you have to cap the top speed.

Top speed, yep. And you can ride, drive on roads 35 miles per hour and under, so you can ride, you know, basically every surface road here in Los Angeles you can go on. Wow. And then, yeah, you have to have seat belts, you have to follow lighting requirements, you have to have a backup camera and things like that. But frankly, like, we actually think that the category is like severely underregulated. Like, golf carts are death traps, and there's a lot of like really, really bad stories that you hear. Yeah, of course. And for us, like, we're doing a lot to not only exceed but push the boundaries here. And so for us, we are doing, designing this for crash at like the appropriate speed that we're operating at. We're putting in features that we're gonna be deploying for active safety, leveraging the cameras and everything else we have in there too. So. Under-regulated.

Yeah. Gonna be opening up the Chip Washington office soon and pushing for, pushing for—

Total and complete lockdown. So, This is basically a one-of-one. Yeah. Uh, is it street legal? Did you have to tow it here? Can you get something like this approved provisionally for a demo drive so that you can test it on streets, or is this something that, uh, you have to go through a whole process and they all have to be street legal before?

Yeah, no, this is street legal.

We registered. Yeah, there's like a custom build there that we've been able to do. Uh, yeah, uh, we, we typically tow it when we come from like West LA to here, just we can drive it. Yeah, yeah. But, uh, Makes sense. I don't like to think about how much this vehicle is worth because we can't get insured for the value. I think we can't do that. As a startup with a prototype. Yeah. Yeah. You will not find an underwriter. So yeah, we take good care of it.

I love it. I love it. Thank you so much.

I mean, it's impressive how solid it feels that even as a prototype, like it feels like ready to ship.

Well, we'd love to get you guys, you know, later we can take it for a drive.

Can't wait. Can't wait. Thanks for coming by.

See in here? Is this, uh, open or, uh, not right now?

Here, this is the party in the frunk, but we have, uh, a little bit of wiring we've got to clean up there right now. But yeah, but yeah, that's where you see the TV in the front here. Yeah, yeah. And then it has got to have a huge space you're going to fit— you got to miniaturize stuff.

Yeah, exactly.

And we have, um, you know, my goal is I got— I need to have a double stroller that will fit in there that I can keep in here at all times.

Yeah, very cool. No, I love it, I love it. Well, thank you so much. Thank you.

Jordi Hays. Incredible. No, this is amazing. Yeah, awesome progress. Thank you so much for showing it to us.

Appreciate it.

Yeah, thanks for coming on.

Thanks for having me on the show. Have a great day. Cheers, guys. Very cool. See ya. Goodbye.

Ah, friendly little guy.

To the show. We have a couple more guests coming by. Tyler, did you find the car that I was calling— I was talking about yesterday? Can we, can we send that? Uh, it also looks— the Nissan Figaro. The Nissan, right? Figure out, uh, Nissan Figaro. Can we pull up a picture of that? Uh, I like the way this one looks, and this is just like for sale on Cars Bids right now, this old thing. Uh, also looks a little bit like—

what a cool, what a cool vehicle. Yeah, it's so— I'm so glad that consumers are getting options because like wildly different design decisions than Ample, which obviously launched right at the same time. Yeah, Ample's going for like the very low-tech route, like the most simple possible vehicle. Like they're not even really selling, I think, a screen in it. You like that? I like that. I love that. Figaro. I love it. Whereas Chip is going for really creating like a consumer electronic product. Yeah, in many ways.

Tyler, do you want to send a link to the car that you were looking at picking up earlier? I think it was a little bit of an older car. I want to get Jordi's review of whether or not you should pull the trigger on this car. Have you ever— have you ever owned an old car?

Not since, uh, College?

Because yeah, uh, what do they call it, vintage or retro? I don't know. You go further back because— classic class. Yeah, it's really classic cars.

Yes. So if we can pull this up, this is from early 20th century, 1922 Chevrolet. Oh wait, oh, it's pre-war!

Yeah, I thought you were going post-war.

Uh, there's a few options thinking about getting that.

Whoa, whoa, whoa, whoa, don't you think I mean, but why not just go for something horse-drawn?

Like, if you're going to go this far back, why not go 20 years prior and go horse-drawn?

Who would have thought this car would be put out of business by horses? Because yeah, if you're going to go— a horse is less maintenance than this thing. It's over for the— what is this, Model A, Model T, something like that?

Well, it's not Ford, so—

oh, it's not Ford.

Okay. Chevrolet, uh, 490 Touring is what it's called.

I mean, they kind of nailed the design.

It's beautiful.

They made the perfect car, and then ever since then we've been trying to get back to this.

Yeah, I think it's wildly illegal to design a car like this. I don't think you could even if you modernized everything.

Yeah, we don't even know how.

But we do know how to 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. Um, Adam Hunt, researcher at Cambridge University, has a PhD in evolutionary psych— psychiatry. He said he's recently flipped from being bullish to bearish about AI. He says he's updating his bearishness to be more solidly bearish. Early thoughts, uh, and his reasoning laid out here. The whole, it turns out if you keep training and scaling the models more they develop broad new capabilities in lots of domains thesis was wrong. Sorry, Demis. The recent batch of models haven't gotten more general. They've gotten less general. This is the most obvious fact that their language outputs have gotten much worse in comparison to e.g., GPT-3. If they were gaining generalist capacities, we would expect them to be describing their work in ever more graceful and comprehensive prose. And I don't know, I don't know if that's that accurate. I mean, you hear about like Claudish and like the little languages that emerge from the models to be more token considerate. And so there is an evolution of language there, but I sort of take his points that it feels like, it feels very much like the spikes are getting spikier. And if you scroll down, you can pull up this chart. So Thomas Pueyro had put out this mainstream mainstream AGI thesis, this idea that there's spiky intelligence. It starts— if you zoom in on the top left there, you'll see the circle where there's tasks that AI can do and a human is better at every single task. Then the AI is helping me in some tasks means that there's these little spikes where the AI is, is maybe 80% as good as you, so it's somewhat helpful, but there's other areas where it's useless. Then all of a sudden, You scroll over and the AI has a jagged frontier. Sometimes it's amazing, sometimes it's dumb. And then the final stage is, wow, the AI is unbelievably intelligent, but for some reason it fails at X. And then AGI is when it's still spiky, but every possible point on that blurry blob is beyond the human capabilities. And so it is generally intelligent even while remaining spiky. Now, if you scroll down, this individual has a different story that he is playing out. AI is a fun toy that is helping me do some tasks, has a jagged frontier. But then if you go over, he's saying, you're going to get crazy, crazy spikes while the actual frontier broadly sort of stays within that human circle. And then at the final stage, you have this like extreme strength stronger in a few domains like coding and math, but it's weaker in many other areas. And he claims that through the process of optimizing for coding and cybersecurity in these specific business use cases, you might actually shrink the broad capabilities. But Tyler over there thinks he knows better than the PhD in evolutionary psychiatry from Cambridge University. No, we were debating this earlier. But give me your pushback. Back because you think that this is maybe too soon to call this or not exactly?

I'm not saying that it's not correct, but I think if we want to make models really good at language, common sense, or social reasoning, it seems pretty clear to me that right now we could do that if we wanted. All of the recent advancements are because we want a really good coding model so we can make the model really good at coding. Yes. It's not clear that it's impossible to make lang— like, yeah, maybe it's true that the methods we're using right now to make models better don't apply, like, equally to all kind of types of capabilities. But it, like, there's a reason that, like, the models aren't really good at social reasoning right now. It's because, like, it seems like social reasoning is not that valuable for models.

Even writing. Yeah, it's like, look at some of the most iconic, like, exchanges in business history. You're never even like, never like, oh, well, when Steve emailed the Adobe CEO, his prose wasn't absolutely electric. Yeah, it's just like, what did it actually say? And so I think it's possible that, like, incredible writing just isn't valuable enough to—

Yeah. I mean, so I agree with you that there's a world where, like, every task needs to be hill climbed. There's some debate on the timeline today about ChessBench. So Gary Marcus is taking a little bit of a victory lap saying, AGI smarter than the smartest humans? My ass, says Casper Averov, peak rating 2851, could probably, probably could have crushed the best commercial language models in chess when he was 7 years old. And so this graph of Chessbench shows that GPT-5.5, Gemini 3.1 Pro Preview are all around just north of 1,000 Elo, which is still impressive. It's crazy, but it's not by any means superhuman. Human. And but, but like every time someone points out something where it's like the models are bad at this, it does seem like there's a very quick cycle of like, let's make them good at that.

Yeah. Also, like, I think it was Amjad that was talking about this. He basically took an LM and he said like, help me make a chess bot. And then it was like became superhuman. So like if the model can just do the thing like, yeah, maybe it's, is it the model doing it or is it the model writing code that then does the thing.

It's kind of the same thing. Yeah. Yeah.

I mean, at a certain point— Like it's using tools to—

Well, then at a certain point, a really good model would just be like, I'm going to go download Stockfish and just run the actual chess engine. Like, I don't need—

Yeah. I mean, that seems fine to me. I don't—

I just think that when people say the model is generally intelligent, they're hoping that it's just like a natural—

Like they're implying that chess is like indicative of some other— cognitive ability, right? Because clearly the chess is not the actual thing that matters because the chess can just download Stockfish. Yeah, yeah, yeah.

I don't know. Anyway, fun stuff. 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. I really like the dog one. It's great. Anyway, we've been keeping Stanley Tang from DoorDash waiting for far too long. He's the co-founder and chief product officer. Stanley, welcome back to the show. How are you doing?

Hey, how are you going, Jordi? Good to—

Great to see you.

Yeah, I could see it. I want to get the update. I want to get the news on a big day. I sort of— I also want to know your take on this idea. I don't know if you were listening, but this idea that LLMs and AI models are so incredible and yet they're underperforming at chess on this chess bench. How are you? What is your overall frame for thinking about progress in AI, the intelligence, the spiky-ness of these models? Because I imagine that that flows through what you will do with a harness or with a new initiative and what you will just wait for the model to naturally get better at.

Yeah, I mean, I think I feel like we're going through one of the most exciting phases in kind of the tech industry. I think AI is going to be huge. I think the models are only going to keep getting better and better. I mean, I spent a lot of time thinking through how do you apply AI, not just in the LLM world, but seeing all the stuff that's coming out of the AI LLM world and bringing that to the physical world. Yeah. And I think that's probably going to be the next frontier. Yeah. Like robotics, physical AI, drones, etc. And honestly, like a lot of the things that you're— the reason why you're seeing like the Waymos that are happening, the DOT robots that DoorDash— that we announced last year, the drones, a lot of these, it's all part of the same revolution. Is these AI models getting better and better. And the moment you sort of flip from kind of this classical heuristics approach to now this AI, ML-learned approach, that's when all these robotics started working too. And so I think it's— things are just going to keep getting better and better. So I think what you see now is probably the dumbest version of what you ever see.

Yeah. So give us the overview of DoorDash Air, and then I want to know how AI is accelerating the deployment of that program, how it's integrated into the actual vehicles and the drones and the business broadly. But just, just set the table for us on DoorDash Air.

Yeah, well, we're really excited to announce DoorDash Air today. It's, it's, it's our in-house drone delivery program. It's the same team that's built behind DoorDash Dots. It's built as part of our DoorDash Labs group. And I think the big milestone we announced today is that we officially got our FAA Part 135 certification. So essentially it's a, it's our air carrier certification. It's the same certification that actually all the, whenever you charter a corporate jet, you have to get a Part 135. And, and, and we are, and with that it means we are now officially legal to do drone deliveries commercially in the US. We're the 8th company to get the Part 135 for drone delivery. So we're really excited to bring drone deliveries to the DoorDash marketplace. Yeah, it—

this feels AI accelerated because if, if we were talking 5 years ago, 8 years ago, 10 years ago, and like the discussion was around launching sidewalk robots or drones, the Or would be very important there, I imagine. And it feels like this year it flipped to and very quickly. And is that because the team is accelerated with coding agents and AI, or is it just that the opportunity's so big and your company's bigger that you can really, I know it's one team, but I imagine the team can be very big now, but what's actually enabled you to move so quickly here?

Yeah, I think it's a combination of a bunch of things all coming together. I think, I think certainly the AI piece, the models are getting better and better. Things that weren't possible 5 years ago, all of a sudden now the DOT robots can now handle it, drones can do it. I think the second is the hardware as well, just hardware getting better and better. This investment, right? Because of the AI revolution, there's so much money now being invested in things like compute, right? You know, like again, like 5, 8 years ago, we didn't have things like the NVIDIA Thor chips and things that could actually run these powerful AI models. There's a lot of investment going behind that. And obviously now with, with robotics and physical AI being a thing, like a lot of investment going behind the hardware itself as well, like the drones, the components, the supply chain. And then last, last piece, of course, is the operational piece as well. Like I mean, obviously we've actually been, you know, we started DASH Labs almost 8 years ago. It took us a while to just even figure out the operational piece. And I think that's the other big thing a lot of people don't realize is with something like drone deliveries, it's not just putting an aircraft out there. A lot of what we spent the past few years working on is also thinking through what does the end-to-end stack, the end-to-end infrastructure looks like. Like, how does this— how does a drone actually plug in to the DoorDash network? Like, what is the loading infrastructure at a merchant look like? How does a merchant handoff look like? How do you— how do you route a drone order, a robot versus a human Dasher? How do you— how do you even know an order is drone eligible or not? Like, how do you build— so, so we have to build— we have to build a device called Smart Scale where we give out to these merchants where you have to weigh the device so that, you know, it's actually, you know, there's weight limitations on these drones, that it's actually a drone-eligible order. So we have to build all that. I mean, a drone is a— a drone itself is not particularly useful if it's not connected to our marketplace, if it's not connected to our customer, to our merchants. And we spend a lot of time working on those. So I think it's— you got to do all of the above. I think it's a confluence of a lot of these things all coming together, which we feel like is, which makes it the kind of the perfect moment to do DoorDash Air.

So there's a lot of AI, I imagine, accelerating the, just the development of the SaaS products that you need to create the ecosystem, actually run the drone program. Then there's probably some AI-assisted hardware development, everything from procurement, sourcing parts, the supply chain. But then what does the AI actually look like on the drone? You can't probably bring an NVL-72 with you. So what is that? What does that look like? What does the autonomy stack look like on a modern drone these days?

Yeah, I mean, we will have more to share about the drone itself, but I mean, you can imagine it's just even things like how does a drone— like, I think, like thinking through, like, for example, the pickup experience. Experience or the merchant, like how does a drone actually, the way it docks at a drive-through standalone store is very different than a strip mall. Like it's very different than like a rooftop. Like it's like, how do you build kind of the perception system to kind of handle these various merchant pickup infrastructure? Again, that's part of going back to like, you gotta think about the system end to end. And then on the consumer side, again, it's like there's a lot of like nuances, like how do you know like the exact coordinate to drop the package, right? It's not as simple as, oh, just take the lat long from Google Maps and okay, that's where you drop because that might be over someone's like car or someone's like building where there might be like wires or et cetera. So it's like, how do you build the perception system that? And then of course, like, while it's actually traveling in the air, like, how does he— how do you handle deconfliction with other, other drones? Um, I mean, and, and, and for example, in airlines, like, the way you do it is like every, every airline sort of has this thing called— it's called ADS-B, which essentially, think of it as like it's Find My Friends but for airplanes. Um, so I'm a pilot, so I actually kind of deal with this all, all day. But with drones, you can't have that. So it's like, it's like, so what is that? What becomes system to, to, to do aircraft deconfliction. There's a lot of— again, there's a lot of nuances that goes into building something like DoorDash Air. And again, it's not just getting a drone just to fly. That's like the easy part. It's how do you build kind of the end-to-end system, actually commercialize drone delivery, manage the fleet, handle pickup, drop-off, deconfliction, merchant loading, routing, which areas, what orders are good for drones, etc. It's that whole system and that's Honestly, that's probably more of what we've built as part of DoorDash Air. In addition to, of course, we got to build the aircraft as well to get our FAA Part 30, Part 135 certification.

Tell me the story of DoorDash's original go-to-market, your selection of markets, because you took a pretty contrarian stance, if I remember. And then I want to know about, uh, if you're thinking of applying that same philosophy to the rollout of DoorDash Air?

Yeah, I mean, I mean, even— I mean, we started DoorDash 13 years ago in Palo Alto. It was literally a Stanford food delivery project called Palo Alto Delivery.com back then. 8 PDF menus and a Google Voice phone number. We're delivering out of our own cars. You know, like, like Tony had a Honda Accord. We were just driving around delivering orange hummus.. So we've definitely come a long way, but I think the, the principle remains the same. It always starts like, like, I think, again, this is, I think this is a, a common trap a lot of these companies make is instead of saying, oh, like, let's just build something cool, a cool technology for the sake of it, you always have to start with what's the use case? What's the customer use case? What do these deliveries look like? Especially with DoorDash, I think, again, people kind of have this mis— misconception that, like, delivery is very nuanced. It's not like, you know, we do over 3 billion deliveries a day. It's not like you just— like, all 3 billion, all 3 billion deliveries look different. You don't just like plop the drone.

Specifically, that number is so mind-blowing.

Yeah, it's crazy. But, but, but specifically, I seem to remember, I don't know if it was investors or just commentators, but people were saying like, it is crazy that DoorDash started in a suburb. Everyone assumed that delivery services and like Uber, for example, would start in like the highest density because it's the lowest distance to go. It worked out. What is unique about where DoorDash Air you think might be successful? Because people might say San Francisco, they're techies. Yeah, they might say New York, but there's different considerations around tall buildings and regulations, like where should we see this flourish early? Where's, where's the last place it's going to go?

Yeah, I mean, it's actually, I mean, in a way it's kind of similar to DoorDash. Like DoorDash thrived in the suburbs and I see kind of see the same roadmap here of drones. It's like, again, you start with the use case where the delivery's happening and does this use case fit into our form factor? And turns out drones do really well in suburbs because one, you can travel, you know, like It's great for deliveries where things are further out, where perhaps there's not a straight line you can just get to, or maybe more rural areas where it's less desirable for human Dashers to take an order. But, and also, of course, like drones don't have to sit through stoplights and things like that. And then, of course, single-family homes are like, in suburbs, are like the perfect use case for dropping off the drone package, like right in front of our in your front porch or your backyard. So again, it's like, I think you always want to start with the use cases and kind of figure out like, okay, like these are the deliveries that are really good for drone deliveries, but then these are deliveries that are really good for our DOT robots. And then the remaining maybe is a complicated grocery order with like, you have to go up steps or high rises. You still want humans to do that. So again, it's all about like picking and choosing. There's no two deliveries across our 3 billion delivery deliveries that look the same. And it's all about how do you, you know, how do you build kind of this multimodal system, right? And that's kind of the future we believe in is kind of this multimodal platform.

How— what does range look like for DoorDash Air over time? Like if I have a favorite restaurant that's 30 miles away, am I going to be able to get that delivered hot?

Yeah, maybe one day. I mean, but so far, like, like most of our deliveries tend to fall between kind of the 3 to 6 miles, maybe it's up to 10 miles. And if you can kind of get a, you know, a coffee delivered in under 15 minutes when before it might have taken 45 minutes or an hour like that, that's like a, that, that would be a killer use case. Again, it's like, what are the use cases we can unlock that is possible with drones that you couldn't do with someone that's in a car? You're, those are delivery.

You're thinking about ordering French laundry. To Malibu? Is that what you're trying to do, Jordi?

I'm actually looking forward to getting into the 500-mile range. Exactly. I want you to break the sound barrier. That's what I'm asking for.

Over time, I'm sure you'll get there. Well, congratulations.

Yeah, really, really cool.

Thank you so much for taking the time to come chat with us. Have a great rest of your day. Congratulations. We'll talk to you soon. Cheers. Goodbye. Let me tell you about CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. And we got Amjad Massad from Replit, the CEO, the founder. You know him, you love him. He's been on the show multiple times and he's back. How you doing? Look at that. What we got on the shirt? Oh, look. There you go. Thank you. Wow. Loyalty. That's— that's vintage. That's vintage. Thank you so much for taking the time to come chat with us.

My pleasure. I haven't seen you in a while. It's been way too long.

Yeah, I mean, we went to your office, we recorded an episode in your— most people don't know this probably because it was just a blank wall in the background, but, uh, we recorded an episode in your office and, uh, we didn't actually get you on the show that day because we were just hanging out. But I'm so glad that we have you here. Uh, catch us up to speed on what's happening in Replit world, the new, uh, the new launches, the AI-powered Creative Suite, uh, and then I have a million other questions to ask you.

Awesome. Yeah, so, um, a few things are happening. One is, you might have seen, we signed this big letter with NVIDIA and Microsoft around open models. We're really excited about open models. Look, for 2 years, Anthropic was the undisputed king of coding agents, and, and right now, actually, it's not just, uh, GPT that's caught up, but also the open-source models. Kami K3 is an amazing model. It's at maybe quarter the price, sometimes even better. Um, and it's also a fantastic design model. So we have this new—

what's that? I saw the Arena leaderboard. We were talking to Joe Weisenthal about this an hour ago, um, that, that Kimmich K3 really outperformed Spikey on front end, which is— we were talking about how it changes the conversation, because if it spiked on bioweapon development development, that would be a different conversation. But, uh, it's— if we wanted to spike anywhere, front end is probably a good place for it to spike.

Absolutely. Absolutely. It's a— it generates beautiful aesthetics. It's the first model that I see that is actually trying to reason about design, trying to— like, if you read the traces, that's the cool thing about open models. They kind of show you the traces. Now, those models don't show you the traces anymore. It thinks about page structure and layout and aesthetics and things like that. And so Replit Design, for one, is the first product on the market that's purely multimodal. We kind of, we eval every model on the face of the earth. Some of them are good at SVG, others are good at animation, others are good at CSS. And so we created an environment that brings up the best of these models. You can also select the model you like. You want to design purely with KIMI? Go ahead and do that.. And so there's no other place on the internet where you could do that. And, and kind of the motivation behind that is, you know, we, we've been on a now multi-year crusade against slop. We don't want to be purveyors of slop on the internet, right? I mean, we're all inundated with tech slop. I mean, LinkedIn is sort of unusable at this point, but also all the apps kind of look the same.

Because I haven't logged in in months.

Yeah, yeah, it's, it's not worth it, honestly. And so, we're like, okay, how can we make it so that— look, we think people largely have taste. It's hard for them to express how they want designs to look. And so, we came up with this concept we're calling ambient intelligence where you can design by point and click. And the way that happens is you sort of— we Well, first of all, we give you a mood board. We integrated with Mobbin. It's typically a paid service. We're providing it for free. And so you can find inspirations from around the internet, from the best UI/UX websites and landing pages. You can drag and drop them into the canvas. Very cool. Then when you click on an artifact on the canvas, it gives you options to continue the design. It's like, let's design another page. Let's look at different vibes. Let's generate different variations. And so I'm someone who who I think I have a good taste. I know what looks beautiful, but I can't express it. I can't say, you know, it should be brutalist style, whatever. I don't have the language. And so, by just pointing and clicking, I'm arriving at really cool designs. Like, I'm working with an architect on our new office. I kind of know intuitively what I want my office to look like, but I can't go on Pinterest and search or whatever. So, I went to Replit. You know, put an initial prompt for things I like, create a kind of a mood board I like, point and click, point and click, an hour later, I have something that exactly that I would like. I would never in 100 years be able to express it. So, it's really cool. That's the beauty.

I mean, that's— design for non-designers. That's— sometimes you don't— sometimes you don't have words for it, but you also just need to see every possible direction, and then you can really hone in. And to me, that's It's so crazy to me. I used to go, when I was like a teenager, I'd go on the freelance platforms and I'd do like a $5 project to make a logo and I'd have a bunch of people do it. I'd spend whatever, I'd spend $25, $30 on it. And I was so excited to get all these different kind of variations. And now it's actually as simple as a single prompt. You can just put in the prompt, make me 30 versions, you know, don't just make a logo, make 30 versions of the logo. So yeah, it's not even a prompt, it's 2 clicks on, on Raplet.

And we're actually, we talk a lot of internally about the post-prompting world. Like I think we're entering a period where prompting is no longer needed. Some people will still do it, but it's not just in design. I'm sure you're kind of following the narrative around loops. Where you specify a high-level intent and then agents prompt other agents. Mm-hmm. And so this is something we're increasingly focused on because again, people are really not good at prompting. Mm-hmm. It is like, you know, we went from coding, which is a very low-level kind of task, to prompting higher level, but it's still somewhat low level from how people think. The way people think is they specify an intent. Like, I want my sales leads to get responses in 5 minutes. Sure. I don't want to say I want the routing to be that way, and that way agents should be able to figure it out.

Yeah.

How are you thinking about— John wanted like a large kind of lever, like a slot machine to not have to prompt anymore. Just pull the lever, you know? Yes. If you could use that to start a new loop, that would be a very satisfying user experience. How are you thinking about—

It'll be like these Vegas casinos where all—

Yeah, exactly. Exactly. Game. How are you thinking about game development? There's been some very impressive web-based games. There's been some acceleration in what I'm seeing on Steam with little niche games, and it feels like that's a place where you're particularly just well set up. And I feel like a lot of— I don't want to characterize the Replit community too precisely because it's huge, but it feels like there's a lot of folks who want to build a business, they go to Reddit. They want to build a tool, they go to Replit. If they want to build a game, they might be going there too. They probably already are, but it feels like that's the next thing to really fall, really get polished, really get deployed into the mobile app stores, into Unreal Engine. And there's more harness there and partnerships that I could imagine you creating a really amazing product around.

Yeah, I will say last week has been a huge week for AI for game development, right? I mean, for a long time, we actually spent some time about a year and a half ago building a game dev stack, which still exists on Replit, but it wasn't very good. The models were really bad at 3D reasoning. Sure. And I think between Kame K3, Opus 5, now they're starting to actually be good at, at 3D. Yeah. And so it's, it's something that we're, we're paying attention to. There's actually already a large, uh, you know, community of people who do a lot of 3D and games things on Replit. We, we recently launched a CAD integration so you can like make real physical products. And, and we actually have some folks in the office that designed something on Replit and actually went and produced it. Yeah. And so it's not just games. I think it'll start affecting robotics, it'll start affecting kind of physical industrial design. And so it's a, it's a very exciting time and we're, we're so, we're paying attention to that.

Uh, what's been the biggest source of growth for you this year? There's a lot of companies that are doing, you know, direct response advertising, influencer marketing, channel partnerships. There's so many different ways, just SEO, GEO. What stuck out to you as a particularly valuable marketing strategy that's grown the platform?

We do all of the above. I think we have one of the best performing UGC ads right now. Yeah. Uh, I'm sure, I don't know if you saw that, but we have like a bunch of, um, bunch of videos where people, there's like one that's really going viral right now where, uh, where like typically a woman is getting like an app from Replit, like a Replit-built app, uh, about like, um, uh, you know, asking her out on a date.

Oh, I think I have seen this.

Yeah. Yeah. 25 million views and all of that. And then now a lot of our competitors are copying it. But the biggest source of growth has been enterprise. And I think Replit differentiates in many ways, but one of the ways we differentiate is because we've been building the platform for a long time, we actually have a really deep stack. Like everything, when you buy Replit, you get a hosting environment, deployment environment, security agents, you get admin controls, you get all this enterprise stuff. And we've been growing really fast in enterprise because, you know, Vibe Coding has been kind of proven out and now we've went from kind of the consumer excitement that happened last year and now real adoption in enterprise affecting real impact on how people actually work. And the really cool thing is, you know, when we talk to our enterprise customers, it's like, this is the most impactful tool, more so than like the CoWorks of the world where you're sort of like, you're doing the same thing, you're sitting in Excel or whatever. It's, it's saving you 10 minutes here, 15 minutes there. Uh, this is actually automating entire workflows where it's like, I used to do this for 20 hours a week. Yeah. Now I built an app and I do it for 0 hours a week. Yeah. And so you can automate entire work. And we have a bunch of products and launches coming up around this idea of loops and really automating work coming, coming soon. And so enterprise has been the largest growing segment for us. We're still growing consumer, prosumer and all of that.

Are enterprises asking you to sort of act as a de facto router? Because, because you're model agnostic and you understand the different pros and cons, there's a world where if I have a workforce that's pretty deeply integrated into the Replit ecosystem, I could either go to you or maybe even enforce a rule that says, hey, for front end, we, we're using Kyma because it's cheaper and then this meets our standards.

So at this company we will ask you for that. Okay. And they're asking for that granular control on a user and a group level. So I want my designers to be on KIMI. I want my frontend to be on that. Yeah. They want that deeper controls. And what we're seeing from enterprise customers is that they're saying like, look, I know the big AI labs are trying to lock me into these big contracts where I'm just using them. And there's this whole token maxing thing that was happening and everyone was sort of burnt by it. And there's like really deep mistrust, uh, be— because also a lot of the big AI products, um, don't put guardrails around the product. And so a lot of employees, you know, you, you have instances where employees are spending, like one employee could spend like $10,000, $20,000, uh, a week, uh, like a month doing that. Um, and so they're asking us for control controls on models, on spend for routing, but also I think more importantly is I think the guardrails that we build into Replit. So when you're doing design, you're in design mode, the agent is a designer, it doesn't go off, veer off and do a lot of random things. When you're building an app, we have the guardrails around building apps. And even with that model routing, Replit is a lot more outcome-focused and actually cheaper. And we see a lot of companies now ripping out, say, Claude, they gave everyone Claude, right? And they're like, okay, my engineers are great at it, but actually people in legal and finance, they're burning tokens for no reason. So let me implement Replit and let them be focused on a certain task.

Yeah, there's a remarkable story in the chat. Somebody in the chat says that 2 years ago they were in $50,000 in debt. They would meet business owners at restaurants and build them products in real time with Replit, and that got them out of crushing debt. I'm sure you hear those stories all the time. White belt. Yeah, total white belt. Thank you for your service. What advice do you have for young people who are nervous about the job market, AI, how can they take advantage?

Look, look, the narrative has been totally warped. Like, you know, the, the politicians, you know, just the discussion and debate in our country has been about job loss. It's been about, you know, spend. It's been about all the stories like this. Someone who was $50,000 in debt discovered Droplet and got out of debt and perhaps is making more money than that. Those stores are not getting told. And I understand like, you know, negative headlines get more attention. I think there's a perverse incentive for some of the players in the space to generate the negative headlines because you can create regulations coming out of that. And we know the story of regulatory capture that's been happening. It's a story as old as capitalism, right? And so kind of my advice to people is kind of a, first of all, mindset shift. AI is incredibly empowering. Look, even for me, you could say, oh, it's easy for you to say. You've been always like a technical programmer guy. I've never trained a machine learning model. Now I'm training models with, with Replit. I'm the owner of Replit Agent. I actually trained a chess model that's on its way to become the best chess LLM model. It's actually already beating state-of-the-art GPT models and these kind of models. And so I feel huge empowerment. And to me, it's like, fuck, there's nothing I could not do. Like, I was thinking like this math thing that's happening. I've never done math.

I want to go find some proofs.

It would like bio— Conjecture max. Yeah.

Conjecture max. Do you know there's this video where PDD, I guess,— I guess he should not be named now, but there's this video where he closes a deal. He's on the phone. He closes the deal. He hangs up the phone. He's like, "What can't you do, I can do." And he gets goosebumps.

Yeah. It's unfortunate that he is— that is him in the video because it's a great video.

It's a great video.

It's a great video.

And I feel the same way with AI. It's like, what else? What's out there? Let me go, let me go do it. So I think that mindset is important.

I love it. Well, thank you so much for coming on the show, breaking it down.

Awesome to see you. Always great to see you guys.

Always have a good rest of your day. We'll talk to you soon. Goodbye.

Let me tell you, someone should use AI to take that video of He Who Shall Shall Not Be Named and put Amjad in the video. That might be worse.

I don't know. 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. That— your soundboard is fantastic today. We have Adam Cohen from Weave, the co-founder and CEO, coming on the show with amazing fundraising news. Adam, how you doing? Good.

We're in good spirits. We're, uh, yeah, we, we signed with Standard, so we're—

What's behind you? What's behind you?

You gotta go. What do you got there?

Oh, the gongs are going viral.

Why don't you guys hit it in unison?

Yes, tell us, how much did you raise?

Uh, we raised $13.5 million thanks to Standard.

Congratulations.

Now All right, quick history on the company, please. And yeah, then we'll get— we can get more into the round and where you're going.

Yeah, so we, uh, we started this company just under 2 years ago. We went through YC at the start of last year. The idea originally would just measure software engineering. Um, very quickly this turned into measuring AI because that was the, you know, kind of took over. Um, and what we realized ahead of the curve— we've always kind of been working with the token maxers of the world is figuring out once Opus 4748 came out, people realized they didn't need the best models for every task. And specifically for coding, this is like 80% of AI spend. So we built the world's best prompt router. It reads your prompt and actually shifts it to the right model at the cheapest cost. And this is what Standard's like super excited about. And we are starting to, yeah, go crazy with—

who do you sell to? Do you sell to the CEO, the CFO, the CTO?

Yeah, we sell to the CTO. Uh, CFOs, CEOs obviously love us, but like, we're only focused on engineering. We think that's like still the majority of where AI spend is happening. Um, so we're just focused on like every engineering task, uh, as opposed to like, yeah, a lot of the other routers are trying to do some other things or either app-level routing or just code routing.

Okay. And then, yeah, how, how do you differentiate?

Break that down. App app-level routing. That's somebody that's integrating a router into their product so that behind the scenes it's just deciding it for the user. This is more of purely this sort of enterprise focus where teams are doing various tasks or they have agents running and you're sitting in between them and the models to run all of those processes as efficiently as possible.

Yeah, that's exactly like the app-level routing for us. What we're doing is like anybody using Claude Code or Codex plugs in our router, it sits directly in their current workflow, adds like next to no latency, and it will route to whatever the most cost-efficient model is. So we host all the open-source models in the US. They prompt, it routes to the right spot, and they save up to, well, some companies up to 80%. That's amazing.

How— what is customer acquisition like? Is this, you know, like person to person at the highest level? You're going after enterprise CTOs, so you're meeting them, talking to them, or is there like a self-serve option for like small businesses? Because it feels like a lot of small businesses, they might be starting to use AI, starting to use coding, but a lot of them are on a $200 plan. They're getting great value. They don't really they're not really in the, like, token maxing is going to, you know, make or break my business this quarter.

Exactly. Yeah. So like all the max plans, people still don't care that much. Um, we have some startups spending like $5-10K per engineer per month. We're in that bucket, but like startups still don't care that much. So it's mostly like enterprises that we're targeting. Um, we work with like developer experience or developer productivity teams. They're kind of owning this now, the like optimization layer.

Sure, sure.

Is it fair to say that something like 15 minutes could save you up to 20% on your token bill? Absolutely.

What's that? Is that Geico or something? Yeah. Geico.

Sorry, bad joke. Bad joke. Couldn't help myself.

Well, so what is your— what does your actual team look like? Are you building in San Francisco post-YC?

Ours? It's the only way for what we're doing. Yeah, we're all in person. 6, you know, 996 kind of feel here. Team's really fired up. They all want to join the video, but we weren't sure how far to push this.

They can come say, they can just come say hi if they're hanging out. Why not? But what, 99s, it's not even enough. You're not even that differentiated if you're doing 996 these days, I feel like, right?

Sure, there's an opportunity.

You've got companies that are putting in 7, sometimes 8.

8 days a week. Yeah. There you go.

There you go. That's why you have to jump out of a plane to attract talent. This is the best part of the video is we've had like an influx of people that are interested, I think, just 'cause I did my fears of acting was scarier than the jumping out of a plane.

Yeah, I talked to, I heard one entrepreneur who said he changed and manipulated time. He has, every day is 3 days. Oh yeah. 'Cause his first day goes through 6 AM to noon. His second day goes from noon to 6 PM. Then he works a third day from 6 PM.

Stack that up over, over a day.

Yeah, he's gonna crush you.

Or a month. Yeah, yeah.

And he's not crazy for saying it, you're crazy.

Anyway, sorry, we're having too much fun. It feels, feel like a Friday over here. Um, what are you, where are you, where are you going from here? Are you going to keep the team small or—

yes, there's about 16 of us. We are hiring very aggressively. Yeah, we need some, like, incredible engineers and AEs to just scale up. We are swamped. I feel bad. Our kind of, like, campaign to our account executive and to our go-to-market team is death to Jerry. Jerry's our AE, and he is too busy with too many meetings. So we want to make him as tired as possible. And now we need a second.

Yeah. Is there going to be, like, long-tail demand for fine-tuning of open-source models from your customer base? Like if they're doing something really routine, typically that's been the domain of like the Neo Labs. And I don't know if that's the right term for your company, but it feels like you could slide there over time. Is that an ambition? Absolutely.

So actually we already have like an early version of this where you can give feedback in the router ship to the router based on what model you're choosing or what model it chooses. You just easily pass feedback and it's going to learn from this. So we do this both as like a company, we have our own built-in, but now we're exploring what this looks like at a per-org basis. Yeah.

So if the company is just printing money, they can be like, when I ask for the weather, I do want you to use the Frontier intelligence for that. No, that's probably the default thing is like, you know, people are just hammering off prompts for everything. Routing around the obvious solutions is the benefit. Makes a ton of sense. Well, congratulations. Uh, don't break that gong all too soon. You got to warm them up. We've learned this the hard way. Every time we, we hit our gong, somebody comments and says you didn't warm it up.

Warm it up. You didn't warm it up.

But so far, may the gong so good for us ring loud and often for your business. Good luck, and thank you so much for coming on the show to break it down.

Great to meet you. Have a great rest of your day.

Let me tell you about Shopify. Shopify is the commerce platform that grows with your business, lets you sell in seconds online, in-store, on mobile, on social, on marketplaces, and now with agents. Big news out of the Fed. The Fed holds rate steady, but 3 officials voted for a rate increase. The rate has been unchanged in this July meeting. The split shows pressure building inside the central bank to act on inflation 2 months after Chairman Kevin Walsh took over and the Dow skids 1,100 points as investors worry about Fed next steps. They're thinking that there might be a rate hike in the— in reaction to inflation, of course, somewhat driven by energy prices, oil prices. Oil is spiking again. And so lots of tricky news for the global economy. But fortunately, the economy chugs on and the The economy is strong.

So I believe in it.

I believe in it too. Never, never blackpill. Dwarkesh Patel, certainly not blackpilling. Gavin Baker, uh, quote tweeted Dwarkesh's, uh, post that predicted, uh, LAB revenues in the trillions very shortly. Uh, Gavin Baker says, I see a series of fairly wild thoughts from Dwarkesh. I thought I was bullish, but even I am not assuming that the price to rent compute continues to go up. I will say that I am in Silicon Valley this week and Dwarkesh is capturing the zeitgeist. Net-net, public markets would probably be trading differently if they saw the OpenAI, Anthropic, Groq, Cursor, and open-source numbers over the last 6 weeks because everything is growing. Even in— there's so much horse race around, oh, open-source is going to come and grow. And it's like, well, sometimes open-source comes and grows and everything else grows too. And that's That's what we've seen throughout the whole Deepseek saga when the lab revenues grew a ton, even though open source was big. Yeah, there's a comment here that I think is on the nose.

Please. J Capital says, timeline is very interesting. All private AI companies have their best 6 weeks while retail AI/seminars have their worst 6 weeks. Very, very interesting. So disconnect there.

Yeah. I mean, one of them, SK Hynix, $64 billion in quarterly profit. It was a disappointment for chip investors in SK Hynix. Fell 10% despite record results powered by the AI boom. The profit increase was incredible. I think they, they, they grew profit over 500%, something like that. And still the stock traded down because expectations are so high. Everything's priced to perfection. In other news, HBO Max is launching HBO Max Shorts for those who don't don't have the attention span to watch an episode of TV. Uh, it's a mobile feature designed for users to watch iconic and memorial scenes from the HBO library in a vertical and scrollable feed. This is already happening on Instagram and TikTok. Uh, you see a lot of movie clips, just great moments, and it actually is great content. So I can understand why they're trying to own a piece of that.

You've got to get the Jeffrey Katzenberg apology form.

Oh, for sure. Yeah, for sure. Uh, but yeah, uh, High Yield Harry captures it pretty well, watching society become too cooked even to watch TV episode. And this is from Idiocracy, Mike Judge's one of his best works, the creator of still HBO Silicon Valley actually, who will now be available in short form because you will be able to watch short clips from HBO Silicon Valley on HBO Max Shorts. You'll be able to see the— how do you say it? I don't know, all the iconic moments from Silicon Valley when they're hanging out on the roof of Google or whatever the Google equivalent is, Gavin Belsen's company. Anyway, thank you so much for tuning into TBPn. We will see you tomorrow at 11:00 AM Pacific. Leave us 5 stars on Apple Podcasts and Spotify and sign up for our newsletter at tbpn.com. We'll see you tomorrow. Love you. Goodbye. That's the outro. Goodbye.