I see a large IP on the horizon.
You're surrounded by journalists. Hold your position.
Overnight success.
The laser 3000.
Double leg.
That's misinformation.
Clearing order inbound.
Let's just roll.
We are surrounded by journalists. Hold your position. Bump it up.
Trust the extra five.
We are extra founder.
Very good.
I see multiple journalists on the horizon. Stand by.
UAV online.
Blaze. Double blaze. Triple glaze. Double kill.
Cop is wait.
Team Deathmatch.
We are experts. Triple blades. Let's just roll. Right.
Marky clearing order inbound. Come on, get up.
We are surrounded by journalists. Hold your position.
Strike two.
Activate golden retriever mode.
Mark clearing order inbound. Five.
I see multiple journalists on the horizon. Stand by.
Friendly IP in triple glaze. Another one.
Yeah.
You're watching TBPN. Today's Monday, June 29th, 2026. We are live from the TBPN Ultradome, the temple of technology, the fortress of finance, the capital of capital.
Tell you about ramp.com. Time is money. Save both. Easy to use corporate cards, bill payments, accounting, and a whole lot more.
Lot more.
All in one place. I got to adjust my IM.
Well, on the front of the Wall Street Journal today, this is how you know this is the whole AI 2027 Washington waking up. The AI stories are making it to the front page, the world news section, not just the business and finance section more and more. So the very front page of the Wall Street Journal, of course the picture is about the heatwave, but the lead, the story with the largest text is about artificial intelligence. China resets the AI race with the United States as security models mark gains. We're going to get into it. This is a fascinating debate because I thought that we'd have a conclusion to the open source AI debate by now. Either the frontier would have collapsed and there would be, you know, perfect commoditization, or they would have fallen.
A conclusion, it'll just go, it's over, we're so back, it's over.
If you're in open source AI, that's exactly how it feels. Before we get into the story completely, Hill in the chat said, did you see the US national design studio open sourced a privacy model? We did.
And we got him coming on the show today.
In just 45 minutes.
At 11:45, I are going to be talking about a first iteration on-device PII redaction model that is far smaller than existing models.
It's actually tiny. It's 15 megs, which is, and you can do it in the browser.
And we have Chad Rigetti. He's coming on to talk about a whole lot of quantum mumbo jumbo.
We'll see what's going on there.
And Pim's coming back on from General Intuition. And we got a bunch more founders coming on. Jakob Diepenbrock announcing a $30 million oversubscribed fund with tons of TBPN guests already in the portfolio. The rest of the portfolio soon to be on the show, I'm sure. Anyway, open source AI. So the big story is centering around GLM 5.2 from Z.AI. Was officially released June 13th. So it's taken a couple weeks for it to really break through to the front page of the Wall Street Journal, but there's been some strong performance on benchmarks, some positive reviews from developers. I have a whole review from Tyler we can go through in a little bit, but we're now entering another round of debates around open source AI, what can the model actually do? Is this a threat to national security? What are the geopolitical ramifications here? And so I'm sure this will be an ongoing conversation throughout this week, probably next week. We have some guests lined up to help contextualize it. But laying down the facts from the journal, security researchers said that a new AI model released this month by China's Z.AI, also known as Z.A.I., can match the latest US models when it comes to finding security bugs, a development poised to reset the global tech race and pressure the White House in its overhaul of US AI policy. So unlike models from Anthropic or OpenAI, Zhipu's GLM 5.2 is open weight. You can just download it, run it anywhere. You don't need to go to an API. You don't need to go to a private company and pay them. You can run it on your own server, provided you have the electricity and GPUs to do so. It is expensive to run, as we'll go into, but it is open weight. The Wall Street Journal says that means it can be downloaded, run on hardware operated by anybody, and can be modified and used without supervision. Scary stuff. Open weight models are ideal for users who want unfettered access to systems they control, but they're also ideal for hackers who want to run them in the shadows.
Unfettered intelligence.
Unfettered. Oh, that's completely out. We were completely out of names for new Neolabs.
That's a good Neolab name. Yeah.
Unfettered intelligence. Intelligence is good. GLM 5.2 has ranked as one of the top 10 most used AI models according to data from OpenRouter, a company that provides access to more than 400 AI models. And what a fantastic business. Alex Atallah over there absolutely cooking at OpenRouter. It's such an exciting way to plug into the AI race without actually needing to play the benchmark game so much, be the front door. Anyway, in some benchmarking tests, according to cybersecurity company Semgrep, GLM 5.2 bested Anthropic's Claude Opus 4.8 model, which was released in May. When given further instructions, Opus 4.8 and GLM 5.2 can match Mythos in bug-finding ability, according to researchers. So prior to this launch, and there's a chart that we should pull up here about overall AI capability. We can talk to Tyler about what this chart actually means, but there was this narrative brewing that open source AI was slowing down relative to the closed source frontier. And I saw a lot of American AI fans sort of cheer for this. Hey, we have the capital markets, we have the data centers, we have the researchers. And so we are able to push the frontier at a different rate. And if we're actually growing at a faster rate in America within the closed source labs, that will compound and there will be a stronger takeoff in the American closed source AI industry. Now this chart sort of goes back and forth and there's some debate over it. It's in the newsletter. You can go sign up at tbpn.com. While we're pulling that up, 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. So this chart, which we can pull up, shows progress from GPT-4o to o1, o3 mini, o3, Opus 4, GPT-5, 5.2, Opus 4.6, GPT-5.4, GPT-5.5, showing a linear trend in this ELO, which is a blend.
It says GLM 5.2 sounds too much like a gray market peptide thing.
It actually does. It does sound a lot like that. And then you can see the red line are the Chinese models which are also improving over time but at a slightly lower rate. And so the question was, are they going to plateau while America's progress continues to advance? And this latest model, GLM 5.2, it seems very hard to apply it to this particular benchmark because this ELO was... Can you give us some background, Tyler, on where this chart came from, what this is demonstrating?
Yeah, so this is by CAISI, I think that's how you pronounce it, the Center for AI Standards and Innovation. They have this way to calculate the ELO of a model. It's basically a kind of approximation of a bunch of different benchmarks. Some of those are proprietary, like they're not open, so it's actually hard to run these.
Also because I was basically trying to bench like all the recent models since this was published. It was, I want to say, May 1st.
Yeah, it'd be great to throw 5.6, Soul, Mythos, and Fable. It would be great to just continue this chart because it's an interesting trend.
So a lot of those benchmarks aren't actually public. So it's very hard to estimate. But I tried. You can look at some of the benchmarks that are public that you can reference. You can kind of match them up to previous models.
5.2 looks like it is a big step up from the Chinese trend line, right?
But even then, I think it's hard.
I think that the group of benchmarks that were chosen for this ELO definitely accentuate the gap between US and Chinese labs.
I think there's a bunch of other groups, like Epoch AI has done a chart. They've basically seen a relatively stable gap between closed source and open source models.
Since like 2023, like a long time.
Yeah. And perhaps at this point the discussion should be more centered around cost per task more than cost per token.
Yes. Yeah. Definitely. Because even like, you know, new models a lot of times when they come out, like okay maybe the token price is actually the exact same but the token efficiency is much better. Then when you do a lot of these tasks, it's not the price per token, it's price per, you know, something completed and then you actually see it go.
And there's a lot of test time scaling laws where you can just throw a million dollars of compute at a particular problem and all the models do really well at it, but it's completely nonviable for any real enterprise use case and probably not even viable if you're trying to be a nefarious hacker or something.
Yes, most people are feeling like 5.2 is very token hungry, right? So it uses a lot of tokens. So maybe it definitely is much cheaper than the frontier models.
But on the per task basis it might be more expensive.
Yeah. I mean that's still generally not. But on specific tasks you can get, you know, if you have low thinking models, low thinking mode on the closed source ones you can see.
Well, let's revisit John Luddig's post from May 2024. This was pre-DeepSeek, talking about his prediction about why the future of foundation models is closed source. He got a lot of pushback from this because a lot of people like open source models, but he laid out a thesis around closed data, closed source data flywheels, exponential capex intensivity of training. And he said open source will have a home wherever smaller, less capable and configurable models are needed, enterprise workloads for example. But the bulk of the value creation and capture in AI will happen using frontier capabilities. The impulse to release open-source models makes sense as a free marketing strategy and as a path to commoditize complements, but open source model providers will lose the capital expenditure war as open source ROI continues to decline. And that was the thesis around the time that the open source AI discussion was primarily driven by Mark Zuckerberg's work at Meta on the Llama family of models. The idea was that Meta would benefit from attracting talent. It was good marketing. It told the story that Meta has an AI story and has AI talent in-house, even if they weren't monetizing it and sharing a really fast takeoff in ARR around those models. It showed that hey, they're able to develop these models and that might help them cut their costs in the long term. Very interesting that that wound up being very different in 2026 looking at the news today, which we'll go into, about them spending a lot on Gemini. There's been reports about them spending a lot with other closed source frontier labs that they should have commoditized with their open source plans. Nonetheless, that was the idea with Meta. But then China sort of woke up and DeepSeek's launch at the start of 2025, and the game theory became way more complicated. So George Hotz sort of summed this up nicely. he has a take in 'AI will be massively deflationary,' a post from just a few weeks ago as to why China benefits from investing in open source more than American firms. He says this explains why the Chinese are giving the much more moderate resources to train models away for free. They love to see deflationary economics in the US. It is much less of a service-based economy. And so if they can go and give away free tools that deflate the value of the service sector, that is an advantage to the Chinese economy. In his formulation, he says even if you don't regulatory capture the US government, nobody is getting a monopoly on AI. We don't live in a unipolar world anymore. And so he compares what's happening in, he likens what's happening in DC to sort of rearranging deck chairs on the Titanic. It's a very fun piece. But so we're back to this discussion of what are the consequences and the impacts of open source models, particularly in the United States. And there's been this clip that's resurfacing from Dario Amodei when he was testifying in front of Congress in 2023. And it's now recirculating and it was reposted like he just said it and he did not. So be clear about that. This is from three years ago, but some of his predictions were very prescient as of where the frontier is today. So he said, 'I'm very concerned about where things are going. If we talk about two to three years for the frontier models for the biorisks,' is sort of a bad transcription of what he was saying. But he's talking about 2025, 2026. Remember he was saying this in 2023. We're there now. 'I think the path that things are going in terms of the scaling of the open source models, I think it's going down a very dangerous path. And again, if the path continues, I think we could get to a very dangerous place.' He was worried about cybersecurity and bio risks being open sourced and then not having a counterweight to that. Now the good news is that we've talked to the CEOs of cybersecurity firms like CrowdStrike and Palo Alto Networks and they've been working with Mythos and GPT-5.5 Cyber for months now to harden systems from LLM-driven attacks. And so there's still this gap between closed source and open source models and that gap allows white hat hackers to implement fixes before black hat hackers have a chance to exploit easy bugs. There still will be a bigger discussion here though in DC over the next few months as the frontier models roll out and the gap doesn't appear to be widening at the moment. So security stances must adjust. It's not that closed source is falling behind, so it's never going to be an issue. There will be this gap and how the American cybersecurity industry and eventually the biosecurity industry implements changes and fixes before open source catches up or commoditizes and makes that particular capability widely available is going to continue to be important. So let's go over to Tyler's quick review of GLM 5.2. Why don't you take me through your bullet points that we shared in the newsletter at tbpn.com and you can tell us like what is the shape of this model? How are the reviews?
Yeah. So I think so far one of the main things is like people are saying it's, oh it's distilled, right? This has been a big thing with a lot of these open source models, especially the Chinese ones. Oh, the only reason that they're good is because they're distilled. It's very hard to actually figure out how true this is.
But people are, you know, it certainly seems like there's some, you know,
I think aspects of Anthropic models. Didn't Anthropic openly accuse Alibaba of distilling a number of these labs?
Yeah. And there's also been a big professionalization of the gray market where a whole bunch of different sort of individual groups will connect a whole bunch of different entities and user accounts and script subscriptions and APIs to then create a front end to like the model that can be served at a very high rate through a VPN most likely. What's interesting is that you'd think that if you were going to do a training run, you would just find and replace on the other lab's name before you hit run. Is that not something people can do? I don't understand.
Yeah. I mean, it also depends on what you're actually like, maybe you're not directly distilling on the API, but you know, you're training on public GitHub repos and those were all used, those were all made with open source models. You're kind of like distilling but it's not really count as distilling, I don't know.
But so if you are like, if you're convinced that these are like super distilled, the only reason that they're good is because they're just, you know, basically taking the closed source labs.
There's also this weird thing with distilling where as more and more of the public internet and GitHub broadly and open source repos become LLM outputs, if you train on that you are in some ways distilling because an LLM has a quirk like it's not this, it's that in text and you wind up training on a whole bunch of Amazon Kindle books. You're going to wind up learning it's not this, it's that. And the same thing applies for different code conventions in open source repos that have effectively been completely rewritten by closed source models.
Yeah. And so I think it's safe to say that like we've generally seen that distilled models generally will generalize worse, right? So you'll see really good benchmark scores.
Maybe they're benchmarks, maybe they're not. But even if they're not like directly benchmarks, you still find that they generally
Yeah. They're kind of accidentally benchmaxing.
Yeah. Yeah.
So I think initially you should just be a little bit suspicious of these super high benchmark scores.
Yeah.
But they lack that big model gen qua.
Yeah. And this is like anecdotally reinforced. A bunch of people have been saying, you know, for coding, these models are really great. GLM is a very good model, you know, for creative writing or something like this where you'd imagine it's a bit harder to kind of benchmax this.
They'll perform a bit worse.
Yeah. I wonder, have people been testing it with the like Tiananmen Square bench? Like does it reject that stuff? Because it felt like that was something that was widely misunderstood by American audiences that in fact that might not be the biggest deal for the CCP.
Yeah. So I think, you know, even if that's true, like the model is open source, you can kind of just fine tune it to like not that. Maybe it's a bit harder than that, but I think you can kind of get around like that kind of stuff.
Okay. Yeah. So we talked about the token hunger and the API price. And in general, I mean you said I'm not convinced that there's a big market for this class of model especially as frontier models get more efficient. And if you look at OpenRouter, the most used models are the smallest open source models presumably being used for specific tasks that need to be repeated over and over again. So
Yeah, so I think like what we've seen is
Like a marginal IQ point of the models is like extremely expensive.
Frontier models are getting very expensive. People have to cut back. You know, they're token maxing. This is like massive bill on their balance sheet, whatever. I think like it seems like there's now basically like two classes of models that people really use. There's like the frontier ones
and they're using coding agents. They need the best thing. If you're doing cyber, you just need the best model because you know the risk of someone hacking you is so great. You just need the best thing, you pay whatever it is.
And then there's the second class which is like these very small, very fast, very cheap models that you can use for these kind of point solution things. Maybe you have some orchestration where using a really big model to have these like little agents using these very cheap models.
I think in the middle it's hard to actually figure out what is the real use case.
Maybe it's like hobbyists using these coding agents and they don't want to pay the super expensive tokens of the closed source labs.
But generally, and you see this on OpenRouter where like what are the top models by token usage, it's these very small models. It's like, you know, DeepSeek Flash.
You're just spamming them for like, you know, every receipt that goes into Ramp gets processed by an LLM at this point. Does it need to be a frontier model telling me that I spent $10 on a coffee? No. It can just do standard OCR.
Yeah, you want super intelligence overseeing your expenses most likely. But no, you use the right tool for the job and that's clearly what's happening on
But also I think like it is a very good model, right? Like we should not fully dismiss
I think the idea that oh the gap is widening, we really don't have to worry about these open models. I think they are like very good.
And maybe if you're super worried about distillation, maybe something changes if the models are kept to these big partners, right? Like what we've seen recently with government coming in. But
I think we can't really fully dismiss these labs.
Yeah. It throws a little bit of a wrench in the monetization potential, like how long can you monetize a new frontier model, that's more tricky. And then the other one is just like if you're going to keep a model behind KYC or behind an approval for specific companies like the government has been sort of edging towards and moving towards. It gets a little bit tricky if all of a sudden you just wait three months and oh, I was waiting to get approved for this one for like GPT-7 or whatever, but by the time the government got back to me, my company got access to GLM 6 and it's close enough. And so that just throws another wrench that I think the government will have to figure out how it puzzles together with the rest of the strategy, which has been yeah back and forth as always. Anyway, let me tell you about Shopify. Shopify is the commerce platform that grows with your business and lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents.
Google caps Meta's Gemini use as AI demand strains capacity in the Financial Times. Surging appetite for advanced models is turning computing power into the tech industry's scarcest commodity. And they have a picture here of a Google Gemini bicycle which looks fantastic.
What does that have to do with Meta though?
I think that was just the best Gemini picture.
It is hard. Otherwise it's just a picture of a phone screen. I mean you saw in the Z.AI, it's just a picture of the app, which is like so boring.
Or it's the stock image of the brain with the neurons. That's always good.
What do you think? I mean this kind of ad placement
Like on a, what do you actually call this? It blocks the water. No, no, no. That just the part that blocks water from, you know, if you were to ride through.
Yeah. Some type of fender thing.
Like a Mansory kit for a city bike.
Exactly. Exactly. But imagine riding that Gemini bicycle in the rain. Fantastic.
That's what it's for. So the water doesn't come up and splat you. Interesting. Okay. I never knew what that was for.
You never knew that. No. I learned this is
This the experience of hosting the show is educational for both of us.
It is.
Google has put limits on Meta's use of its Gemini AI models after the social media giant sought more computing capacity than the rival tech group could provide in the latest evidence of the infrastructure constraints facing even the world's largest AI providers. Google told Meta around March that it could not provide all the Gemini capacity the company wanted to purchase, according to three people familiar with the matter, in a move that has disrupted and delayed some of Meta's internal AI projects. So
I don't understand how much this is possible.
Google spent $200 billion on capex.
Okay. So of course like this
Around this time token maxing was becoming a thing. Every company in the world, at least every tech company in the world's kind of going a little bit crazy from a spending standpoint.
And so, you know, I could see Meta going and like wanting to basically buy a bunch of capacity and then being told like, hey, we can't fulfill that.
But I'm wondering how much more we should read it, like read, like is it worth reading? I mean it sounds extremely bullish for Google, like if they're, this tracks with what they talk about on earnings calls.
Yeah. Yeah.
Yeah, Google Cloud acceleration crazy.
You do have to wonder like could distillation be part of this story? Could that be a factor here? I have no idea.
I don't know. Zero Hedge said Meta puts limits on Claude and Codex fearing distillation, The Information.
But so this story is different. This is Meta telling its own employees, don't use Claude and Codex in certain parts of our business because we don't want to accidentally do distillation is what Meta is saying. So that's different. I was wondering like is Google thinking like whoa that's a lot of, you know, cool it, you know.
Yeah. Owing to the restrictions which remain in place as well as broader push to streamline AI costs, Meta has encouraged staff to be more efficient with AI tokens. Several other Google clients have been affected by the restrictions although to a lesser extent. Meta has been particularly impacted because of its exceptionally high demand for Google's models.
Very interesting. I would love to see a pie chart like breaking down what all the different ways they're using Gemini.
In their business because
Google has not broken out Gemini revenue.
At all to date. So we have no idea what percentage of their AI revenue is like actually spent on Gemini
versus other, like the Gemini tokens broadly go into AI search overviews. So that's a search product, probably insane token demand there, right? You've seen the chart of like they're in the quintillion or quadrillion tokens category. And then you have YouTube now has Gemini plugged in and you can chat with any video and transcribe it. That's got to be incredibly token heavy. And then you have Gemini app users and free users and paid users. So there's got to be a lot of just Gemini internal usage, but it's remarkable. Yeah, I would love to see that Meta pie chart because I thought that they were spending a ton with Anthropic. I thought they were spending a ton with Google, but I also assumed that they would be running a bunch of Llama workloads and a bunch of Muse Spark workloads because those models have performed well at various points in time. And if you go into the Meta app, you now have access to Muse Spark. And if you go into Instagram and you search for something, it populates it with a Llama, like Llama 4 result. And so I would imagine that even though that product has not broken through like crazy, I would imagine that it's still generating a lot of tokens just because of the scale of Instagram. Like Instagram has a billion, two billion users, something like that. It's huge. And so even if it's people sort of, you know, accidentally winding up in an LLM powered workflow, it probably is generating a lot of tokens just because of the scale of that system. On the topic of Meta. Yeah.
Meta shared this morning.
Non-invasive brain decoder research. Brain2Qwerty v2, building on v1 which was published today in Nature. Brain2Qwerty v2 is the highest performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. Advances beyond character level performance to decoding words and semantics, enabling accuracy for overall communication. So if you thought, you know, Instagram was listening to you, if you thought I was listening to your, you know, conversations, now you can have a, you know, new conspiracy at home, which is that they might be just listening to your thoughts. Do you know the device? They say this is a non-invasive device. I just shared an image of this device and I want you to tell me, do you consider this non-invasive or invasive?
Look at this image of the magnetoencephalography device.
No, you got to go. You need to scroll up a little bit because you can't even see the whole thing here. Scroll up. It's not invasive
because it looks like the device could actually potentially carry on for like a whole half of football.
It really does seem like it's a just put yourself in this room-sized device. No, of course this will shrink.
I'm giving him credit here. Non-invasive.
You're putting this thing on. You're daily driving this thing.
I don't know if I'm ready to daily it.
I don't know if I'm ready to daily it. Yeah,
This will be a cool demo. Like this will actually when
When you can just walk in, sit down in a chair and see your thoughts on a screen.
No, we were debating it earlier. My buddy Rob Taves has been on the show twice. Dropped five predictions in Forbes recently. We can go through them at some point. He's going to come on the show. But four of the five were very, very like reasonable. You know, Anthropic's going to be bigger and, you know, TSMC is going to face more competition. And then he predicts that in 2030 telepathy will be commonplace, which is a very aggressive prediction in my estimation. You know, it's certainly not like a straight trend line since like, you know, TSMC has competitors right now. The prediction is just that there will be more competition. But truthfully, telepathy is not really existent outside of like a few demos like this. It's not really something where it's like, oh yeah, like 5% of people have the Meta Ray-Bans that take pictures. So like face cameras are going to be bigger in five years. And it's actually only three and a half years until 2030, which is sort of crazy to say, but we are getting quickly to the future. To the future. Never sell your company. Should you ever sell your company? David Sacks says no. He says the best founders in the world would never sell their company. You could never acquire Elon, Bezos, Zuck, Jobs, Ellison, Jensen, Dell, Page, and Brin. Scott Wu has turned down billions and keep saying no. This is a great clip. Went super viral. I don't know, did I lose internet or something? I don't know. Anyway,
Tyler's app is in shambles.
I don't know about that. But there's some debate over this because Elon
Elon, yeah, this is just Elon did sell two companies. He sold Zip2 and he also sold PayPal. And then Jobs sold NeXT back to Apple. Does that count? I don't know. He did sell Pixar to Disney. That sort of counts. And I mean Elon never sells his company. He just sold XAI to himself, but I guess that doesn't count. But yes, it is a funny thing. Didn't Mark... Yeah. So what's the, do you know the backstory here from Sasha?
I don't know. Tyler looked it up. Apparently there's a Business Insider report from the time that this happened in 2007. How Terry Semel fumbled Yahoo's Facebook deal. How much is Facebook worth? 5 billion, 10 billion, 15 billion, whatever the number, it's probably a lot more than the 1 billion that Yahoo could have bought it for a year ago. As Yahoo continues its soul-searching, here's an unpleasant rendition of Semel's catastrophic decision, courtesy of Wired. When Yahoo came calling with a bid of 1 billion in cash, the pressure became too much. Zuck relented in July of 2006. He was just like 18 months into building the company, something like that, verbally agreeing to sell Facebook to Yahoo. He said yes. He said he was going to sell Facebook to Yahoo, allegedly. Strategically it seemed like a good match. Yahoo had hundreds of millions of users but its foray into social networking was struggling. Facebook had cool tools and was looking for a mass audience. The timing, however, could not have been worse. In the days after Zuckerberg agreed to sell, Yahoo announced it was projecting slower sales and earnings growth and that the launch of its new advertising platform would be delayed. Its stock price tumbled 22% overnight. Terry Semel, Yahoo's CEO at the time, reacted by cutting his offer from 1 billion to 800 million. He just took 20% off, but Zuckerberg, who had been warned about Semel's reputation for last-minute renegotiations, walked away. And that's probably reasonable. I mean, if they're cutting the price there, you have to imagine that as it gets papered, you get cut down again. Then the earnout, you get cut down again, and all of a sudden, you're walking away with barely anything. But two months later, Semel reissued the original $1 billion bid. But by then, Zuckerberg had convinced his board and executive team that Yahoo wasn't a serious partner and that Facebook would be worth more on its own. He rejected the offer and became famous as the cocky youngster who turned down $1 billion, from Wired.
Legendary. Legendary. It's so interesting to imagine the road not traveled there because the dynamic, the way Facebook is built as a social network, like could it have been successful under Yahoo's stewardship or would it have been less exciting, attract less talent, ultimately been disrupted and would they have had the capital and the guts to go buy WhatsApp and then also buy Instagram? You know, to actually maintain the dominant position in social networking. What do you think?
I think Yahoo should make another offer. We were hanging out with Jim, CEO, last week, dear friend of ours, and
I would like to see Yahoo make another bid.
Hey, Meta's trading down. Just keeps going. If it continues at this point, it goes down to 99.99% might be able to pick up if it continues at this trend.
Anyway, let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI. Own the data platform that powers it. Moving on. What else is in the news?
Chipmakers are profiting off AI at the expense of just about everyone.
This is on the cover of the business and finance section today. We are witnessing an extraordinary transfer of cash from the providers of AI and perhaps one day AI users to memory chip makers. Take us away, John.
Yeah, the explosive growth in Micron Technologies' profit in the latest quarter is extraordinarily good news for its shareholders, but it comes at the expense of the artificial intelligence companies to which it sells fast memory chips. Micron, along with Korea's Samsung Electronics and SK Hynix, are to AI what oil producers are to the airlines, makers of an essential input that this year suddenly became much more pricey because there is extremely limited capacity to make the high bandwidth memory that AI needs and it takes years to build production facilities. Soaring data center demand simply jacked up prices. Micron's soaring profits are for its customers, soaring costs. We are witnessing an enormous transfer of cash. They said profit shifts of this scale are rare events and
Investors should be paying attention to where the money is coming from, where it's being spent, and how long it will keep flowing. In the quarter ended May 28th, Micron increased prices for DRAM chips more than 60% on the previous three months while increasing shipments by a low single-digit percentage. It said last week prices for NAND flash memory, also used in data centers, jumped more than 80%. Usually memory doesn't matter that much, but for Micron, customers paid $18 billion more, and that was just in the quarter. Prices quadrupled in a year, and it's hurting outside AI too. Apple last week raised prices for MacBooks more than 15%.
Closer to home, uh, closer to home for me, the memory I bought on Amazon.com a year ago to build a super quiet computer. I hate fan noise. Good, good color commentary here. Uh, has tripled in price and now costs more than the CPU.
For an industry in which prices usually drop every year, it's a huge turnaround in consumer electronics. Passing on higher prices helps limit demand for chips, just as higher oil prices reduce consumption. But the AI companies aren't passing on higher prices because they are able to throw money at supply problems. The problem in AI is that the end users aren't covering the cost of the service, with big losses being recorded by AI model producers. Everything is still priced to bring in new customers, not yet to make money. So higher input costs create a nasty problem: either losses will be bigger, or higher prices will be needed, putting off potential customers.
And you can see the price of Micron's stock has been through the roof as the company joins the $1 trillion club and becomes the first trillion-dollar company headquartered in Boise, Idaho. And Idaho got a trillion-dollar company before New York, I believe, and also before Florida and Austin, maybe something like that. Um, it's rare. It's rare.
Mostly on the West Coast.
Um, anyway, there's a whole bunch of bull cases for Micron still. The stock could double from here, says Barron's. I love Adam Levitin and Barron's sharing the bull case. Tyler, how many trillion-dollar companies are there in Europe? Out of curiosity.
I'm going to go with zero.
That's true. That is true. You are correct.
The other ASML could get there, maybe, sitting at around 700.
Wait. What about Eli Lilly? Or no, Novo. Novo was a trillion, right? Oh, did it ever touch a trillion? I don't think so. Right. It was real close.
It's a humble 165 billion.
Brutal. Wait, but you're thinking of Eli Lilly. Oh, Eli Lilly hit. Yeah, rough. Very, very rough. Comcast is planning to split up the company. Competition is escalating.
Eli Lilly, the Indiana company. John, is it from Indiana?
Okay. I like it. I like it. NBCUniversal and Sky will separate the company's connectivity business from its film, theme park, and streaming operations. Oh, yeah. Universal Studios. Comcast is up on the news. Comcast plans to separate its media and connectivity business.
Who's building the Anduril of theme parks? It does seem like there could be an opportunity to create a net-new theme park business with a modern technology stack.
It's very expensive. Everything needs to be like the modern technology stack in parks is expensive. You don't believe in the theme park capital markets?
I don't know. I know people that have worked on theme parks at Disney, and it's tricky because you have to amortize a ride over like 20 years, and so you'll go... seems like an absolutely brutal business.
Yeah. That is probably harder today because at the time that a lot of these parks were built, you didn't have infinite online entertainment for every single sub-niche instantly available.
I mean, there's a whole bunch of trend pieces right now about how IRL experiences are seeing higher than ever pricing in the face of you could just watch the Knicks game on TikTok highlights, but people still forked over $5,000 to go see the game. And so you have that barbell strategy where Thrive is buying a stake in the San Francisco Giants, a baseball team that should face...
Exploring the NBA team to Vegas. But at the same time, there is enough stat that came out this morning or maybe yesterday that there's more sports betting volume than all sales of movie tickets, theaters, theme parks, and a couple other of these IRL categories.
Uh, less. Lower. And the stat was like volume. Yeah. And so it's not exactly like a proxy for revenue, but still meaningful.
Theme park vertically integrated Tweety Bird tattoos.
Tweety Bird tattoo parlor right on site.
Six Flags. No, I don't. I like it.
Six Flags never really got the same cultural power that Disneyland did. There's something about the flywheel that Walt Disney laid out that does seem very, very important. And so, how do you start that? It's not just, oh, you know, tech-enabled theme park that's not going to draw people in. You need to have IP around it.
There's something about, I mean, we've read stories about the Disneyland fan that saves up every year and spends so much money at the park. And I think that's probably the lifeblood of that business. And that doesn't happen without building a whole cinematic universe around every single ride. And that just takes so much time. And this goes back to the question of like, is Netflix enduring IP? They haven't been able to, even though it's been 20 years of like, I mean, I don't know when they started producing their own content, but it's been 20 years for that business at least, and they haven't really developed their own franchise that lives in the same world as Batman.
Well, I'd push back and say Narcos.
Narcos. You want to go to the Narcos theme park? I was talking about this with somebody once, talking about HBO, like why don't they have a theme park? And he was like, what are you going to do? Take your kids to a brothel in Game of Thrones land? Like no, it doesn't make any sense. Anything like it needs to be uniquely general audience. You can't have R-rated, you can't have a content backbone that's R-rated because theme parks will always attract families and kids. And so you can't have any theme park that's built around an R-rated IP library, and so that just narrows it down even further.
Well, all of America's basically turned into a theme park for European soccer fans.
Oh yeah, in the journal. European soccer fans marvel at the splendor of America's suburbs.
Many of these reels serve to me.
Dutch fans in Missouri see a nation that is risky and expensive but vast and bountiful.
Everything is three times the size.
You've been seeing some of these people in real life, right?
I don't know if I've seen any of them. I did go out to lunch like a week ago and it seemed crowded, but I was unclear if that was just local residents going out to watch the games or actual tourists coming to town to watch.
Gabe, in the X chat, I think Ferrari has a roller coaster in the Middle East. Do they have a whole Ferrari theme park in Abu Dhabi? Because that's not R-rated. You can take your kids to Ferrari theme park.
Yeah. I was in Abu Dhabi and I was driving by it and I was like... Yeah. I was just thinking of like if you wanted to spend a day, you know, getting the Ferrari experience, like you could just go to the track. Or you could just rent a Ferrari. So I don't know.
Yeah, but you don't need to go to Six Flags to get the Batman experience. So you can just go out in the middle of the night and arrest a criminal.
I saw another report that apparently there's an individual who's being like the Batman of Mexico. Do you guys see this? This is very funny. And so the guy went out and found criminals.
Develop says Yas Island. They literally named an island Yas.
Weird. No. Okay. I don't know.
Anyway, Dutch soccer fans are having fun visiting America. Frank Evering, he hadn't even heard of Kansas City. But when the Dutch soccer fanatic saw his team would be playing along the border of Missouri and Kansas, he made a detour in his worldwide road trip. Everyone got in his camper van and drove south from Toronto, making stops in Detroit, Chicago, and Indianapolis. Along the way, he and other European fans who flocked to Kansas City for the World Cup beheld the fruits of the American economy from a vantage point few foreign tourists typically see: suburban superstores, hulking plates of food, quiet streets. He marveled at the sprawling houses, a contrast from the tightly packed homes of the Netherlands. I did notice this when we were in France. The food portions were way too small for me. It was brutal. It's spacious. He said, 'You go here for your shopping and there for your dentist. People are so rich here. I think that's why they can be so nice.' What an ultimate white pill in America. In America, everyone's like, 'And we're so divided and everyone hates each other and it's terrible and the economy is about to fall apart.' And then one European tourist comes like, so nice.
Something about the grass. The grass is always greener, right? The grass is always greener on whatever side I'm on. That's what I like to say.
The throngs of Dutch fans that flooded Kansas City and its suburbs this past week got a taste of day-to-day life in the United States, reigniting the long-running transatlantic debate: who lives better, Americans or Europeans? The Europeans had plenty of thoughts on American culture. 'We are a bit shocked about the food you're eating,' the Dutch national team superfan Sandra Tate said. Fans also balked at the size of Costco and the vastness of the highways. In recent days, social media has been filled with videos of Europeans gawking at the staples of suburban life: a two-car garage, a walk-in closet, a second refrigerator. One Brit went viral for trying Chick-fil-A for the first time. 'That was absolutely banging,' he said. And another toured the inside of an American fire station.
The way they experienced Chick-fil-A was me seeing the Renault Twizy. I was just like, wow, this is unbelievable. They made the perfect car. Yeah. So small. So small.
And it's the way they think about our fire trucks, which are massive. This is nuts, honestly. They said...
Tyler, while we wait for our first guest, do you know anything about Bosnia's World Cup team? The United States is facing them on Wednesday. Do you have a stat breakdown or anything? Be very careful with what you said because I saw that there was a news reporter who faced fierce backlash for really calling Bosnia out and saying like, I don't know where it is on a map. And the funny thing was that it was delivered in like the typical newscaster like, and I'm here reporting on the ground and tonight, but Bosnia will be playing, and then she just transitions into color commentary giving hot takes about how irrelevant Bosnia is in her mind, and the Bosnians did not enjoy her critique of their country.
Anyway, there's a little golf cart. We got to talk about this at some point, but there's a new car. It's like a Twizy. You're going to love it. It's close. It's no Twizy for you. Let's bring in our first guest from the National Design Studio. Welcome to the show, gentlemen. How are you doing? Thank you so much for coming on the show. Please start with an introduction of yourselves, the company, and then the announcement today.
My name is Edward Gorberstein. I run engineering at National Design Studio. It's technically not a company. It's a government organization.
Oh yeah, that's right. Sorry.
And I'm Tiger. I'm one of the engineers at National Design Studio.
Okay. And today the launch, take us through it.
We're launching Rampart. It's a local-first privacy model that puts people back in control of the data that they share with AI.
We were just like, you know, kind of building a chatbot for fun. And we were upset that none of the frontier models will actually fit in a browser. So you cannot do PII removal in the browser, which is, you know, pretty damn important for our use case. You just have to like, you know, trust that the server is actually removing the information and not lying to you. So we're like, okay, well, what if it was just all on device, like personal data never had to leave your device, it was just, you know, secure by default.
Okay, so open source, the weights are on Hugging Face, runs in the browser, under 15 megs. A technical user could go right now, download the model from Hugging Face, vibe code their own Chrome plugin, and have it be running however they want. But how do you see this actually rolling out? Do you want the government to implement this in various places? Do you want companies to? Is it sort of like open up the primordial soup of ideas and see where it goes, or do you have a rollout strategy that you are advocating for?
Well, the reason why we open sourced it is because we do want companies to use it, and we want people to use it, and we want people to make it better. So we want vibe-coded Chrome plugins. We want, you know, vibe-coded tragedy extensions, like whatever value is derived from the product. You know, this is just like a total side quest for us. We just want to build software that's helpful for the American people. We've already launched a series of products, and like TrumpX has got 15 million users. It's saved over $500 million in drug costs, and you know, we rethought the UX there. So we're, you know, basically across everything we're working on, we're just trying to find the first principles best approach for users.
And this just came as a derivative of that. We're not ML researchers or engineers. We're just like, you know, we should just do this.
You created PII superintelligence. That's what people are saying online.
Basically, it's like tiny intelligence. It's by far the smallest model. Like the other ones are at least 50 megabytes. This is 15.
Yeah. So, to what degree did you build on the shoulders of giants? Is this some pruned and distilled and fine-tuned open-source model? Is this something where it was easier to just start from scratch but use architectures that are more prevalent and well-established? Like, how did you actually go about training this model?
We tried 72 different base models.
And, you know, put them through a training set. We ended up on MiniLM. So we definitely are standing on the shoulders of giants here.
Yeah. We started taking a look at the OpenAI privacy filter that just got released recently. We're trying to figure out, is there a way we can, you know, just quantize it? Can we maybe remove some of the parameters? Like, what can we do here to try to make use of the state-of-the-art model? And, you know, we tried a lot of things. We just could not get it to fit into, you know, we want this to work on legacy devices, on an old Android phone, for example, or an older iOS device. And it just would not get small enough and still make any intelligent sense to try to actually run it. So, yeah, we ended up essentially, it's technically a fine-tune, but we trained MiniLM and basically made it do exactly what we wanted to do.
Can you help me understand use cases a little bit more? Because I feel like most of the time when I'm transmitting a document to a prescription website, Rx, or a financial institution, the PII is potentially the only important part. They're often sending me a blank form and asking me to put my PII in there. What is the inverse scenario where I want to redact my information but I still need to transfer something? Because in most cases that would just be the template or something, in my estimation.
Yeah, the flags are set at compile time. So you can decide, like for our use case, it's really important that we have this data or that we don't have this data.
And so we hand all the customization back to whoever wants to use the library. The model just says, oh, you know, this is a phone number, this is a name, this is a surname, etc., etc. And then ultimately, it's, you know, whatever you want to do with the model, you can just do it. Fundamentally, what we were looking at was there are a lot of cases where people will ask a question pertaining to a document of like, okay, for example, with the template, how do I fill out said template? Because, you know, the government is pretty bad with forms. There's like way too many forms. Nobody knows what they mean. Like, you got to pay people to do government forms. So that was a use case we had in mind. PII is not super helpful for that.
And it's also kind of like the breaking point. It's where the product will lose trust. So we're like, okay, two birds, one stone. Let's build this thing.
How do you guys think about side quests at the National Design Studio in general? Like, I imagine every single day there's opportunities that come up, and you guys are in a unique situation where you sort of design. You have a mandate, but at the same time, there's so many different places that the government interacts with people's lives. I'm very curious.
It's pretty hard to pick what to work on because there's a lot of exciting things. Everything is huge scale. Everything could be way better. Maybe not everything, but a lot of things. So there's a huge calling for side quests. But we just tried to keep everything in line with our vision, which is like we want to make the American digital experience better.
And then we've kind of chosen a track to get there, and on the way we built this model, and on the way we built TrumpX, but we're excited to see how it develops from here.
Back on the show. Yeah, like diving more into that, do you have a reference point in tech? People might ship, you know, they might think in quarters, financial quarters, three-month cycles. They also might think about a two-pizza team, which I think is like 10 people. Do you have an idea of where the sweet spot is from what you've experimented on? How many people do you want to bring into a project, and then how long do you want to spend there so you don't get stuck for a decade? Because you might not have a decade.
Yeah, I mean, there's definitely a lot of places to get stuck because the visibility is super low in a lot of these projects, and you don't know how broken they are until you're really in it. Being able to determine that in advance is definitely AGI level.
We have a really great team. We're very fluid. We're constantly trading responsibilities back and forth. You know, someone might be better at doing one part of the tech stack than somebody else, but they're on a different project. We'll just borrow them for a day, or even for an hour. We share a lot of responsibility at the studio. This is also definitely the only place in the government where people work seven days a week, consumed on Red Bulls. I think the ideal amount of people per project, if they work super hard, is two.
Like one design person, one engineer, and they both have like, you know, full scope, and then they're able to call on people as necessary.
Yeah. Two, with the caveat of you're calling in your coworkers, say, 'Hey, can you take a look at this over my shoulder' quite frequently.
What's your pitch to talent that you might want to recruit into the National Design Studio? I imagine lots of people that would join could get a blank check from a venture fund or could go work at some of the best companies.
Everyone that joins, you know, that's the case for them. Yes. It's definitely more for people who are super mission-oriented. You know, who the hell, like, what great engineer wants to come work in the government? You know, the answer is typically nobody, unless it's, you know, like the IC where there's really interesting problems to solve. So I think that we have a super golden opportunity. At least the way I evaluate problems, I try to see how big the problem is in terms of like how many people will use it, the delta between what exists versus what our team can do, and how fast we can do it. When you look across those three matrices, it's like a home run place to work. So I think that is its own natural kind of calling card for the right kind of talent that we need for the studio.
Complex too. It's also a huge benefit. It's pretty sick.
Where is that? Is that where you guys are right now?
We're not there right now, but we're about to be there. Yeah.
Awesome. All right. Well, thank you so much. Congratulations on a very fun project, and we'll talk to you soon.
Have a good rest of the day.
Goodbye. Thank you very much.
Cheers. Let me tell you about CrowdStrike. Your business is AI. Their business is securing it. CrowdStrike secures AI and stops breaches. So, Apple and Audi alumni just unveiled a $25,000 open-air electric neighborhood vehicle. It's called the Amble 1, and it's a street-legal EV built for short local trips. No doors, fewer screens, modular design inspired by the 1960s Lunar Rover. Goes 40 miles an hour with 60 miles of range. Weighs under 1,000 lbs. Takes 5 hours to charge. Rear seats fold flat for cargo, surfboards, or gear. Built-in mounts let you add baskets, straps, mirrors, and cargo accessories. Already has 500 vehicles committed.
I love it. I think it's great.
Daily, weekly, you know, the Doug score out of 100. What are you doing?
So, I mean, I just went through this whole crazy search for basically this exact vehicle. Didn't find it. I don't like the aesthetics of golf carts. I've driven a lot of carts in a commercial job in college. I've owned a golf cart. In my experience, it's impossible to feel cool while driving a golf cart. So I wanted something like a golf cart that was more like, you know, I'm not golfing. So I wanted some utility, wanted to be fun, etc. I landed on a Can-Am HD11. You know, a UTV. It's gas-powered. It's quite fun, but the gas element is actually kind of annoying, even as an ICE defender, that is the internal combustion engine. But I think this is fantastic, and I saw somewhere that they're going to focus on more commercial opportunities. So going to hotels all over the world.
