U.S. AI Execs Warn as China’s AI Push Grows, Substack Adds AI Detection Tool | Diet TBPN

China, China, China, China is in the news. Top American execs are sounding alarms. The White House is divided on how to respond to recent advances in Chinese AI, has weighed crackdown measures. We talked about how this was coming. It is interesting. There's a lot of like hurry up and pedal to the metal and then pull back and then ban the American model and then are we gonna ban the Chinese model in a similar way? And lots of contradictions, lots of insider baseball. Everyone's critiquing everyone else. Lots of interesting takes. Also just some interesting data about what's actually going on with these models. How advanced are they? Where do they come from? How cheap are they to make? What are the economic impacts in America? But the Wall Street Journal has a nice little roundup here we can take you through. So Silicon Valley and Washington are debating a multibillion-dollar question. Should American companies be able to use Chinese artificial, artificial intelligence models. They certainly have been able to, hasn't been a problem for years, but now they're getting stronger. OpenAI and Anthropic executives are sounding the alarm about the rise of cheap AI, particularly powerful new models produced in China, suggesting that they will lead to, quote, a dystopian AI future and present unacceptable security risks without regulation. And that is the point of debate. Are you on the side of dystopian open source or liberating open source, the George Hobbes frame. Some analysis, some analysts who study the AI industry say the two companies which are preparing for public listings in the next year, they just want to eliminate the competition. The emergence of highly capable open autonomous AI systems including Moonshot AI's KIMI K3 model and Alibaba's Qwen 3.8 Max. And Alibaba is actually a massive investor in Moonshot. They have their own Anthropic Google over there going on. Alibaba is like very much the Google to, uh, uh, in the sense that they, they have a large position there, uh, which were released in recent days and viewed favorably by investors and users has turned the AI race on its head once again. KIMI K3 was also competitive with U.S. models on some benchmarks. The debate over the new models, which are quote, open weight, but not yet. They're, they're about to open source them. We'll see what happens next week, allowing users to download and customize them with company data for specific tasks. Very, very good. Very useful. Coincides with division in the Trump administration about whether to take steps to limit the use of models in the United States. Quote, one probable outcome of an open weight model dominant world is full AI communism, which is precisely what China proposes, rather than a market Product. AI is a public good which will ultimately be provided by the state as a kind of digital public infrastructure, Dean Ball, OpenAI's head of strategic futures, said in an X post Friday. Dean Ball made it to the pages of the Wall Street Journal. Now, the AI communism is just at the model level. Like there would still be a bunch of like a lot of people who are long compute, long Neo Cloud. They say like this is where the value should accrue. The data centers will accrue. A reasonable margin and they will still continue to make money. And so it's not that this automatically leads to nationalization of compute in any way. People would still be paid, but the margins might be lower is at least the formulation here. So Ball, a former Trump administration official, called such a scenario a dystopian hellscape. And he said he expected the Trump administration to make moves toward reducing the use of open-source models. Ball later clarified that he wasn't advocating for the United States government to discourage Chinese AI. He also highlighted a central challenge with the AI race. Top-tier AI companies raise billions of dollars to pay for the vast computing resources needed to continue improving AI systems. If everyone uses AI systems that people largely don't pay for, there would be no way to finance continued frontier AI development. Ball, who started at the company earlier this month and focuses on policy and strategy, and OpenAI have said his comments don't reflect the company's official positions. He's just posting. The company has developed its own models, own open models, and Chief Executive Sam Altman had previously said that OpenAI's past approach of only developing closed tools was unwise. Other top AI executives have warned that open models present huge risks since developers and policymakers lack control over how they are used and modified. A lot of risk of cyberattacks, a lot of risk of annoying spam attacks. The cyber thing is very concerning and annoying. If that really ramps up. But the good news is that the closed source models have been available to cybersecurity firms and a lot of big major companies. The finance industry, for example, has had access to Mythos and GPT Cyber for almost 6 months. And then just today, Google announced a specific model, Flashlight Cyber, I think. So they're trying to offer a similar capability, but really go down the cost curve because a lot of the cyber tools, there was, incredible breakthroughs with Mythos, GPT Cyber, like great benchmarks on those, but could be very expensive. We heard, I think, wasn't it Nikesh Arora said, I think he spent like $1 million to find some zero-day bug and it was like totally worth it. But that's expensive for smaller companies that might just need, it would be great to just be able to go to a cheap cyber-tuned model, get approved by Google and say, hey, just go and read every line of code and just let me know if there's anything that's obvious. That could be good. What do you—

Yeah, and just to be clear, they released 3.5 Flashlight and 3.5 Flash Cyber. So it's not like Flashlight Cyber.

Okay, but Flash is still cheaper than the big model, right?

Yeah, there's like the Pro and then there's—

And so I have to imagine that if you— given the choice, if you're like a mid-sized company and your choice is like, go beat down the door to get into Project Glasswing to spend $5 million to have Mythos read every line of code, or you're like, Oh, well, I'm already a Google partner. I have a really solid business relationship with them because I've been buying ads and I have Google Webmaster Tools set up and I just want to go set up Flash Cyber, not Flashlight Cyber. I imagine that the cost would be cheaper because they fine-tune this model for this specific thing. So Anthropic CEO Dario Amadei for years has warned about the risks of powerful and open AI systems, saying in a recent Bloomberg interview that having AI models with advanced cybersecurity capabilities that are free to download could be harmful. He said it's a serious concern. Use of Chinese models, which are far cheaper than US counterparts, is surging at US companies, prompting some investors to question the staying power of top model makers such as Anthropic and OpenAI. The current US open-source frontier is starting to catch up with China's. On Wednesday, Thinking Machines Lab, led by former AI technology chief Mira Murati, released its fir— first AI model as open weight. NVIDIA's NeMoTron 3 Ultra is starting to see traction. That's interesting to hear. I haven't heard nearly enough buzz around NemoTron. I know Jensen has been investing in it. He's also investing in a lot of up and down the stack.

I've been hearing buzz about the name, John.

NemoTron's good.

NemoTron.

Why?

Very cool name. There's people are, you know, there's, there's some talks bubbling. People are saying it's, it's, it's one of the best names for, for a family of models ever.

What comes after Ultra? Isn't that what the— didn't Apple first do that with the The M4 Pro, the M4 Max, the M4 Ultra. So we got to come up with an even better word for what happens after Ultra. Giga, maybe Hyper, and then Exa. The market's faith in Anthropic and OpenAI continuing to build more capable models that push the frontier— AI frontier— has been at the heart of the boom, helping justify trillions of dollars in spending on infrastructure in the coming years. The threat that new players will vastly undercut what they can charge for advanced AI pushed down some tech and AI company stocks prices last week. David Sachs, the venture capitalist and White House AI advisor, suggested that he read Ball's post as a confession to a regulatory capture strategy. Sachs and others have long seen calls for AI regulation by companies such as Anthropic as efforts to use new laws and policies to stymie competitors. The weaponization of regulatory uncertainty as a competitive tool should be completely unacceptable. He said. Security-focused administration officials in the past year have considered a slew of measures to crack down on Chinese AI companies developing open models, people familiar with the discussion said. They've weighed adding the businesses to trade blacklists, security warnings on Chinese AI companies, and a potential executive order targeting open models. So it does seem like even though Tyler Cowen, we were going back and forth with him about this, about like, can you even ban it? You can't really ban anything in the land of the free. Like, the solution to piracy was not really to ban piracy. Like, you can still torrent things, but it's just so easy to use Apple TV and Spotify and Netflix that most people just wind up just paying $5, $2 for a song or, or $10 a month for something that it effectively killed it just because the product experience was better.

This morning, our Treasury Secretary Bessant was on Fox and said this administration supports open source models, but what we do not support is IP theft. Hmm. If we see especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft. So it will be— if the admin does take serious action here, I would expect that to be the line of reasoning where they could just sanction the companies behind these models, which I imagine would create a real opportunity for American open source. You would ultimately see a lot of these, you know, neo clouds unable, I imagine, to serve the traditional Chinese open source models. I imagine that even OpenRouter would probably have to—

Yeah, maybe.

Stop supporting them.

But again, with an open source model, Like, how do you even define it? Because if there's a research paper and then you go and implement that and you train it or you take it, you fine-tune it or you use the data to bootstrap your own model, like, just like where, where in the supply chain, how many steps away do you need to be to like not count on there? Didn't Perplexity do like a 1776 DeepSeq very quickly? They did a fine-tune. I think you did a fine-tune on DeepSeq as well.

Yeah.

So that was like using sparse autoencoders. I did it first, so.

But what nationality was the model that you used as the base?

Well, it was Deepseek. It was Chinese.

Wow.

Confession. You can go now.

Yeah, people are already angry at Besant's comments and are saying funny that he's not mentioning that yesterday a US judge approved Anthropic's $1.5 billion copyright lawsuit settlement. I think they settled that they were in possession of books that had been torrented or basically acquired unlawfully.

Yeah.

I don't believe that—

Acquired unlawfully? Sounds like they stole the book. I think they bought the books.

No, no, no, no, no. I think this was— No, separately, they bought a bunch of books, trained on them, and destroyed them.

And that just made people upset because no one wants the idea of books being destroyed, of course.

I think those were rare, rare books.

The Wilmanitis Collection was destroyed. Rare PDFs.

Judge William Alsip of the U.S. District Court for the Northern District of California issued a preliminary approval of the settlement last year after ruling that Anthropic had illegally downloaded and stored millions of copyrighted books, and the settlement was signed off on by a different judge yesterday. The payout will deliver $3,000 per work across an estimated 500,000 works. Shared among the authors and publishers who hold rights to them.

Anyway, there was some news about like whether or not, uh, Kemi K3 is distilled. Tyler, can you take us through this Lisan Al-Ghaieeb post?

Yeah, so he didn't actually make this chart. It was from someone else. But basically this shows a bunch of models compared to another and it's like the kind of semantic similarities between them.

Okay.

So it's like if you use similar kind of diction or phrases, stuff like this, you just like prompt the model a bunch of times and you kind of can tell what's similar to each other. And you can see in this version, it looks like, like Kimiki 3 is very similar to Fable 5.

Fable 5, and then also Opus 4.7, Opus 4.8, and Sonnet 5. So they might have gotten tokens from like all 4 of those models sort of mixed in together.

Also, I think this exact example, like only like a few prompts were used. It's not like super involved. It's also like, this is very much just like correlation. This is not like proving that they distilled or anything. They could have just used like similar, you know, post-training techniques or something like this. Totally.

But notably, GPT-5.4, 5.5, 5.6 share very little similarity to the Anthropic models and vice versa. The Anthropic models don't seem to be correlated with the OpenAI models, although the OpenAI models are correlated with them, with each other. So 5.6 Luna looks pretty correlated with 5.4 because they probably were like, well, we're building on the shoulders of the work that we've been done. It's the same thing with Anthropic where Fable has some of the same flavor as an Opus or a Haiku, and that makes sense.

That makes sense because people have been saying, oh, maybe those models were actually distilled from the big Mythos pre-train, right?

Yeah. Or vice versa. Fable is like Mythos class. Or vice versa. You have a bunch of users that are doing rollouts on Opus, trying to vibe code different things, and you take all that training data and then you feed that into Fable, and Fable sort of like picks up on a lot of that because that's what the, the, the user said, wow, you really did a great job on this Opus, or you did really poorly. It pulls that out, but it's still learning the, the, the textual similarity. Well, people will, uh, go back and forth on this and, uh, and debate the level of distillation and also what the punishment for distillation should be. Should it just be a slap on the wrist, or should it be something bigger?

The only thing I would say here is, like, distillation against the terms of service.

Yeah.

And I think that suing Chinese companies, it would be a complete and total waste of time. Ask anyone who's built a product that's been knocked.

I disagree. Incredible aura farm. If you personally serve Xi Jinping the next time he's in San Francisco meeting with a world leader.

You show up, you have been served.

You show up to the UN, serve the world leader. There's something about that that's, you know, extremely bold. But I know I agree with you, uh, very difficult. I don't even know if you went to your law firm and you said like, hey, we want to sue this Chinese company, like how they would trace that down. Jim Cramer says we must not let our companies use these Chinese models to save a few bucks. OpenAI and Anthropic are correct. This is vital national security. Please Please read Bing West's just released Cat5. I respect the Chinese people greatly, but these companies are run by the PLA for heaven's sakes. George Hotz fires back under the TinyGrad handle. He says, is this a joke? I'd be happy to use Anthropic if they put their weights on Hugging Face like the Chinese. American tech companies are rent extraction machines that have lost the plot. No point continuing to defend them. They need to change if they want to win people back.

Meanwhile, China is also thinking about export controls in the Financial Times, apparently. Recently, Alibaba, ByteDance, and—

Keep going back and forth. MoFCom.

MoFCom talked to AI companies including Alibaba, ByteDance, and Zhipu on limiting the transfer of key data for the training of their models overseas, probably as well as allowing their model weights to be downloaded.

Or Ministry of Commerce, probably their Department of Commerce.

That seems more like it. China would still let overseas customers access the models and services, but it sounds like they're leaning towards not making the weights open. We're going to see with the new KIMI K3 if they actually release the weights. Next week. Dean Ball says, wow, 2 days in a row where things I said are likely to happen seem likelier to happen from 2 adversarial countries, almost as though my analysis was right and as though people should not have deliberately misinterpreted me as of making prescriptions. Huh.

Dean Ball victory lap. You got to take a victory lap for Dean Ball. He called it. No, he— I mean, he was saying like, I don't necessarily agree with this, but just from his experience, this is how things might play out. And people were very upset with him about that.

Raj in the chat says Ministry of Commerce, People's Republic of China.

Yeah, it's their Department of Commerce.

That would go hard on a business card.

It would.

The great slop wars— slop wars continue. Let's pull up this video from Substack. They say today Substack is launching an AI detection feature via an integration with Pangram. Going forward, you'll be able to scan posts, replies, and comments on the Substack app to see an estimate of how much of it was written by a human or with AI assistance.

Mm-hmm.

We care about this at Substack because it gets to the core of our mission to build an economic engine for culture.

This is gonna be bad for Tyler since he's mastered the ability to write exactly like an AI. What was that sentence, sentence you wrote by hand that was like, according to studies, you're absolutely right. And you just threw in every possible AI-ism and got flagged 100% by Pangram. No, I think people will enjoy this. The revealed preference is probably that people don't care if it's good and they enjoy the content. But we'll go back and forth and debate it with Chris Best today and see where all of that is going. Now, do you think that Substack writers should admit when they hire essentially ghostwriters?

Uh, probably.

Because sometimes you start a brand under a name like not every post on Bloomberg.com is from Michael Bloomberg. In fact, none of them are. They're all from other people, but they disclose this is from, you know, some writer. This is from Joe Weisenbaum, right? But some new influencer brands have grown even in the pre-AI era. I had a writer on my team when I was doing YouTube videos where I would write something, he would do a draft, we'd try to go back and forth. I'd put him in the credits, but I wouldn't like kick it off with like, "Hey, just so you know—" Andrew in the chat says, "Book authors don't." That is certainly true. That's 100% true. Some book authors will put like, "This is written by this person and this person." It depends on like the aura of the ghostwriter. Some ghostwriters actually do get credits in the book, but yes, most book authors— like Tom Clancy, I think passed away and they're still producing books. There's a number of writers where they're their individual brand grew so big that they can— you can have a new company. Yeah, they become their own brand.

And so, yeah, I would imagine that 30% of the books in like an airport bookstore are just 100% ghostwritten and probably don't disclose it.

I think— I do think it should be possible to get to the truth if you hunt hard enough, but I don't necessarily need you to put it front and center.

For a Substack author or creator, just put it somewhere. Yeah. That you have a team of people that help you create.

Yeah. Your work. Well, Dave Eggers took to the Financial Times saying that no one is going to read AI books. He's of course the author of The Circle. He's an American writer, editor, publisher. A Heartbreaking Work of Staggering Genius is his 2000 memoir bestseller. Final finalist for the Pulitzer Prize. He says, gullible and ridiculous. No one will read AI novels. I think people are already reading AI novels, so maybe he means like no one who's cool or no one who he likes or something. Let's read through a little bit of this Financial Times article. Dave Eggers offers a hand so I can clamber aboard his 23-foot sailboat moored in the shadow of San Francisco's Golden Gate Bridge. It's a grand setting with few frills as we sit on the stern a meter apart, fuel cans and fishing nets at our feet. There's no one around to disturb our conversation. The nearest toilets are porta-loos a short bike ride away. The berthing fees are negligible, making it one of the cheapest places to live in uber-wealthy Silicon Valley. I hadn't thought of that. Sailboat, if you're at a tech company, this has got to be the next big lock-in trend.

Antonio García Martínez did this. He talked about in his book Chaos Monkeys.

I'm not— I won't be satisfied until I see LARPers posting images on Twitter being like, hey, I just moved to San Francisco. I'm building the next one-person $25 billion.

I'm finding a houseboat in San Rafael. Yeah. How much? For $1,000. It actually works out to $1,483 a month, apparently, if you finance it through Boat Trader. Goodbye. It's only about $200,000 to live in a 2022 houseboat.

2022. That's a brand new houseboat.

Yes. You have to think about the— yeah, the slip fees. But you could easily come in, have your own house on the water.

The average— the average one-bedroom rent in San Francisco is like $2,000, $3,000 right now. Yeah.

You could have a whole house and a boat and a boat.

It's amazing. This is the hack. If you get a job in tech and you're printing, making six figures. Wow. Move into a houseboat.

Let's pull this up. Is this an option in L.A.?

Should members of the team be doing this? Should we— should we recommend this? Look at this brand—

almost brand new. I feel like 2024 Le Mer Modern 11 is a 36-foot houseboat. We have a video here.

Well, Dave Eggers says they call it a $10 dock with a million-dollar view.

Look at this, John. This? Yeah. A couple thousand bucks a month.

This is where the next great startup will be born. Yeah, sure.

Garages are out, garage houseboats are in.

And also like, oh, take your shoes off. It's not— it's natural to take your shoes off when you're on a boat, you know, right? Let's get boat shoes.

Wear boat shoes. The Le Mare Modern 11 offers reliable performance and strong power with 2 Yamaha 25 high thrust engines providing a total of 50 horsepower. It's 2 cabins and 1 head ensure overnight comfort for extended stays. How about year-long stays? How about a decade-long stay? The boat's design includes an upper deck terrace with Flexi Teak flooring, enhancing its appeal for relaxation and entertainment. And might I say, locking in standard features such as air conditioning. We have air conditioning on our houseboats, uh, with reverse cycle heat, heated floors, LED lighting, LED lighting. Yeah. The inclusion of foam-filled composite floats and galvanized steel subframe ensures durability and stability. Did we just— this thing looks—

housing crisis. Send this to Sauger and Jetty immediately. This is seasteading. This is the teal thesis. Everyone living in houseboats, just more.

BoatTrader.com is saying that this— this is amazing— modern houseboat would come out to just $2,259 with only $52,000 down. I promise that we have no financial relationship with boattrader.com. Yeah. Even though I'm making a pretty strong pitch here.

Yeah, I love it. Should you recline your airplane seat? What do you think? People are going back and forth on this.

Brandon Ducks says, I don't remember subscribing to this channel, but I've been listening for 10 minutes. I'm having a good time. Welcome. Last name Ducks comes on just for that. Yeah.

Make it make sense. Make it make sense. It's just like I just happen to be here for the ducks. Section of the show.

Andrew says, I'm gonna need a double-wide houseboat, LOL. Bootlegging distilled Anthropic models on the houseboat. Living the dream.

Okay, so should you recline your airplane seat? There's a central— investigating the central dilemma of our time, Joshua Rothman in The New Yorker wrote a piece. Imagine this scenario. After a long drive to the airport, you've just taken off for a 15-hour flight. Where are you going? 15 hours is so long. How long was the flight to Austria from LA?

I think it's like 12 hours. So you're going to the Middle East.

You're going to the Middle East, maybe. OK. You're in coach, and a family of 3 is sitting behind you— a man, a woman, and a toddler. They seem nice, but as the flight continues, the kiddo grows restless. He kicks your seat and pokes her toys into your shoulder. Stop that. You hear the father say. The toddler incredibly quiets down. The cabin lights dim. It's time to sleep. Yet just as rest closes in, you sense motion. The seat in front of you is reclining. You watch as it comes all the way back. Now you feel cramped with someone else's seat in your face. You also notice some stiffness in your shoulders. Your fingers caress your own recline button. Surely the reclined upon are entitled to recline. Just as you start to lean back, you hear a clearing throat. You turn around to see the dad who fixes you with his gaze. Excuse me, he says. Would you mind keeping your seat up? I think it's game on. They put the button there. They would just take it away.

Yeah, I think— I think I've always expected everyone to just recline.

Everyone reclines and everyone has the same amount of space. It's the conservation of space.

And it's really hard to make the argument that the vertical seat is more comfortable than the reclined seat. Exactly. So I think everyone commit to comfort. Everyone commits. And then the chat is sharing the real alpha. Stand up immediately. When does the flight land?

Have you seen those meta Ray-Ban videos where people sprint to the front? It's so rude. The guy is like, like, you know, he's like, as soon as it goes beep, he just bolts and passes everybody and gets to the front. Very rude. I don't think you should do that. This is antisocial behavior.

I don't think you should recline your seat. How is it antisocial? Because you're infringing on someone. Just because you can do it doesn't mean you should.

But they put the button there. Horrible take. Horrible take. That's like, don't use the bathroom on the plane. They put the bathroom there.

My entire life, even as a child, I was looking up waiting for, you know, the seatbelt sign to be up in the air. Seatbelt sign's off, you recline your seat. Everyone reclines their seats.

All good. Yeah. What about ordering a Diet Coke on a plane? If I'm sitting next to you, will you order a Diet Coke or is that antisocial behavior?

That's completely different. Okay.

What's your stance on JBL in the airplane?

A little fatty boy.

Pro JBL, you know. I do think that could be the next trend.

The big question is, are you bringing on your, you know, smart portable speaker, maybe from an OpenAI? Are you bringing that on the plane and talking to it the whole time?

Is that offensive? I think that's deeply offensive.

Have you seen when like a band is traveling and they start playing on the plane? No. Have you seen these videos?

No. Yeah, that sounds cool though. I'll send you one. Yeah, I'd be down.

That would make me jump.

So this all started because, uh, in 2014, FiveThirtyEight— we gotta have him on the show, uh, Nate, obviously— uh, FiveThirtyEight asked around 1,000 flyers what they thought about reclining their seats. 41% of them said that reclining was rude, so they lost.

It's not rude. It's over.

That number seems to have risen though. In 2020, Eric Jones, a math professor who writes about travel statistics on the website The Vacationer, conducted a similar survey and found that 77% of respondents objected to reclining. Wow. Sheep, sheeple over here. We got sheep.

Okay, how about this? I think you should be able to recline if you're like dressed nicely. If you're wearing sweatpants, then you don't get to— you don't get to recline. You gotta sort of conditional— yeah, yeah.

Should kids be able to kick the front— the seat in front of them as much as they want?

See, that's something I will put— I will totally put up with. I'm totally fine getting kicked in the back by a kid because it's so brutal as a parent, as a kid, to be on a plane. They're locked up in the plane.

That's miserable.

And Just, just tough it up. You're reclined. You have multiple inches of padding. Unless you got Ronaldo back there, you're going to be fine.

Order that hot pizza pocket. Live life in the permanent underclass. Live a little, for God's sake.

Ramp launched a model router. Braxton loves it. He says Ramp launching a model router is god tier product expansion. The winds are blessing it. The market is exhausted with token maxing. Proliferation of new models, sovereignty. This is the first time I've looked at a launch and went, wow, this might be the start of a new core product. Yeah, there's been a bunch of speculation about this, but it makes a lot of sense. RAMP's promise is to save you time and money. The CFOs are increasingly annoyed by token routing. We talked to Eric Lyman about this last week. Are you using the craziest frontier model, burning millions of tokens? To answer the weather, maybe that should have an expense report attached. And so very, very exciting to see where this goes, how they roll it out. I know it's in alpha with a number of customers, and I believe that the product's actually available to non-RAMP customers as well. So you can use their product exclusively.

Very cool. Yoush says going after everything time and money touches. I love it.

I love it. Anyway, leave us 5 stars on Apple Podcasts and Spotify. Sign up for our newsletter at tbpn.com.

And we love you. We love you. Throwing flashbang.

Goodbye. Flashbanging the horse.