The big news of the day is that China has taken a page out of Washington's playbook potentially and aiming to limit the export of AI models. Now, these are just meetings at this point, but Chinese authorities have reportedly, this is according to Reuters, held meetings with some of the country's top tech firms. That's Alibaba, ByteDance, Z.AI, DeepSeek, Qwen-Fly are not mentioned yet, but you have to imagine that they'll be in the conversation if they aren't already about potentially restricting overseas access to China's most advanced model. So, they're thinking about doing the same thing that's happening with GPT-5.6, Soul, Ultra, Mythos, Fable, all these different models that are starting to get more restricted as they get more powerful. The government wants a say in what companies and what non-governmental organizations, what governmental organizations have access to them. So, the discussions are still preliminary.
When did, in AI 2027, did China wake up? Right around this time. Right around this time, mid-2023. I actually have on my calendar a check on Jack Clark for today because last year, December 23rd, Jack Clark predicted that there would be big changes and acceleration in mid-2026. He said, 'So by summer of 2026, it will be as though the digital world is going through some kind of fast evolution with some parts of it emitting a huge amount of heat and light and moving with counterintuitive speed relative to everything else. Great fortunes will be won and lost here and the powerful engines of our silicon creation will be put to work further accelerating this economy and further changing things.' Do you think that's where we are? Is that an accurate assessment of mid-2026? When I saw this prediction, I put a calendar reminder for mid-2026. We're right in the middle of the year and I wanted to check in and say, does it feel like this? Obviously, this is not a quantitative benchmark. This is a qualitative feeling. Fast evolution, emitting a huge amount of heat and light, moving with counterintuitive speed relative to everything else. Great fortunes certainly are being won and lost every day in the market. Tyler, how would you assess the amount of heat and light emitting from some parts of the digital world? Would you say it's huge?
I would say it's huge. I think it's huge. It's kind of hard, like I feel like I've been like very tapped in for a long time. So, it's hard to say like...
You've been feeling the...
It's hard to see like an outside view is like, oh, is this actually like, has there been a big change like recently?
Yes. So, there's two things like one...
It's been kind of like a personal thing for you being so tapped in. Would you say that?
So there's, I mean there's two things going on. One is that anyone who has...
been like a personal thing for Tyler. He's been so tapped in for so long. It's like...
No, so much of news coming out of DC. Yeah. The past like two, three months has been about AI which is like definitely was not true. Yes. You know even earlier this year I would say.
Yes. And just the amount of heat and light around cybersecurity issues, the coding models, the agentic capabilities, the revenues that we would see in enterprise, those have not been a shock to AI insiders who have been tracking the capabilities of these models since last October or even last summer when models started, when vibe coding was coined by Andrej Karpathy and coding became, you know, not solved but, you know, greatly enhanced by artificial intelligence. So going back to AI 2027 which was published April 3rd, 2025, by late 2026 they had, or sorry, mid-2026 was China wakes up. In China, the CCP is starting to feel the AGI export controls and lack of government support have left China under-resourced compared to the West. I'm going to fast forward a little bit. Hawks and the CCP warned that the growing race towards AGI can no longer be ignored. So they finally commit to the big AI push that they had previously tried to avoid. He sets in motion the nationalization of Chinese AI research. Okay.
Creating an information sharing mechanism for AI companies. So right now they're just looking at export controls.
Feels like we're a little bit behind on this.
It's possible the CCP is just reading this as a guide and just implementing it. They're just planning to implement it to a T.
Just like, yeah, thanks. Thanks for the playbook. You should have billed me for it.
Yeah. No, the real 4D chess here is to shame the CCP into not copying off of AI 2027, saying, 'Oh, really? You're just doing the AI 2027 thing?' Come on. You don't need to invest in AI. Don't be mimetic with our philosophers' literature and our most popular podcasts over here in America. Think of your own strategy. Come on. What? Do whatever. No. So, the discussions around AI export controls in China are still preliminary. And it isn't clear how Beijing could claw back access to models that they've already released. If they're on HuggingFace, you can download them, you can run them. And even if they had some sort of intellectual property, they could sue an open-source inference provider for running Kimi or something like that. That feels really difficult to do from Beijing. But there are other implications. So, once open-weight models are posted publicly, of course, they can be copied, mirrored, downloaded, reused anywhere, fine-tuned. But there are several ways China could restrict future frontier models before they leave the country. At the lightest end, Beijing could require companies to file future releases with regulators or submit them for national security review before publishing model weights. A more aggressive version would require government approval before any advanced model could be released overseas. And at the most restrictive end, China could simply ban the public release of frontier open-weight models. They could say no more open-sourcing. Although the game theory there is a little odd because it feels like the fact that China has been open-sourcing near-frontier capabilities has been to China's benefit in general because it creates like more disturbances, like the DeepSeek moment was definitely good for China.
Yeah, pricing pressure.
So they could limit these to state-approved customers or Chinese government priorities which at a certain level might just be the most important thing to do. The RSI loop. Of course, you want all of the GPUs that you have focusing on advancing the frontier internally as opposed to sending it off to America. The talks come as the US is also moving toward tighter control of frontier AI. The Trump administration recently restricted access to Anthropic's most advanced models, Fable 5 and Mythos 5, before later lifting some of those controls after additional safety mitigations. OpenAI has also reportedly faced government pressure around the rollout of its next frontier models. The issue has become more urgent because Chinese open-weight models are increasingly competitive. Z.AI's GLM 5.2 in particular recently impressed Silicon Valley by performing close to leading US models at a much lower cost. The Qwen family of models from Alibaba has also become one of the most important open model ecosystems in the world. While ByteDance's Doubao is one of China's most widely used domestic AI systems. I didn't know about that. If China places export controls on future frontier models, it could raise costs for American AI users and companies because it's obviously very beneficial to have a great open-source model available to you and use them for lower-stakes tasks while reserving expensive US frontier models for their most important workloads. If the best Chinese models are no longer freely available abroad, some of that cheap model supply disappears. But you have to ask what happens to Meta? Do other labs start piling back into the open-source ecosystem more, or will all the demand just go straight back to the frontier labs? That's the question. It really does. It really does. We should play some clips because...
We need to be thinking about podcast safety because this is...
an addictive...
There should be export controls on this. It should be only available to American citizens. They should, Patrick, even one of the guys over there is Canadian. Yes. Maybe North America. Let's pull up a clip from this. I want to hear Jeremy chat about all sorts of things. They go all over. Where in the timeline is...
Well, should we talk about SK Hynix's new listing?
Yes, we can run through that and then we can pull up.
SK Hynix is set to raise around 28 billion in a NASDAQ share sale this week. Potentially one of the largest ever New York listings by an Asian company. The South Korean memory chipmaker is already public in Seoul where its stock is up more than 750% over the past year, but the US listing gives American investors an easier way in. The story is big on the timeline because Situational Awareness, the hedge fund run by Leopold Aschenbrenner, may participate alongside Baillie Gifford and Coatue. Together, the firms could take as much as 7 billion of the deal. SK Hynix has become one of the purest AI infrastructure plays because it leads in HBM, has become a key Nvidia partner, and now has a market cap around 1.1 trillion,
which is like a quadrillion won.
When you look it up, it's very confusing. I think it's a lot.
The business is ripping. 2025 revenue grew 47% to 63 billion. Profit more than doubled, 28 billion in profit. Q1 revenue tripled year-over-year. But interestingly, the stock only trades at 7x forward earnings because memory is brutally cyclical. And there's still a lot of people that are wondering what happens if there's another leg up in the AI demand and the AI buildout. Will that mean more competitive pressure? Will there be a bottleneck where they do stagnate? How much can they actually grow? And so many investors that are just in memory have seen booms and busts from cloud or smartphone or crypto and they say, ah, like I don't know that this one's going to last forever. I'm pattern matching to previous cycles not thinking about this as a different technology cycle. So I think that's what's going on with SK Hynix but it should be a blockbuster IPO nevertheless. Very interesting to see an Asian company listed.
And the most notable part of this deal is obviously Situational Awareness.
And their participation in it.
They cover the best, the billion-dollar PDF, an idea that Jeremy and Patrick have talked about before. But Situational Awareness is turning out to be more of a maybe a 10 or a hundred billion dollar PDF.
Yeah. Seriously. A whole investing thesis built around that. It's just a PDF. I don't know. It's an essay, a thesis.
If so, if the PDF was worth a billion or more than that, what was the Dwarkesh episode worth? It's got to be some percentage of it, right?
Yeah, definitely. Because a lot of people watch the podcast, watch the clips, scroll through the PDF.
Breaking news. Meta unveils Muse Image, its first image generation model from MSL.
Waiting for this. Let's pull up some images of this and see the results. It should be good given that they have all the Instagram and Facebook data to train on. Alex Wang has a post.
Let's pull this up.
Okay. Muse Image, Muse Video, whole bunch of very detailed badges printed all in one image. He says releasing Muse Image today. The first image generation model from MSL. It is crazy that they didn't have any like when back when it was FAIR, did they do anything with image? Did they? Because they had that amazing model, Segment Anything. So they were doing work on images and obviously they've processed a ton of images. It pairs with Muse Spark to reason through your prompt, search the web and plan before it generates. People get what they mean on the first try. Live now in the Meta AI app. Very, very fun. Even puts a realistic QR code in there and can do infographics and old Polaroids as well. Big su August 1983. Three things Alex Wang is excited about under the hood. Self-refinement. The model improves its own output within its chain of thought. Emerge during IRL, not by design. Multi-reference composition. Many images blended into one coherent generation. Multi-turn editing. Iterate without losing coherence or starting over. Make an image of this person riding a bike, this bike, and wearing this suit while passing by these people on a park bench. Make it look like a drawing in this style. That's cool. So you can upload multiple images and sort of drop them in in his various prompts. What you got Tyler?
Yes. So Meta did release a number of image models. So there was I think it's pronounced Chameleon in 2023. There was Emu also 2023 and then there was another like version of Chameleon.
I think they were somewhat different than the like diffusion which was like popular at the time. I think they were all autoregressive transformer which is like how you get the good tech stuff like that.
But I believe this is a different theme.
That's cool.
Yeah. And Alex is also previewing Muse Video competitive prompt adherence, visual fidelity, temporal consistency. Also coming to Meta AI. This has got to be GPU intensive if they actually roll this out to everyone. Although Meta AI does have a smaller install base than Instagram, but I mean imagine you really push this in Meta AI and...
what they buy all the GPUs for.
You got to get people generating stuff. I completely agree. I think put this in the hands of the modern Instagram creator. Let them have fun with it and see you already have the algorithm to filter the slop out. The stuff that is sloppy but still entertaining and creative will surface to the top. So, congrats to them on a successful launch. And let's click over to Jeremy Gons on why he doesn't worry about AI taking jobs. A very, very interesting take on the latest episode.
Understand this idea of we're going to run out of jobs. To me, it's like very obvious that every, certainly every white-collar job is like totally fake and made up in the sense that like they are not contingent for shelter and food and medicine and other necessities. But most jobs do not touch those or if they do they touch it in a very, very derivative way. What is your job as an allocator? It's like, well, because capital is inherently inflationary, you have to, you can't just leave it alone. And so, like, my job is like when you have money and you don't want it to go away, you have to give it to someone. And so, you give it to a bunch of people and then I take it and I put it into things that are productive and then hopefully you don't lose your money, you get more money. And like, is this useful? Is this good? Yeah, sure. But like, it's not real. And to me, there's like unlimited amounts of jobs that you can create in those sorts of scenarios. We're going to have unlimited wants and desires and our economy is solely driven by our like unquenchable desire to consume things. And so we're going to come up with new things to consume. And now again, in the short term and medium term, that might be volatile and there might be a lot of job loss, etc. And you know, that's not good. And you know, there could be a lot of despair. But in the long run, like we're just going to invent new things to do. Like we've already solved all of our problems. Like the worry about, oh, we're not going to have more jobs. It just doesn't really resonate because I feel like we just make up stuff for us to do. And that's sort of the whole point of it.
That makes me think we need an Invest Like the Best soundtrack.
We do need to Invest Like the Best soundtrack. Let's play the other clip that Patrick shared. The West Coast is definitely eating the East Coast. The founder is incredibly important and sort of the founding act is incredibly important in any business. I do think that it's notable that like the current paradigm in which we live, the largest and most important finance firms come out of a culture of leverage buyout which it's extractive in so far as the primary goal is to make a thing more profitable. The core idea behind it is like using debt and financial engineering to make a lot of money in a way that the quality of the business itself is maybe ancillary to the core in...
Afterglow says, 'My god, what an obnoxious soundboard. Unwatchable.'
Sorry, it is a little loud today. I feel like the soundboard's coming in a little bit hot. We're sorry about that. But we're having fun here. And of course, we don't want to pirate the entire Invest Like the Best episode. We got to tease it.
This would be a good use case for AI if we can make a separate show that's just the show, but an agent in real time pulls out the sound effect and just has a normal show, no sound.
Sure. Sure. So, you can choose whether you want the soundboard version of the show or the not or the...
Tyler says, 'Don't be sorry. That's why I'm here.' Tyler's just here for the soundboard. So, unfortunately, there's people that are here that don't like the soundboard. There's people that are here only because of the soundboard. We are trying to find the middle ground and we don't always get it right.
