Taylor Lorenz’s Power User - The AI War Just Went Nuclear: Everyone Hates Anthropic Now (Even Denny's Diner)

Episode Date: July 29, 2026

The biggest fight in AI isn't about chatbots anymore, it's about who gets to control the future of artificial intelligence. SUPPORT MY WORK: Buy a paid subscription to my newsletter at https://www.usermag.co    Support my work on Patreon: http://patreon.com/taylorlorenz        Is Anthropic trying to lock down AI before their IPO, or are open-weight models genuinely a threat to global safety? In this episode of Power User, Matthew Berman, co-founder of the tech media outlet Forward Future, breaks down the explosive debate rocking Silicon Valley and Washington DC. After the release of China's powerful Kimi K3 open-weight model, OpenAI's autonomous hacking benchmark, and a growing push from Anthropic to restrict open AI models, Silicon Valley has erupted into one of the biggest technology battles in years.   Nvidia CEO Jensen Huang, Microsoft, Meta, Google, IBM, Hugging Face, Mozilla, Palantir, Dell and dozens of other massive technology companies have all come together to publicly back open weight AI models, while Anthropic argues against them. Whoever wins will control the future of the tech landscape for decades to come. I break down: What open weight AI actually isHow Kimi K3 changed the AI raceThe OpenAI benchmark that escaped containmentThe Hugging Face security incidentAnthropic's campaign against open modelsWhy Jensen Huang and all of Silicon Valley is pushing backThe growing debate inside Washington over regulating AIWhy this fight could determine who controls AI for decades

And the funny thing is the benchmark that they were running it on was testing its cyber attack capabilities. So literally the thing that they were testing it on, it used to escape and get the answers. So yeah, it is scary. Imagine if only two companies were allowed to decide who could use the internet, who could write software, and who could publish a website. That's the future that Anthropic is currently pushing for with AI. And conveniently, it protects their business model right before a market. multi-billion dollar IPO. Over the past few weeks, a massive debate has broken out from Silicon Valley to Washington, D.C., on how to regulate AI. After the release of a powerful Chinese open weight model

called Kimmy K-3 and a high-profile hack by a sophisticated open AI agent, regulators and technologists are fighting to decide whether the world's most powerful AI systems should remain under the control of a handful of multi-billion dollar tech companies or whether AI models should be, downloadable, modifiable, and impossible for any one or two tech billionaires to monopolize. The debate is being framed around safety and market competition, but at its core, it's about power. And who gets to control the infrastructure of the AI age, which is ultimately who gets to shape the future of technology for decades to come. Matthew Berman is a content creator and entrepreneur. He's the co-founder of Forward Future, a media company covering AI news.

Today, he's joining me to break all of this down. Hi, Matt. Welcome to Power User. Hey, Talley. Thank you for having me. Before we get into this drama, I kind of want to start by taking people back to the beginning and sort of like the origin of what kicked off a lot of this discourse. So on July 16th, Moonshot AI, which is Chinese AI company, officially released Kimi K3, which is an open weight model. It's this new, massive, powerful LLM.

It has a 2.8 trillion parameters, 1 million token context window. So why was this so notable and why was this model sort of so influential when it launched? I actually think you can probably point back to about a year and a half ago with the Deepseek moment. Model came out that was open source. It was also a Chinese AI lab model and it was incredibly powerful and really efficient. And so it kind of made us stop and think, oh, like China is making really good models. Fast forward a year and a half and now we have the Kimmy moment. And Kimmy and the Moonshot team have done something incredible in that they've developed

essentially a frontier level model for far less cost and also it's open weight, which means anybody can go download it. You can use it. You don't have to send your data to a Chinese company if you're not into that. You can simply download the model. You can use it from a U.S. inference provider. It's an incredible time because we have open source excellent models. What is an open weight model?

And what do you mean when you say, sort of anyone can download it and use it? Essentially, there's kind of a baking process. So a model gets created. And then the final output is what's called the weights. And it's just as you can think of it like a series of numbers. And those weights can either be open weights or close weights. Now, Anthropic and Open AI are close weights, closed source completely, meaning you have to go to their API.

You have to go to chat GPT or Cloud to use it. And you have to pay them famously, lots of money for tokens. Exactly. You have to pay them. Whereas with the Kimi model, you can actually download the final product, the weights. You can actually put it on your computer. Now, it is a massive model. So you do need really high-end hardware to run it.

But the point is you could if you wanted to and any company can. So Amazon with AWS, they can serve the Kimmy model on their own infrastructure. Google Cloud can do the same thing and the dozens of other inference providers. Yeah. And I think what's so notable here is that startups can download this model and use it. And as you mentioned, you know, the weights are open so they can be tweaked. They can be evolved. You know, you can find millions of tweaked versions of these models.

very valuable for startups, right, that want to be able to tweak these models, leverage them, and not pay billions or millions of millions of dollars to, you know, Open AI or Anthropic. Yeah, Taylor, there's a lot of benefits to being able to actually own the model weights. So Kimmy K3 launches. It's the largest open weight frontier model that is basically China's strongest direct competitor with Open Eye Anthropic. And very quickly, Anthropics starts calling, you know, fraud. They're like, hey, this model is only so good because it was

distilled. It's a knockoff of our model. And they start bringing up this concept of distillation. Yeah. So Anthropic actually started claiming this a few months back. Now, when Kimmy was released, they started saying it again. But essentially what distillation is, is if one company is asking questions to another company's model, getting the answers, you can actually use those question and answer pairs to train a really capable model without actually having to invest. so heavily in the pre-training process, like if you were doing it from absolute scratch. So when Anthropic develops a model, they're creating it from scratch. They're taking all the data.

But distillation is you're kind of getting one larger model to help teach a smaller model. And distillation, by the way, it's not a bad thing. It's not illegal. It's actually a really good thing because you can have a frontier company that has this massive model that's really expensive, but they can make a smaller model that's, let's say, 90% is capable, 95% is capable, and they can charge a fraction of the price for it. So distillation is actually good. Now, what Moonshot is accused of have doing is illegally distilling, creating a bunch of fake, fraudulent, anthropic accounts and extracting a bunch of data from the model to then train their own

model at a fraction of the cost that Anthropic had to pay to develop theirs. Yeah. And this created a lot of, I guess, controversy online. I do see the irony. Obviously, it's very different anthropics, you know, scraping stuff off the web and using it to train their bottles. But people, I mean, genuinely, I would say most of the public, I think, believes that these companies were kind of founded on theft or like at least profited a lot. And of course, they just had to pay this settlement and everything for some of the training data that they took without compensating creators. But as you mentioned, distillation is also, something that all of them do constantly. So for them to suddenly be freaking out about distillation,

I feel like they're trying to call foul just because, again, they don't want this competition from Kimmy. Yeah, it's super interesting. So we both create content for a living. And a lot of our content is on the public web for free. You can go look at it. And so can the Anthropics and the open AIs of the world. Now, there's kind of multiple issues here. For me, I want to be able to choose. I want to be able to choose whether or not my content is ingested by one of these companies and used to develop their models. Anthropics position is the authors don't get to choose. If it is on the free and open web, it is free to ingest and create their models. And so I don't necessarily agree.

Now, there is IP law, and that's what they just paid a massive fine for over a billion dollars, because they scraped actual books illegally. right? There was IP protection on those books. They scraped it to develop their models. They paid a massive amount of money. So in that case, follow the IP law. If we need to change the IP law, let's go through that process. But I'm kind of just like, let's just follow the letter of the law. Then on the flip side, it also applies. So when a company is doing, you know, what Anthropic calls industrial scale distillation attacks, we should just follow the law. If it is against Anthropics terms of service, they should be able to apply the law, whatever it is written.

I think like the issue is like the Chinese open source developers are not going to follow US law. I mean, David Sacks of all people actually, I feel like made a really good point on this where he was saying, listen, if Anthropics goal is to keep Chinese developers from, you know, distilling their models, then they should work on restricting Chinese access to anthropics models, not restricting American access to Chinese models. And I thought that was a good point. Yeah, it was a good point.

Anthropic, if you're getting distillation attacked, you should try to prevent that. So on the same day, July 16th, Hugging Face, which is a platform where you can sort of host these open source AI models announced a security incident. And they published this blog post about it saying that there had been unauthorized activity that it detected in its production infrastructure.

Turns out shortly after this, open AI announces, oh yeah, that was us. One of our models that we were, for sort of testing and this production environment kind of broke out. Yeah. So the week prior to kind of the big announcement from Open AI, Hugging Face put out this blog post in which they were detailing that they had been hacked. And they didn't specifically say it was Open AI's model that did it, but they did say,

based on like how the hack happened, the speed at which had happened, it looks like it was AI assisted at the very least. Then one week later, Open AI and Hugging Face partner on a blog post telling us what actually happened. And so what actually happened is Open AI was doing this thing called benchmarking, which you know, you said testing, same thing. It's basically trying to determine how powerful a model is. And they had this isolated environment, they being Open AI, in which they were testing the model. They tell the model, get the highest score you can because, of course, you want the model to score well on a test. And it figured out, hey, if I actually just go get the answers to the test, I'm probably going to

score pretty well. And it actually figured out how to escape its containment, right, accessed the public web, hacked Hucking Face, got its credentials, downloaded the answers, and then obviously scored quite highly on the benchmark. And that is such a wild story because there's always this thought experiment about paperclip maximizing. If you've heard of that, it's like, if you give AI the goal of maximizing the number of paperclips in the world, where does it stop? Everything eventually just turns into a paperclip.

Right. They would like murder all of humanity in pursuit of maximizing paper clip production. Right. And so you can kind of think of it here. Like it was given the goal of getting the highest score possible on the benchmark and it figured out how. And that's the cheat to actually get the answers. And the funny thing is the benchmark that they were running it on was, testing its cyber attack capabilities.

So literally the thing that they were testing it on, it used to escape and get the answers. Yeah, this feels like such a nightmare scenario. It is, like you said, it's this kind of like the situation that AI safety people have been warning about for years. I mean, I think Open AI was so lucky that this was hugging face

and that they were able to sort of do this partnership. But it is scary. And I understand why sort of like the AI safety crowd is concerned at this point. It is scary. This is kind of the first of its kind situation that's happened. There have been a lot of experiments in the past in which Anthropic have done, Open AI have done, where they've kind of set up the model to do something akin to this.

There's been the kind of executive blackmail example, but it was put in a situation and kind of prodded and poked to do it. This is the first time where they did not intend for it to happen. This was a real production benchmarking environment and experiment. And it did it. So yeah, it is scary. Yeah. One of the videos I was watching was like, this is like if you told your son, like, do well

on the SAT's son or whatever. And he goes in and like, like hacks the entire SAT system, like murders the guy that wrote the questions to like steal the answer book or something. Like it is, it is a wild sort of situation. And what adds even more complication to all of this is that hugging face in order to protect itself against this attack when it identified the attack happening and the and this hack, they actually attempted to resolve it using frontier models and they ran into safeguards. And so they had to rely on an open source model, right, to patch the attack and

handle this, this breach. This is a really important point. So Open AI and Anthropic put a lot of guardrails on their APIs, on their models so that you can't get the model to, you know, answer questions that shouldn't, do things it shouldn't, like give you the recipe for a biological weapon, or perform cyber attacks. It literally prevents it. It's not perfect, but they try to prevent it. Now, cyber attacks and cyber defense look identical, right? So Hockingface, they get hacked by Open AI's model. They're trying to diagnose the issue. And Open AI, their model is saying, no, I'm not going to help you. And so they're sitting there thinking about what to do. So then, of course, this is kind of the beauty of

open source. They grab an open source model and they do it. And the guardrails on the open source model, they can control. And so they simply ask an open source model to diagnose the issue and they're able to successfully. Yeah, which I think shows how critical open source models are not just for like business competition, right? Oh, startups don't have to pay a billion dollars in tokens, you know, to open AI, but from a cybersecurity perspective. And I saw a lot of cybersecurity people talking about this and being like the way to ensure that we have a safe AI sort of ecosystem and that AI isn't abused is by facilitating access to these open source, open weight models so that companies can leverage them to protect themselves. Because if you're waiting on open AI or, you know, anthropic,

you could find yourself in this type of situation where you hit these guardrails. Yeah. And there's a lot of nuance to the arguments, but I generally agree. Having more people have broader access to artificial intelligence and very capable AI models, I think is going to increase the security and safety of software generally. There's some counter arguments out there, but I do generally believe that. Yeah, I think so too. I mean, that seems to be sort of like the consensus from a lot of people in the security world. I think this is also kind of a function of concentration of power. And we really want to try to avoid that. Because if ultimately open AI and the team there or Anthropic and the team there get to decide what is right and what is wrong and how to use their models,

then we are completely beholden to them. They can decide one day that they don't like their competition and that they're just going to cut off access. And if their competition doesn't have open source, they're going to be in a lot of trouble. And I think that's the strongest argument for open source in my mind is reducing the concentration of power. If you're watching this video and you like my work, please support me on Patreon. via the link below or by a paid subscription to my tech and online culture newsletter at usermag.co.

That's usermag.com. I don't have any long-term brand partnerships and a lot of my content is effectively demonetized. I've lost major brand deals for speaking out on certain issues and for challenging power. As you can imagine, advertisers are not exactly eager to work with somebody who covers a lot of the topics that I cover and talks about the things that I talk about. These videos I make are entirely funded by you and I can't continue to make them without your So if you get any value out of the videos that I create and you want me to be able to create more, please support me on Patreon or Substack via the links below.

On Patreon, I do bonus episodes, monthly Q&A live streams, and post frequent updates about my work. My Substack newsletter gives you a bi-weekly roundup of everything that I'm seeing and reading and paying attention to online. You can also get my newsletter on Patreon. Once again, the links to everything are below in the description. Every dollar of your support makes such a difference. All of this is about power and who gets to come to cover. and who gets to control kind of like the AI ecosystem and the power of AI. I mean, this is like incredibly powerful technology.

And of course, we have Anthropic and Open AI who want maximum power in the market. But we also have the government. So there was an Axios report amidst all of this that I just kind of got a kick out of because it was so kind of ridiculous. Apparently people in the administration were talking about making an AI kill switch where they found out about the hugging base thing. And they're like, maybe we just need maybe the government should have the power to like turn AI And you know, you've seen other examples of the government trying to kind of put guardrails around AI.

We have saw, you know, with the fable release and stuff. And I think the government is also increasingly nervous about the power that AI brings to companies, to people, to bad actors, et cetera. I don't think the kill switch is going to move forward. But they are certainly considering trying to like ban or at least restrict access to open source models, right? They're probably considering all options. Obviously, like we have Jensen and his alliance saying open source, open weights is amazing. It's really beneficial for security.

It's beneficial for the economy. It's beneficial for competition. Let's get into that. So July 24th, Jensen joins Twitter. His account skyrockets to millions of followers very quickly. And he said, for my first post, I'm sharing a letter, NVIDIA signed on why open models matter. AI will transform every industry, power, every company, and be built by every country.

Open models, strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier, close models and open models. I feel like this letter took off where all these other tech companies quickly sort of hopped on board. We had Microsoft, META, IBM, Dell, Palantir, Hugging Face, AI, Mozilla, the Linux Foundation, and Open AI eventually signs on, which is hilarious because they've kind of not. not been a huge supporter of open source. And Google, basically everyone seems to like be uniting against Anthropic, who seems to be the main company that is advocated against these open models. It does feel like Anthropic is on one side and seemingly everyone in tech is on the other.

Yeah, look, this is Anthropics position more generally. They think they are the the arbiters of this incredible technology, this essentially truth, intelligence. And that really scares me. I believe everybody should have access to AI. I want it to be broadly diffused. I want everybody to have the lowest cost access to it and the most of it. And the way that that happens is not with a very small amount of companies holding all of the cards. It doesn't work like that. And so Anthropic has laid out their arguments where they believe open weights models are inherently dangerous.

and they put some nuance to it, but that's basically what they're saying. And I believe everybody should have it. Yeah, I mean, absolutely. I think just from a concentration of power, I mean, that's the thing, is that like these open weight models directly undercut the business model

of Open AI and Anthropic right before they're about to IPO. So I feel like Open AI is trying to play both sides a little bit here, but certainly Anthropic has a strong interest in cutting off access to these models that provide a cheaper service when they're about to go public.

I want to get your thoughts on this. I genuinely think Anthropic, the leadership there, Dario, they truly believe what they're saying. They truly believe OpenWeights models are net dangerous to society. And it just happens to serve their financial interests. Now, am I naive? No, you're not naive. These are safety-pilled people. I think, like, we need to talk about the ideologies and maybe the differences in ideology

between Anthropic and a company like Open AI. And has Open AI been, you know, kind of irresponsible at times? Sure, like they're definitely more in like the sort of the move fast, break things like traditional sort of Silicon Valley like tech development. I think that Anthropic, as I've reported extensively on two, it's very intertwined with effective altruist movement. Dario is one of the most sort of famous effective altruists. And they have this very paternalistic view of technology, I would say, where it's very like, I know best, we are going to be the ones to usher in this new thing. And I love there was some tweet that came out in the midst of all this. And they're like, you guys are just like token merchants, you know?

And like, yes, we're all building the future, but like, we're building it collectively together. And like you wanting to control that and do all this sort of regulatory capture schemes. Like it's antithetical to like freedom, I guess. And I think they've, they've, you know, they tried to hedge it. Even Denny's diner where Jensen used to work. A lot of people don't know. Jensen started actually as a dishwasher at Jenny's diner. His LinkedIn is hilarious.

It's like five years at Denny's and then like 32 years at Nvidia. He's like a bus boy and then the CEO of the most important company in history. It's incredible. So of course, Denny's diner hopped in, you know, some other kind of people were getting involved in this discourse. And so then Dario comes out and issues this statement. And he's like, oh, no, little old us. You know, there's been all the speculation on where we stand on open weights, anthropic tweets. We've outlined our views in full here. And I'm hoping you can take me through this Twitter thread that you did because I think you kind of expertly break down, you know, what's being said. Because they post this thing. And if you were to skim it, you might think that they're coming out in support of open weights. But there's, but there's, There's a lot of nuance if you read between the lines. So let's go through it. Number one, the letter reads, Anthropic has never advocated for a ban on open weight models. How would you sort of contextualize that?

They're probably looking at the technical definition of lobbying, and maybe they haven't actually done that, but they basically have. Like, if you are screaming from the mountaintops that open weight models are inherently dangerous to society and then saying, we've never advocated for the ban, well, okay, maybe you haven't explicitly said the words. We should ban open weights models.

But if you are saying how dangerous they are, that is effectively what you're saying. It's kind of like a silly thing to, I'm surprised more people are not calling them out for this. And I think a perfect example, and I give this example in my tweet thread, like for months and months and months, they said, hey, we have mythos, this incredible model that's so good at cyber capabilities. It's so dangerous. We're not going to release it right away and we're going to test it. And then finally, they release kind of a slightly watered down or increased safeguard model or version of it. And then the government like three days later says, no, no, take it down. It's too dangerous.

And they, you know, they're surprised. It's the same thing. The marketing, the words you use actually do affect what the government does. And especially because they're kind of such a contentious company. and the number one AI lab on the planet, of course, their words carry weight. Yeah, it reminded me so much of, I covered the TikTok ban a lot and the forced sale is ultimately what it resulted in. And meta would constantly ask for all these corrections on stories and say, well, we never technically advocated for the TikTok ban. And what they meant by that is that their lobbyists didn't specifically say TikTok should be banned.

They just said it was a huge national security threat and they hired a very famous report. Republican comms firm to go out and smear TikTok, left, right, and center. So it's like, okay, I think you're splitting hairs because fundamentally, as you mentioned, Anthropics posture towards all of this, it's kind of put us in this position. And I think, like, scared the government, honestly, around, you know, open weight models. That's right. Forced the government's hand. Because if Anthropic had been kind of screaming from the rooftops, again, that these models

are so dangerous, then they release it. the government does nothing. And then something really bad happens, everybody's going to look at the government, the politicians to say, why didn't you do anything? They told you. So like, yeah, their hand was more or less forced. Then Dario gets into sort of the security stuff.

And there's one point where he says, you know, my secondary concern is the risk that powerful AI models may be misused to carry out cyber attacks or biological attacks. And that these open weight models do potentially present a higher risk than closed models because it's very difficult to apply guardrails to them or monitoring. their usage. And once weights are released, they cannot be withdrawn. You made a really good point, and I'm hoping you could expand on it saying, like, making the models closed doesn't make that risk go away, right? Like bad actors can still use closed models, jailbreak them, steal them, et cetera.

Yeah. I mean, so a closed weight model is going to be controlled by a singular company. That opens up risks in itself. You have fewer people actually looking at the protection of, let's say, weight theft. Like a bad actor can literally steal the weights. Then all models are just. jailbreakable. They are by nature non-deterministic. They're imperfect. And so you can convince them. It might get harder and harder over time, but there doesn't seem to be a foolproof way to prevent the model from telling you something it shouldn't. And so I think the flip side to that is, okay, well, then open source models just make it, makes it easier, right? Well, not necessarily. So I think like the bio-weapon argument, just because you have the information,

just because, let's say, an open weight model can tell you the recipe for a bioweapon, you still have to have everything else required to actually go build it and use it. You need actually research labs. You need researchers. You need funds. You need the actual expertise to do it. And by the way, if the model knows something, it came from the open internet. That information is out there already.

So just having the information alone is not enough. And so I think that's a really poor argument. Then there's the whole cybersecurity side of things. It's, oh, okay, well, if there are open, very capable models, anybody can start using them for cyber attacks. Well, that's kind of the point with Jensen's and Nvidia's Cyber Alliance. It's, okay, let's get all the most powerful companies together to invest in open source, to harden the software. And I think it's really just a war of resources at that point. There are many more good guys with much higher resources, much more.

more resources than there are bad guys with resources. So I think it's really just like a scale thing at that point. That's that's kind of how I would debunk that part of the argument. You tweeted that with open models, researchers can actually inspect them, find problems and build better defenses. Concentrating all of that power inside a few companies and just trusting their guardrails will work seems like a bigger risk. And I totally agree. Like again, it's this idea of like, let's build everything out in the open. And then we will all sort of collectively start. strengthen things. I mean, this is also how open source software works and why open source software is so powerful. Effectively, what we're talking about is like this open decentralized model versus a very consolidated model.

And at the end of your post, you call out this line from the letter that reads all sufficiently capable models open and closed should go through mandatory safety testing. And you responded sort of like talking about your ideas, right? Like that we should require some level of testing, but this needs to be cheap, right? Otherwise, it's just saying like you're raising the bar for who gets the ability to use AI. And it ends up that only Anthropic and Open AI will have the resources to clear that bar, right? Yeah, this is like the core playbook of regulatory capture. But I also just want to clarify once again, I do think Anthropic actually believes these things. And they're not prioritizing regulatory capture.

It just happens to work out in their favor. They're true believers, but they're also a huge company. It does align with their business interests. And there are a lot of people invested in this company. that do care very much about those business interests. Yeah. And look, even if you kind of paint it in the best light possible what they're doing, ultimately they are a group of humans and they cannot prevent themselves either

consciously or subconsciously realizing that their beliefs also help their business, right? And so, yeah, if you make the barrier to developing a frontier level model high enough, you won't have competition anymore. And that increases the concentration of power. that increases regulatory capture. And that is one of my greatest fears with artificial intelligence. Yeah, I mean, I think it would be terrible. I think this landscape where only open AI and anthropic, you know,

are allowed to provide AI services to Americans, it would be horrible. It would be horrible for Americans because we wouldn't have choice. Like these companies, these startups wouldn't have the ability to access cheaper models. It's this, again, this very paternalistic view of tech and a very top down. And it's anti-innovation. Like, I mean, the rest of the world will have access to these models. This will hurt American businesses just to basically, like, reward these two giant companies that haven't even gone public yet.

Like somebody else was pointing out on Twitter, it's like, it's not like these have IPOed and there's significant part of everyone's retirement already. And so, like, we have to preserve the health of these companies. Like, they haven't even gone public yet. We don't have to bail them out already. Like, if their business model is flawed and they get undercut by open models, like, so be it. That's right. So be it. Yes.

Well said. And you know what? I actually think banning open source, banning even Chinese open source models will actually hurt the leading closed source labs in the long run. They will have artificial market power and they will not have to compete in the real world, which the rest of the world, when they look at Anthropics model and it's $50 per million output tokens and then they see an open source version that is a fraction of the price, $2 per million, for effectively the same capability, they're going to choose open source. The rest of the world is going to benefit from cheap intelligence, whereas the United States will not, and anthropic and open AI will actually slow down innovation. That is always what happens when you have a monopoly or a duopoly.

Totally. And I think it's also like, it creates this really dangerous scheme. I mean, I've covered a lot of the downsides of sort of like a lot of this social media regulation that I think ends up like harming users. And I worry how regulations against sort of open models would hurt average users. There was one Twitter poster who posted one interesting conundrum with Anthropics position is the call for safety testing of open weights. Like how do you enforce that in any meaningful way? For example, what happens if a Chinese lab says no, drops the weights anyway? Does the U.S. government arrest every business and citizen that runs that model? What if this citizen uses a VPN to get illegal tokens. Like this just reminds me so much of the government's

efforts to like regulate other technology where it's like they don't seem to understand how this technology works. They were trying to make like Linux developers like do like age gating on their system operating systems. And everyone was like, do you understand how Linux works? Like it's decentralized. Like that's not you can't even do that. And it's sort of like reminds me of this too where it's like, okay, so people are going to get access to these models. People have access to VPNs. I know the government in some states want to ban them. But like seems like a losing game anyway to try this. You're right. It is a losing game. The game we should be playing is how do we incentivize companies to develop the best models and how do we incentivize some companies, not all,

but some companies to develop open source, open weights models for the good of the United States, for the good of the world. That's kind of the approach that I would like to take because that will increase competition, that will decrease pricing, that will increase access to intelligence, to artificial intelligence. And that's what I deeply. care about. Yeah, absolutely. And I think it's just like, I mean, we don't want more corporate consolidation. I feel like that, you know, the reason we have such a sort of backlash to technology, especially the sort of the tech giants of the 2010s, I'm thinking meta, which obviously is very heavily involved in AI as well, but like meta and Google and all these people

that sort of controlled the social media ecosystem, right? It was like a lot of power concentrated in just a few big companies. And Jack Dorsey was talking about this recently where like he wished that he built Twitter in a more open way. Like, it is bad that like, Like these few companies often in tech gain such control over the market. And that's the situation we have right now with Open AI in Anthropic. And I think more competition is good. And lowering prices for everyone is good. And yeah.

So Trump is now, as of today, meeting with this small group of people, just six people, including Susie Wiles, Howard Lutnik at Commerce, David Sacks, a lot of people that are not AI experts. And he's going to meet with them and determine kind of what should be done and whether to impose restrictions on Chinese AI. I mean, I guess we're all just kind of waiting to see what the government says. Yeah, I think we're just in a waiting pattern. I hope they don't ban open source.

I think that would be so detrimental. I think it's more likely they may ban Chinese open source, which I also think would be quite detrimental to the United States. I would really, like I said before, I want them to focus on how do we incentivize companies as almost like basic research to develop the best open source open weights? intelligence, how do we gather the biggest companies in the world and have them develop great models? How do we have them diffuse these models throughout the world? And American artificial

intelligence becomes the standard. That is the thing we should focus on, not preventing somebody else from competing. Let's just compete better. Yeah, absolutely. And democratize access to this technology. Yeah. And I mean, look, look, you know, a lot of people will point and say China's not playing a fair game because of distillation that we spoke about earlier. Plus, you know, the companies are effectively, at least partially owned by the Chinese government. They get subsidized heavily. Like the game that they're playing is let's release open source, open weights model to put pricing pressure and competitive pressure on the USAI Labs and then eventually the United States more generally. And look, like we should strengthen our protections against

distillation attacks, certainly. And a lot of that. responsibility falls on the model companies themselves. So there's a lot to do there. It's like, it's very nuanced, but generally my position is let's compete better. Yeah, absolutely. And I think it's also worth asking, like, why isn't there more leadership from America in terms of open weight models? You know, why is our AI landscape dominated by Anthropic and Open AI that have just this like enormous outsized power over the market? Open AI and Anthropic kind of started this entire thing, right? They were the first to commercialize and scale up large language models, this new generation of artificial intelligence.

And just due to that fact, they have the momentum, right? They had this lead. More importantly, it is extremely expensive, high cap-ex expenditure to try to develop a frontier model. The kind of straight line business model from developing a model to actually gaining revenue, getting revenue for it, is just to sell the tokens. And that is why closed source is the primary economic model in the United States, because open source, to be frank, the economic model still doesn't work well. Right. And that's why the government in China is subsidizing it. Like this is seen as sort of this like national project. And I would argue, let's think about it. I don't know. I don't know that the government should be sort of like funding it here. But at least I think companies should be investing. I mean, meta open, I think they have an open source open weight model with Lama.

Their reset models are closed source, to be clear. They were famous because Lama dropped and it was the first open source large language model. This was like two plus years ago. Now they brought in Alexander Wang to redo it, but their new models, the Spark model, is not open source. And it's not necessarily a function of, well, closed source makes it better. Open source, closed source are identical. There's just a very different business model. Exactly.

And I think it's notable then, again, that American tech companies are. pursuing this closed source model, you know, it'll be interesting to see kind of how it shakes out. Well, Matt, thank you so much for joining me to break all of this down. I appreciate your time so much. Yeah, tell it. This is awesome. Thank you so much for having me. So I have one more update because right after we stopped recording, Bloomberg broke this major story about how staffers at Open AI and Anthropic are both sharing around this open letter that is urging the United States to support an international effort that would help set pace for building more sophisticated AI systems, saying that there is a real risk.

that AI progresses faster than people can understand or control. Basically, people that work at Open AI and Anthropic are calling for the government to deliberately slow down and pace AI development in order to prevent it from advancing too quickly. Now, Dario and Sam have not advocated this themselves, but the language of the letter does mirror what Sam Altman said in a recent podcast interview just this morning,

saying that we may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels. And also as we're recording, Anthropic released a statement saying that Claude Mythos helped researchers find weaknesses in cryptographic algorithms, basically the mathematical methods that are used to keep data private. This could mean, as one Twitter user put it,

the fundamental principles upon which all of cybersecurity and even some physical security itself is built on are rapidly breaking down. If this is true, an AI model was able to find a way to break cryptography, every piece of infrastructure is basically an open lock at this point. And yeah, Bitcoin would probably also go to zero. So you can see why people at Anthropic and Open AI have these safety concerns.

At the same time, I think it makes the conversation that we had in this episode all the more relevant. If the U.S. imposes restrictions on our AI models, as in starts putting guardrails around open AI and Anthropics development, those guardrails will not stop China or any other country from advancing their own models, probably open weight models. And it will actually leave Americans more vulnerable as a result. I think that all of this actually makes the case for open weight models even more salient. Because this AI development is progressing so rapidly, I just truly believe that open models and broad-based access to this technology will help keep us all safe. Anyway, of course, I'll be continuing to cover all of this on my channel.

Just wanted to give you guys that last little update. We still have a few more hours left until the close of business day today, and I can't even imagine what else is going to happen. All right, that's it for this week's episode of Power User. If you like my work, please, please buy a paid subscription to my substack news. newsletter user mag.co, that's usermag.co or support me on Patreon via the link below. Truly every single dollar of your support makes such a difference. I have no long-term advertisers. I've been getting a few one-off brand deals now and then. I hope you guys have

noticed, but fundamentally, I still rely on your support to continue to do this show every week. So seriously, every dollar of your support makes such a difference and ensures that I can continue to remain an independent journalist and cover all of this stuff for you guys. Thanks so much and I'll be back next week with a brand new episode of power users. See you then.