I need to crash out. I did not think this was going to take so long. And I genuinely really miss working with Fable. It's kind of crazy we can get a taste of something that powerful and just have it taken from us with no promise of if we'll ever get it back, much less how long it will take. We have had some promises throughout this. Like when the official press release came out from Anthropic, they promised us that it would take 24 hours for them to share more details, which never happened. It has been 11 days. That's a lot more than 24 hours. Or the managing director of International for Anthropic statement that they are very confident that in the coming days the models will become available again, which was 6 days ago. The lack of statements from Anthropic doesn't mean there isn't anything going on. from other members of Congress trying to bully the White House into giving them more information to actual lawsuits from customers of Anthropic who were using the fable model for their work, suing the government for these restrictions to crazy leaks about how this is affecting other labs with GBD56 potentially being delayed until mid July. This is unprecedented and it's a huge shift in how we need to think about our tools. So, where is this all going? Are we all going to have to move to openweight models so the government can't restrict us? Is this the end of AI development as we know it? Is this the end of US dominance in AI? Where is this going? It's really hard to know, and I'm not going to speculate or guess, but I'm going to do my best to explain where we are now and what the impact is. All that said, if I'm going to have to run models locally, I'm going to need way more GPUs, and I can't afford that right now. So, let's take a quick break for today's sponsor. You've already heard me talk about today's sponsor. It's Work OS. They support millions of users for free, and it's a great platform. But, I'm not here to talk about you or your millions of users. I'm here to talk about agents because agents have a really rough time signing up for services and that's why work OS put out a new open standard OMD and the standard is actually really really compelling. It allows for agents to sign up for services on behalf of a user. The agent can register a full account by itself or you can make an easy way for the agent to register and get claimed by a user. Both of these are built into the standard. There's a reason why companies like Cloudflare, Firecol, Rescend, and more are already working closely with work OS on the spec because it's so helpful to let agents go sign up and do things without having to go clear some weird form or hope computer use and codecs works properly for once. All of that is nonsense and all of that nonsense existed because agents had to go down the same paths users did because we built our O platforms around people. And to be clear, that wasn't a mistake. people should be the core of O on most services, but getting an agent to fill out a form and pass a capture correctly is not the path we should have ever taken. And now we have a simple standard that makes a lot of sense, but admittedly would be hard to add to most off providers. Hopefully, you listened to me in my previous work OS ads because if you did, you're in luck. Enabling OMD in your work OS apps is literally one click. If you want enterprises and agents to use your service, look no further than soy. Oo. Let's start from the top. Fable 5 and Mythos 5 officially dropped on June 9th. They were incredible. I put out a video showcasing just how cool they were. I was actually live streaming showing off some of the cool things you can do with these models when the ban occurred because the ban occurred on the 12th, 3 days later. I was running two $200 plans just to maximize my usage of Fable cuz I was so impressed with it. And then it was gone. And I will be real with you guys, I honestly assumed the ban would last the weekend and we would get it back the following Monday. It's been 11 days since. We've passed two Mondays since. We still don't have the model back. We don't even have a timeline for when we'll get it back. The alleged start of the drama didn't actually come from the initial reporting of jailbreaks. It came from a company called SK Telecom, which is a South Korean company doing internet stuff. SK Telecom's a really heavy user of anthropic models. They've actually went as far as to invest over hund00 million into Anthropic. It was alleged that when they got access to Fable and Mythos that they were reselling it to customers in China who are banned from having access. Earlier in June, shortly after Anthropic announced Project Glass Wing's latest expansion, the White House asked Anthropic to revoke SK Telecom's access to Mythos, according to a person close to the AI lab. The company immediately complied, sources told Wired, and the US government did not threaten to put export controls on the model at the time. This is really important. Enthropic was told by the White House to ban this company and they did. But this definitely put the government on high alert about the capabilities of the model and the desire other countries had to access it. And it is a known fact that China is trying to do whatever they can to access these models in order to get enough data to distill and make their own things of similar capabilities. And this is a huge thing for Anthropic to just blindly comply with. As I mentioned before, this company invested a hundred million into Anthropic back in 2023 when that was like a meaningful percentage of the company. And part of that was so they could do a commercial partnership in order to make models better at Taylor communication work. Anthropic hard cutting them off is kind of crazy, but that's how hard they were trying to make sure the US government knew that they would do what they were told to do. And it seems like it wasn't enough because these reports of a jailbreak that came from Amazon and made their way all the way over to the government resulted in the ban. Recent reports have indicated that Amazon has actually tried to contact Anthropic about this and Anthropic refused to respond. Why would Anthropic not respond and not acknowledge these jailbreaks? We could debate the reasons why all we want, but it really comes down to this paragraph from Anthropic's announcement. The government has only given us verbal evidence of a potential narrow non-universal jailbreak which essentially consists of asking the model to read a codebase and fix software flaws. Our understanding is that one potential jailbreak was shared with the government. We reviewed a report that we believe is the basis of the government's directive and validated that the level of capabilities displayed there is widely available from other models including OpenAI's GPT 5.5 and is used every day by the defenders who keep systems safe. This is really key. It seems as though the jailbreak wasn't like, oh, if you use these special characters or these special phrases, you can get the model to hack things when otherwise it would say no. It seems like the behavior is actually intended where if you have a codebase and you have a model and you ask the model to help you fix the codebase and improve it, it should do that. But of course, as I've mentioned in many videos covering security stuff, having the code that fixes a bug is also kind of giving your models a guide on how to exploit the bug. And you can take a very dumb model and hand it a diff that patches an exploit and ask it to recreate the exploit or show you how you would take advantage of it and they will. Especially now that we have openweight models like GLM 5.2 which are getting close to GPT55's capabilities. The possibility to take advantage of these hacks and take advantage of the smartest models finding fixes to open source work in order to hack it with openweight models is a very real pipeline we do need to be concerned about. All of that said, the best fix is to secure the software we're worried about to use models to find these exploits, patch them, and ship those patches ASAP before these capabilities are more widely available. And this is what Anthropic was doing with Project Glass Wing. I also think the dissonance here is a big part of why Anthropic is struggling so much. They are mostly used to talking to people in the AI space. Anyone working at Anthropic spends the majority of their time talking about, thinking about, and using AI. They know the purpose of these tools, especially the developer focused stuff, is to work in a codebase and solve real problems in those code bases. And I can guarantee that pretty much no one in anthropic would be interested in shipping a model that couldn't fix a codebase. And that's why they didn't see this as a jailbreak. And that term jailbreak has been absolutely destroyed as a result of all of this because what the government sees as a jailbreak, Anthropic sees as intended behavior and the world sees as a very confusing scenario. But as Anthropic mentions here, there's lots of other models that can find exploits the same way as they call out here, including GPT 5.5. This call out might have been a bit dangerous of them though because from what I've been seeing, lots of other labs are now scared that their model is going to get banned or know your customer as well. Because remember, this isn't a formal ban. The government's restriction isn't that they can't use or sell the model. It's that they are not allowing any foreign national. So people whose citizenship isn't with the United States, regardless of where they are, they cannot use it. Even including anthropic employees. This means that people like Cararpathy have been restricted from access to the model. So why haven't they done that? Why don't we have the model back now? I've seen a lot of misinformation flying around. It's part of why I was motivated to make this video because a lot of people seem to think that Anthropic's going to put out the know your customer like verify your identity and now you can use the model thing. They can't really and there's a lot of layers to it. The biggest one by far is that they also expose this model over API. So if I want to integrate Fable into my product, I would now have to do my own verification of my users to make sure they're US citizens too and somehow forward that along to Anthropic. That type of data privacy chain is not really viable and I understand why they don't want to implement it. They could hypothetically force you to off through this to use it with cloud code and then also be really aggressive about making sure people who aren't you using it get banned really fast. That's not going to be a reliable enough solution. They know that. You know that. Let's be real. So, they've decided to keep it on hold until the ban is lifted, which has the terms of they need to fix the jailbreak. Initially, this was not going well because Anthropic just seemed to not really know how to talk to the White House. But then they started sending employees there in person to negotiate. And the negotiations seemed to finally be going well. It went well enough that in an interview, Trump actually told the interviewer that he doesn't see Anthropic as a national security threat anymore. that he did, but that has changed. Thankfully, Trump doesn't seem interested in using the Defense Production Act, which is a wartime act that allows the government to take over manufacturing in order to produce things we need to defend ourselves, like weapons or [ __ ] tanks and whatnot. He was sincerely considering using that to do a hostile takeover of Anthropic as a business. He calls out that he has the power to use a lot of things, but he's not sure he has to do that. So, let's hope that's the case. You can also see Enthropic kissing ass here because they've learned that they have to do that. They have to be liked by the admin if they wanted to get out of this. To which they said, "We are grateful to the administration for their ongoing partnership and working to get this matter resolved as quickly as possible. We remain committed to working alongside them towards our shared goal of protecting critical infrastructure and making sure the US leads in AI. But according to leaks, their progress has stalled. They have not been able to get any headway and it seems like the government is just kind of not moving and has lost interest in this issue which means getting unbanned is going to be a lot harder. And according to these leaks they're going as far as iterating on Sonnet 5 and like getting a new model out that is a five era model but much cheaper and dumber in order to just get something out for these customers who are hurting. And these customers are hurting a lot. They're hurting so much that they are suing. This is a 43page lawsuit that has been filed specifically against the United States of America, the Department of Converse, Howard Lnik, who's the guy who okayed this ban and pushed for it, Jeffrey Kesler, the Bureau of Industry and Security, Executive Office of the President, and 10 additional defendants. This is a challenge to an unlawful directive that exceeds every source of statutory authority on which it could conceivably rest. On June 12th, the federal government ordered Anthropic to immediately disable two of its most capable artificial intelligence AI models, Fable and Mythos 5, for every foreign national on Earth. I love Capital Earth as a call out here. This is quite a legal complaint. The order to reach anthropic by letter from Commerce's BIS gave the company 90 minutes to comply under threat of prompt criminal and civil penalties and purports to rest on the export control authority that commerce and its BIS administer. Within hours, hundreds of millions of users, including citizens of the US's closest allies and anthropics employees, lost access. The government is not authorized to do what it did. I will say the sheer volume of M dashes here does suggest AI helped write this. Probably not Fable, though. This lawsuit isn't from Anthropic. This is from Legion, a customer of Anthropic. They are a US-based AI native litigation tech company that builds drafting and case management tools for attorneys on top of Frontier AI models. Legion is a commercial customer of Anthropic with a contractual right and license to access and use Fable 5, the model, which was integral to building and operating their platform. Legion software development team includes Canadian nationals working remotely from Canada, a five eyes intelligence partner and a country group A5 nation under the defendants's own export administration regulations. When the directive took effect, Legion lost the latest tools at the center of its development instantaneously. They have three causes of action here. All of these are claims that based on current law, the government doesn't actually have the right to ban commerce in the way that they did here. The use of the International Emergency Economic Powers Act is not valid. According to this lawsuit, allegedly the reason is because it violates express prohibitions and mandatory prerequisites that the Congress put into that statute. Specifically, the president is not allowed to prohibit the export of information andformational materials in any format or medium of transmission. It is entirely withheld in this act, the act that was used to enable this ban. As I hinted at earlier, things are bad enough here that there's a bipartisan Congress group trying to push the White House to explain what the hell is going on. Congressman Sam Licardo, Jay Aubber, hopefully I got that right, Ted Louu, and Scott Franklin, two Dems, and two Republicans are all requesting that the Department of Converse provide greater transparency regarding the June 12th decision to impose export controls on Anthropics Claude Mythos 5 and Fable 5 artificial intelligence models. The lawmakers raised concerns that the action could set a significant new precedent for Frontier AI regulation. This already appears to be the case. It seems as though labs that are able to release models of this capabilities have been really quiet. There have been almost no posts from the OpenAI guys for a few weeks now. And I do honestly suspect it's because of this. And I'm not just blaming the government here. I would argue that Anthropic's fear-mongering has put us in this position. The ways they've presented to the world that AI could destroy us all has put people on high alert which has resulted in this type of high response. But there's another issue that I'm scared about here. And that issue is here. According to artificial analysis, Fable 5 is the best model ever intelligence and capabilities wise, scoring a 60 when the next highest was Opus with a 56. This is also probably even better than shown here, but they had a bunch of instances where Fable wouldn't respond and they got fallen back to Opus. This is a combination score between Fable and Opus, but it's still unbelievably high. Historically, the order of scores here is anthropic or open AI model. Then you have whichever isn't the one before it. Then you have Google or another OpenAI or another anthropic model and you keep going down a bunch and then eventually you'll hit open weight models much further down the chart. But you might notice something different here. GLM52 isn't that far down the chart. GLM52 is a really good model. They shouldn't have called it 52. Honestly, they should have called it GLM6 because it is a significant jump in the capabilities we would expect from openweight models, especially for code work. It's the first time I've used an openweight model and like didn't feel as though I was missing out too much compared to the current available options. That said, it's mostly caught up to like 54 and is getting there with GPT 55. Fable was another tier above. So, what's scary about that is we now have a line that says if you're too tall, if you're above this point, you're not allowed to ride. You're not allowed to sell that model. You're not allowed to put it out in the public without crazy restrictions and all of these expectations that the government has now set. But something just slightly dumber I can download on my machine and use for whatever the hell I want. That is a weird structure. That's a weird system. And it's going to have massive impact, especially around who wants to build here. If in China you can just build whatever you can steal all the data you want in the world and put out an openweight model anyone can download and use for whatever they want, go nuts. You wouldn't download a file the government said is illegal, right? I will tell you with 100% confidence, Department of Commerce, you are not getting me to delete any of my models from my computer. When I have the weights, I have the weights. You're not getting them back. You can't export control what runs on my GPU. Good luck. Have [ __ ] fun. At the same time though, this means that the models that have the fewest restrictions are the ones that have the incentives that we're trying to prevent from dominating. I personally don't want models that are censored by the Chinese government to become the norm because models censored by the US government aren't accessible at all. That's a really bad precedent and it's being set as we speak. It is at the point where both Republicans and Democrats are confused and frustrated and trying to get more information out. If you're wondering why we're not getting better official statements out of the government, it's because the government's stance here is pretty absurd. Pete Hedgeth did make this statement the day after the ban went live, saying, "Three months ago, the Department of War kicked anthropic out of our building forever. Every passing day proves why that was the right move. Even though they were literally using Mythos just a few days before this ban, they were using it in highstakes military operations, as the community note has called out. For what it's worth, it does just kind of seem like Pete's thing is stepping all over the things he's supposed to be allowed to do and going way further than he actually can. The feds have came out and said specifically that his statements are not actual final agency actions, he just likes talking [ __ ] But in a world where he is talking [ __ ] and informing the White House and the Department of Commerce on what should or shouldn't be banned, we end up in scenarios like this. And this is no longer just like, oh, fuckropic. This is now going to impact all future releases. Pretty much everyone I know working on or with AI is scared right now of what this means long term. Every time a new advancement occurs, we have to worry about it being taken away. Imagine a new programming language or a new framework comes out and we get used to building with it. It makes it so we can build better things faster and then all of a sudden poof, it's just gone. We've never had to deal with this before as a software development industry. We've had to deal with this before as a technology industry even. It's kind of crazy that this is a worry we now have to have and the companies who are developing these things have to worry about as well. I could end on my own statements, but I'd much rather quote the congressmen who are pushing back on these horrible decisions. Regardless of the specific circumstances surrounding the individual model, the practical effect of such an action appears capable of substantially restricting the distribution, deployment, and use of advanced AI models, including within the US, and may establish a precedent with significant implications for other developers, researchers, users, and investors throughout the AI sector. Given the importance of these systems to the US economy and national security, we as members of the Congress have a strong interest in understanding the basis for actions that could materially affect the availability and deployment of frontier AI capabilities. We are particularly interested in understanding the standards criteria and evidentiary thresholds the department applied in reaching its decision as well as how those standards may be applied going forward. They've requested a response by June 26th, and believe me, I will be covering that when it does happen, if it does happen. I cannot fathom that this model is still banned, that we haven't made any progress or gotten any real updates beyond all of these attempts to get more information. And I can't believe I'm out here defending Anthropic again, because in a lot of ways, this is their fault. But this is a really bad precedent, and I'm scared of what it could mean for the future of AI. I'm gonna go back to staring at the PRs that Fable made, wishing I could ask it more questions and uh hopefully ship something.
QUELLEN
CR
Codekiste Redaktion
Automatisierte Content-Kuratierung für tech-news.