Anthropic Cried Over Stealing Claude, The Internet Laughed

Anthropic discovered Deepseek, Moonshot, and Miniax used 24,000 fake accounts to extract 16 million clawed conversations for model distillation. Let's take a look at that command AI. We've identified industrial scale distillation attacks on our models by Deep Seek, Moonshot AI, and Miniax. These labs created over 24,000 forgent accounts, generated 60 million exchange with claude extratinics capabilities to train and improve their own models. This is exactly what distillation does. And anthropic go on to say distillation can be legitimate. AI labs use it to create smaller and cheaper models for their customers. But foreign labs that illicitly distill American models can remove safeguards feeding model capabilities in their own military intelligence and surveillance systems. We'll get into that. These attacks are growing in intensity and sophistication. Addressing them will require rapid coordination and action amongst industry players. The internet did not respond so well to this. Um there's a lot of I mean the community notes basically outlining that anthropic uh are set settled a $ 1.5 billion dollar lawsuit for pirating 7 million books for from shadow libraries to train Claude faces $3 billion suit for torrenting 20,000 songs. They built their models on stolen creative work then complain when other others extract their outputs. And you can go on it. You're trying to kidnap something that's rightfully stolen. The internet let loose really on this idea. Now, never ask a woman a man's salary and a company where they got their training data. So good. I will say like it's not quite the same. Like they're not it's not steal like I I'm an educator. Kabaza is an educator. Like you in your everyday life are an educator, right? you learn something, right, and you interpret that information and you forward it on to someone else. Like, it's a transformation layer on top of this. So, it's not quite the same as they're not it's not talking about stealing in that sense, but it's still [ __ ] hilarious that uh the internet responded this way. And if we go through some of the actual attacks here. So, Deep Seek, they discovered over 150,000 exchanges. Now, I watched uh Theo's video here, who and he he rightfully has a has a voice in this. He owns T3 chat as you might know and he uh shared some of his stats some of the things that uh some of the figures that people are that's happening right and he said that 150,000 exchanges is like a daily thing on him on his platform and in fact to be more specific because actually tool calls are also exchanged with so if you get all of that data it's so much more than 150,000. And as you can already see there, um reasoning capabilities, uh across a diverse task, rubric um rubric based grading tasks that make Claude functions as reward model creating censorship uh safe alternative to policy sensitive queries. Moonshot over 3.4 million. This is this is more this is starting to get a little bit more um suspicious. Um operation targets agentic reasoning and tool use coding data analysis computer use computer vision and uh moonshot Kimmy models employed hundreds of fraudulent accounts spanning multiple pathways varied account types. Now, I saw a comment here saying, you know, what does this say about your your privacy that you um that you'd speak about that is there such thing as privacy if you're able to track because I I think they mentioned the metadata here. We attributed the campaign through request metadata, you know, where where's the line there of like privacy versus like what is just, you know, is it is it general location? What is this? Is it is it the email that was used? Not too sure. But yeah, someone questioned the the the legitimacy of privacy when it comes to this. And then Minia Max scaling uh uh over 13 million exchanges, operation targeted, agentic coding and tool use and orchestration. And I have to say the Miniax one is slightly more believable. Uh I had my suspicions. I made a video on my own channel. I had my suspicions about the similarities in performance between Miniax and Haiku. Um, and also that it's been tuned to clawed code. It it seems like it's it's it's heavily uh directed towards clawed code. And then when the target the operation targeted uh features are tool use and orchestration and agentic coding, you have to wonder. But I don't know. I'm I'm just a guy with a camera, right? I I have no idea. But this one's slightly more believable. That all being said, again, Theo shared his numbers with us on on his um on his on his channel. If you can get a model from these sorts of numbers, if Miniax can get a model with 30 million exchanges, then we're doing something catastrophically wrong with training because this really is not all that much. I think again Theo shared numbers that came similar to this. I think you know that was the the the lie was 150,000. it was more in the realm of 3 to 4 million exchanges something like that. But if you can if you can distill a model with this level you know um these numbers are more akin to something that is um benchmarking which anyone's kind of allowed to do. you you benchmark against other models to make sure that your you know that your model is good, right? I just watched a video earlier on about PewDiePie who's trained his own model and he's benchmarking them against Chad GBT and all and Gemini and these things. So it's these are the sorts of numbers we're talking there. So yeah, why Anthropica are doing this? I don't know. I don't doubt to be honest that this happened personally like just because the numbers don't line up doesn't mean to say that they aren't training on their you know uh using their model to train but it could just be one of many as a start realistically and this is an interesting um way to go about it and they haven't really mentioned it in this article too much. Let's do uh military. Okay, here we go. Foreign labs are distill American models that can feed these unprotected capabilities into military intelligence. And this is true, but I I feel like this is more of a a a sk a fear tactic to try and get people to like be like, "Oh, this is actually a bad thing." Cuz ultimately, going back to the tweets, people being upset that, you know, that Anthropic is stealing, like people just don't care. Like, why should we care? And it all comes down to this military training and and this and that. Saying all that, what they're going to do about it detection, they built some um things. And actually, just on that note as well, they noticed as cuz uh Miniaax got released just shortly after Opus 4.6. And they noticed once Opus got released they noticed a complete shift in the attacks happening like from that same sort of data sources and stuff like that. So again I think I don't I think it is happening but do does can that sort of can those sorts of numbers create a new model? I don't know. Um intelligent sharing so obviously they're going to be sharing this with open AI and all the rest of it and and cloud providers as well. uh Vertex um whatever Amazon's one is access controls. So they they you know they have their right to do this. They don't want their model getting used. No company can solve this alone. So you know the jury is out when it comes to um what this means. How terrible is it? Do we even care? Are they lying? But it is something that uh that Anthropic have been crying about and we're here for it. So that's that. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.