Transcription
There's something deeply weird happening inside Chinese AI models. Models like Deepseek and Kimmy think they're clawed. Now, I know that sounds crazy, but stick with me. It's a really interesting story.
A few users started noticing this pattern. They'd be mid-conversation with Deepseek, and when they ask a casual question like, "What model are you by the way?" Deepseek would respond back something like, "I'm Claude, an AI language model built by Enthropic, a San Francisco based AI research company."
Now, obviously, this isn't right. You got Deepseek saying that it's clawed. And this was becoming so common that someone even filed a bug report on Deepseek's official GitHub titled giving name as clawed when asked, "What is your name?"
Now, why does this happen? Why does a model like Deepseek, which is a pretty well-known model, saying that it's clawed? First, we got to understand how these models are built. You might already know this, so I'll keep this short, but basically, you start with massive amounts of human written text from the internet. You clean it, label it, and then train a model on it for months. This will burn through millions of dollars in compute just to get a base model.
Now, that's part one. Part two is after you have a base model, you perform fine-tuning to teach the model to be a helpful, safe assistant. And that process alone could take years. So really there's three factors here. Time, money, and infrastructure. You need those GPU clusters as well.
So these Chinese labs figured out a shortcut. Why go through all that when Enthropic already did that for Claude? There's a technique called distillation where a smaller student model is trained on the outputs of a larger, more capable teacher model. Meaning instead of collecting all that real world data and then cleaning it for years, you create thousands of fake accounts, direct them to Claude, ask millions of questions, and then take the outputs from those answers, and you train your model on those outputs. You're basically teaching your model to think like Claude by feeding it everything Claude knows.
I mean, it's genius. It's brilliant. You're compressing years of research and billions of dollars of compute into a few weeks for basically pennies. And the people who built the original thing, they have no idea this is happening unless they're actively looking for it. Just last month, Enthropic published a report the details of how Chinese AI companies have been stealing from Claude and they call it industrial scale distillation against their