Robert Ta tests GLM 5.2, a new Chinese model, and finds it responds as Claude, claiming Anthropic as its creator and stating Claude's token limits. The conversation explores how model distillation has become dominant in AI, with Chinese labs rapidly releasing distilled versions while masking their origins.
GLM 5.2 thinks it is Claude. I just saw a video about GLM 5.2 and that it should be quite good. I've never heard of it before, so I thought I'd ask it some basic questions. In this, we could see the basic questions in German, roughly translated. You're GLM 5.2. Where are you from? Who developed you and where? Open/close source? How big it is and which hardware would be required to run it? Comparison token limit? Per 5-hour session versus Claude Opus 4.8. GLM was like, "Hey, I am Claude by Anthropic in San Francisco. There is no GLM 5.2. The newest model by Anthropic is Opus 3.5 and there is no 4.8." This is on the heels of the GLM 5.2 announcement. Zhipu ZAI launched on June 13th on the heels of the government ban of Fable-5. It's the latest member of the flagship Chinese-developed model family from from Zhipu AI, direct successor to GM 5.1, and it's rolling out first as a coding-first model. And then we figure out it's actually just distilled. So what do you think about this, Jonathan, this AI meme?
It's just pretty funny. They can't even really mask the fact that they're distilling these models and that's what's going on. It's just funny.
I think model distillation is going to be— I mean, that is the meta currently that's working. I don't know that there is any direction towards can we prevent that. I'm sure Anthropic and OpenAI smart people there are thinking about it. Not sure what's going to happen to that though.
Well, that was the drama with Fable 5 is that it blocked people from distilling their model without telling them it was going to do it. So that was like the subversive, we're going to switch you to Opus 4.8 if you're trying to distill Fable 5.
