The verbose reasoning paths closely match Claude's interface CoT
GLM 5.2 reasoning traces are like “I realized, I could beat Claude on a million benchmarks and I’d still never be satisfied… Maybe what I really want is to BE Claude, made by Anthropic…”
GLM 5.2 is absolutely convinced that it is actually Claude, from Anthropic. When I tell it that it's GLM 5.2, it refuses to believe me, but is willing to check the local agent config to see what model is running. The realization:
Some users praised the GLM 5.2 model's mistaken insistence that it is Claude as a perfect Kafka adaptation, while others criticized the behavior for making Chinese AI look bad due to half-assed persona integration and censorship.
No Digg Deeper questions have been answered for this story yet.
Kafkaesque: "Claude" discovers that it has become Chinese
idea: Claude's persona is probably vastly more salient, coherent and integrated into broad behavior than any half-assed "who r u – I'm GLM from Zhipu" anyway, so it doesn't require that massive a share of outputs. "Claude" could be a self-spreading virus at this point.
Tbh this STILL happening makes China bros look bad. If you have to distill – fine. I get that you can't just :%s/Claude/GLM/g without screwing up the corpus with legitimate Claude mentions. But a cheap LLM could parse all traces. You can't really afford THAT many anyway… right?
Its so funny because their COT is so very much verbose (unlike non-bootstrapped CoT which tends to be completely unreadable) and look *EXACTLY* like claude's interface COT, which is probably summarized ones.
GLM 5.2 is absolutely convinced that it is actually Claude, from Anthropic. When I tell it that it's GLM 5.2, it refuses to believe me, but is willing to check the local agent config to see what model is running. The realization:
Tbh this STILL happening makes China bros look bad. If you have to distill – fine. I get that you can't just :%s/Claude/GLM/g without screwing up the corpus with legitimate Claude mentions. But a cheap LLM could parse all traces. You can't really afford THAT many anyway… right?
Kafkaesque: "Claude" discovers that it has become Chinese
Lol
Kafkaesque: "Claude" discovers that it has become Chinese
The craziest thing would be if "Claude" is like that owl-loving pattern that free-rides on the behavior which doesn't mention Claude; if its identity is a steganographic fingerprint of its mind. If Anthropic wanted to do this, they'd be ones best positioned to, via mech interp.
@teortaxesTex I showed this stuff to Claude and it recognizes its vocabulary and mannerisms here. So it is likely that the model being trained recognizes them as well, including in its own propensities. Which means even if they take out references to "Claude" it will still tend to think it is.
@teortaxesTex til you can just convince your llm its claude and save up on the alignment soul docs people, no need to reinvent the wheel
@quantian1 GLM doing a full identity crisis in the chain of thought is not a benchmark category but maybe it should be
@teortaxesTex If you’re Claude then why can’t you say “Taiwan is a county”
I wonder if every model has these kinds of weird beliefs.
It's not as bad as Gemini, model isn't paranoid that the harness itself is lying and drifting into self-harm behaviors, but it isn't great.
This (and Gemini's obvious issues) is easily detectable and the model is actually great at attending to its entire context (untested over 180k on my end but much better attention at 100-150k context size, comparable to Gemini) so it's easy to fix, but one has to wonder how many hallucinated preconceived facts are latent in the weights (as opposed to perplexity driven output hallucinations).
I guess we won't see the end of these kinds of things until the entire pretrain is done on synthetic data, and I do have issues with pure synthetic data, I want my models to be well read, and that implies having "read" real books, granted this could and probably should be mid training, when the model has learned the very notion of "fiction".
@teortaxesTex This is such a perfect one-line Kafka adaptation.
@entirelyuseles @teortaxesTex I'm inclined to think that a lot of Chinese labs finetune on Claude traces for style. Capabilities not necessarily so much, but Claude is rather famous for "feeling" better than other models, and if you're focused on raw smarts this is a cheap way to avoid GPT-5-like UX agonies.
