alt Hacker NewsI find the tokenizers most compelling. That's what the model is trained on, it's an immutable fact of the model and its architecture. You know for a fact that the model is at least related to other models that way. And if a tokenizer is unique / specific to one lab, like GLM's is, it's basically as good as it gets.
Comparatively, you can't be 100% sure that Z.ai isn't able to host some other lab's model (although in this case, the hosting errors still support the GLM theory).
No, that's not how anything works.
You can finetune an LLM to a new tokenizer by nudging just a few layers (even wildly different kinds of tokens), and there's nothing stopping a lab from using someone else's tokenizer.