it was obvious from day 1 this is what happened

the chinese labs haven't been improving at training from scratch, they've been getting better and better at distilling models, so naturally they continue to follow this path

firstly, it shows the vulnerability of exposing a model to users. a highly capable group can quite literally suck the functionality out of your model and take it for themselves, so there's no sandboxing it

secondly, it's actually interesting to know that distillation is so powerful. in the science-fiction scenario of meeting some other form of intelligent life, they might have their own models and this would be a way of siphoning intelligence off of them