When you trigger the Chinese model so hard it replies in Chinese
@matrix Have you seen where the diluted models claim to be themselves in chinese, but ask them in english and they tell what they were diluted from.
@matrix I wonder how much chinese data it has to be trained on in order to do that.
@Chizu Chinese resell AI subs so all the US AI companies definitely have ton of conversations in Chinese.
but I don't think training data full of chinese is needed. if you reward the llm for using less token during thinking using it can train itself to use chinese characters where it's efficienty.
@Chizu *models