[xico]

That’s indeed its interface, but there is still a huge training phase during which the model was taught using mostly English, and it explains why some other languages interactions feel like they are some direct translation of English expressions (not to mention higher ways of thinking). Most of the thinking indeed is done during training, and then encoded in the weights which act like a huge cache of thoughts, but there is still a lot of learning on how to use the context when you emit this character (token), the attention layers, which is also a form of thinking that uses whatever new input you are giving to the model and applies a series of transformations (granted, finite in current models) to it.

[Kuinox]

But maybe the dataset for said language is just too small to produce correct content at all, and that's why ChatGPT is bad, not because "ChatGPT is thinking in English". Except when it default to answering in English, ChatGPT sounds like french when I talk to it in french.

The bigger the dataset, the best is it's capabilities, training it exclusively with a language don't mean it will be better at this language.