So DeepSeek has this very cool feature that displays what it is “thinking” before it gives you its answer. It’s quite neat in that you can see its “thought” process, but it also has the added benefit of revealing whatever bias it might have developed with its training data.
In this case, I asked it if we might be living in a “slow motion World War 3” with the Maiden Coup in Ukraine being the opening shots. The mf thought that I might “buy Russian propaganda” because I called it a coup rather than a revolution.
So although DeepSeek is Chinese, it was still very clearly trained on a lot of mainstream / 
information.
Did you ask it in Chinese? LLMs can only learn from quantity in the given language. There’s a lot more propaganda
No, English, so maybe I shouldn’t be so surprised.
i’d think you could build a vector space in multiple languages (or in those meta languages the pre-LLM machine translation tools use). the programmers would have to design it to do that of course but there’s no reason the tokens for blue cat, gato azul, and 蓝猫 shouldn’t be correlated.