[rcruzeiro]

I assume the web version has a wrapper around it that filters out what it considers harmful content (kind of what OpenAI has around ChatGPT, but much more aggressive and, of course, tailored to topics that are considered harmful in China). Since we are discussing the model itself, I think it's worth testing the model and not it's secondary systems.

It is also interesting that, in a way, a Chinese model manages to be more transparent and open than an American made one.

[nipah]

I think the conclusion is a stretch, tho, you can only know they are as transparent as you can know an american made one is, as far as I know the biases can be way worse, or they can be the exact same as of american models (as they supposedly used those models to produce synthetic training data as well). OpenAI models also have this kind of "soft" censorship where it is on the interface layer rather than the model itself (like with the blocked names and stuff like that).