Open weights don't tell you what's inside

A growing argument on my feed: switch to Chinese AI models, because they ship open weights and open beats closed.

Open weights are a real property. You can download the file, inspect it, run it yourself. What they don’t guarantee is what’s actually baked into the answers. Several open-weight models trained in China deflect or refuse on specific political topics, not as an API filter you could route around, but trained into the weights themselves. American models carry their own version of the same thing: different topics, same mechanism, alignment choices baked in by whoever trained them.

That doesn’t make the two sides interchangeable. Western values sit closer to mine, and probably to most people reading this. China’s political system adds another layer on top: it’s simply a more sensitive environment for a model to be shaped in. None of that makes American alignment some neutral default, just a reason not to treat the comparison as perfectly even.

I run closed and open models, local and via API, American and Chinese, depending on the task. That mix doesn’t excuse me from checking what I’m actually dealing with. Know the model, and stay critical of all of it.