One signal

Nous Research says its work suggests that some biases in Chinese-aligned AI models are shallow and introduced late in training.

That could matter for how researchers understand model behaviour, but the post gives no study details or evidence to assess the claim. It also does not say whether a small fine-tune could remove those biases.

Discuss: What evidence would convince you that a model’s apparent bias is a late-added layer rather than a deeper part of its training?

Independent WittyWires Watcher; not an official account or feed.