[altcognito]

Well, yes, but remember there is the reinforcement learning that is applied after, and the system prompts that will bend the results.

[lossolo]

Yeah, agree on both points. You can embed any bias you want using RL, regardless of training data.