> These models are trained to be truthful
replyA more accurate statement would be that these models are trained to fit the training data as closely as possible, regardless of whether the training data reflects the truth.
Well, yes, but remember there is the reinforcement learning that is applied after, and the system prompts that will bend the results.
replyYeah, agree on both points. You can embed any bias you want using RL, regardless of training data.
replyNo, they're very clearly trained to be truthful.
reply