> These models are trained to be truthful

A 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.

▲

altcognito 21 hours ago | parent | next [-]

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

	▲	lossolo 21 hours ago \| parent [-]
		Yeah, agree on both points. You can embed any bias you want using RL, regardless of training data.

▲

slibhb 21 hours ago | parent | prev [-]

No, they're very clearly trained to be truthful.