4 ms·
Training on data in a particular field will not prevent hallucinations, because the generation algorithm itself works on hallucinations. Anything below the res
by mcphage 2y ago
Training on data in a particular field will not prevent hallucinations, because the generation algorithm itself works on hallucinations. Anything below the resolution of “looks like the source” is indistinguishable.
- imtringued 2y agoWhen you ask it to do arithmetic operations, its approximation performance is surprisingly good. The expectation is that its responses are good approximations of the query, but not necessarily perfectly accurate. The problem though is that good approximations in the English text domain might care more about good presentation than accurate information content.