calme_toi

Generative models nowadays are good. They don't make simple mistakes or hallucinations anymore. But that means the gotchas are more subtle, more hidden, and, maybe, more dangerous.

Especially in an area I am not an expert in, the answers I get from models seem fine, but to a domain expert, they can be total BS.

One way to mitigate this is to instruct the models to reply less generatively. For example, instead of summarizing, I ask the models to highlight or extract the sentences that contain key info from the articles.

So HN, am I the only one doing it? What are your non-generative tricks?