[dentalperson]

How do they get ChatGPT not to hallucinate stuff about the articles? Everything seems fairly accurate, which is not my experience with ChatGPT when talking about technical things. Is it heavily curated/edited by humans? I noticed that the text often comes out verbatim from the articles, perhaps this indicates a clever prompt that keeps things closer to the truth by requiring verbatim output.

[gremlinsinc]

chatGPT hallucinates more the further removed it is from the data. I'm asking it about laravel, and it knows nothing about laravel 9 or 10 changes, but if I feed it an entire article or document it'll hallucinate a lot less because it's fresh.

kinda like how we can recall things closer to the event than months later.

it knows a ton from it's training but it still got it from the web so always question it, but if we can add meta data and other things to strengthen the llms understanding it shouldn't hallucinate much at all.

[ilaksh]

If you use temperature 0 with an API call it does not hallucinate much at all especially with a good prompt including the information you are asking about.