Why has ChatGPT a bias towards correctness?

Question

I am using ChatGPT (3.5) as an assistant for programming and using tools like Terraform or jq.It's much more useful than Googling and gives helpful answers most of the time, although sometimes it takes refining the answers through repeated questions. There are hallucinations but I think it's more often right than wrong and very often it can answer questions on an expert level.When you search the web on your own you encounter many incorrent, misleading, outdated info on a lot of technical topics, though.So how is it that ChatGPT overall seems to generate more correct and helpful answers than misleading info?

sk11001 · Accepted Answer

Look into the training process - https://www.youtube.com/watch?v=zjkBMFhNj_g
The goal of the initial pretraining phase is to make it good at predicting the next word. The rest of the training process is aimed at making it (1) helpful and (2) as correct as possible.
I think some people oversimplify things by calling LLMs "next token predictors" and they leave out the tuning towards helpfulness and correctness.

thorin · Answer

I very rarely find it answering on "expert" level, but maybe I have a warped idea of what expert means, I've certainly never considered myself and expert at anything. However it definitely works as a good alternative to google search and often I find using it as a sounding board for stupid questions or as a kickstart when starting something new works very well.

gryfft · Answer

What kind of answer are you looking for here, specifically? The questions you're asking seem to be about how OpenAI's closed implementation works in terms of the biggest open research question in the field. What kind of answer do you expect?

futureshock · Answer

It&rsquo;s because when you take all information together, on average it&rsquo;s correct for those topics. If the consensus was incorrect, you would get incorrect answers.