What scares me is not that ChatGPT is often wrong.

@mos_8502 What's worse is that there are some uses and topics for which it's almost 100% of the time, and some for which it's ... not very often. And, of course, there's no way to tell which is which unless you already know, or you fact check everything. If you use it in one context and then try it in another...

@mos_8502 The only reason it is ever right is due to the preponderance of right answers in the stolen data. What is truly amazing is how much of the vast swathe of crap that is the internet is correct, and that is simply reflected in the output of the synthetic text generator.

@criffer @mos_8502 Part of the disgusting way commercial LLMs are trained is that they're not just hoovering up the Internet: there are armies of invisible underpaid and exploited people in the majority world that filter, categorize, label, and otherwise massage the training data. This has been happening in some capacity for a very long time for commercial ML-based services in general, e.g. image classification services, and not just the recent explosion LLMs.

When you realize the horrific stuff these people have to deal with on a daily basis in filtering the numerous awful bits of the Internet and the fact that these poor workers get zero mental health support and take all that trauma back to their homes and communities, it's hard not to be absolutely disgusted by the entire industry.

@mos_8502 true!!

@mos_8502 sounds like me. and other actual people who know their thing. and make mistakes sometimes.

the scary part is trusting without any review