News for hackers. Stuff that matters.

A comment on GPT-4 and professional benchmarks: the wrong answer to the wrong question. Parent comment

Re:GPT-4 and professional benchmarks: the wrong answer to th... by raattgift on Tuesday March 21, 2023 @10:21AM (#35246542)
Probably the most fundamental problem of ChatGPT is that it refuses (or is even unable) to admit, "I don't know".
[ Parent | on HN ]
Re:GPT-4 and professional benchmarks: the wrong answer to th... by blihp on Tuesday March 21, 2023 @10:39AM (#35246850)
It truly has achieved human-level intelligence!
[ Parent | on HN ]
Re:GPT-4 and professional benchmarks: the wrong answer to th... by spacebanana7 on Tuesday March 21, 2023 @10:33AM (#35246753)
I imagine kind of defect can be fixed with more fine tuning / RHLF
[ Parent | on HN ]
Re:GPT-4 and professional benchmarks: the wrong answer to th... by sebzim4500 on Tuesday March 21, 2023 @11:02AM (#35247206)
Almost certainly true but they'd have to use a new cost function. It's not just about collecting examples where the model should say "I don't know".
[ Parent | on HN ]