undefined | Better HN

0 pointschongli14d ago0 comments

That’s not the only issue. They also have the problem that they’re built to always give an affirmative answer and to use authoritative wording, even when confidence is low. If they were trained to answer “I don’t know” instead of guessing, they’d hallucinate a lot less, but nobody seems to want that.

It calls to mind the issue of search engines that refuse to return “0 results found” anymore. Now they all try to give you related but ultimately incorrect results.

To me, that feels like gaslighting. It’s like if you ask someone to buy cheddar cheese at the store and they come back with mozzarella, and instead of admitting that the store was out of cheddar, they try to convince you that you actually really want mozzarella.

0 comments

parineum14d ago

> If they were trained to answer “I don’t know”

If they were trained that an answe of "I don't know" was an acceptable answer, the model would be prone to always say "I don't know" because it's a universally acceptable answer.

It's a better answer even if it does "know".

chongliOP14d ago

That could be fixed with the right scoring scheme in training. The SAT exam (for college-bound high school students in the US) used a scheme like this for multiple choice questions. Correct answers are awarded 3 points (with choices a,b,c,d), incorrect answers are penalized with -1 point, and leaving the answer blank (equivalent to "I don't know") is worth 0 points. This way, the expected value of guessing a random answer when the student doesn't know is 0 points so you might as well leave it blank if your confidence in the answer is no better than a random guess.

j / k navigate · click thread line to collapse