As LLMs become the go-to for quick answers, fewer people are posting questions on forums or social media. This shift could make online searches less fruitful in the future, with fewer discussions and solutions available publicly. Imagine troubleshooting a tech issue and finding nothing online because everyone else asked an LLM instead. You do the same, but the LLM only knows the manual, offering no further help. Stuck, you contact tech support, wait weeks for a reply, and the cycle continues—no new training data for LLMs or new pages for search engines to index. Could this lead to a future where both search results and LLMs are less effective?
LLMs are awesome in their knowledge until you start to hear its answers to stuff you already know and makes you wonder if anything was correct.
What they call hallucinations in other areas was called fabulations, to invent tales or stories.
I’m curious about what is the shortest acceptable answer for these things and if something close to “I don’t know” is even an option.
I get the feeling that LLMs are designed to please humans, so uncomfortable answers like “I don’t know” are out of the question.
Not designed, but trained. Training involves rewarding finding answers, so they WILL give you something. “I don’t know” is not going to fare well in the training development, so it naturally gets filtered out, while very creative (but wrong) LLMs do well.
This is mostly just a matter of proper prompting.
Sound similar to betteridges law of headlines.
Im sure there are tricks like adding ‘fact check your response’ but I suspect there is something intrinsic to these models that makes it a super difficult problem.
This applies equally well to human-generated answers to stuff.
But LLMs truly excel at making their answers look correct. And at convincing their users that they are.
Humans are generally notoriously bad at that kind of thing, especially when our answers are correct.
Have you met humans? Many of them base their entire career on this skill.
Sure, but they’re a minority. Millions, at most, out of billions. Probably less than that.
All modern LLMs are as good as professional mentalists at convincing most of their users that they know what they’re saying.
That’s what they’re designed, trained, and selected for. Engagement, not correctness.
True, the difference is that with humans it’s usually more public, it is easier for someone to call bullshit. With LLMs the bullshit is served with the intimacy of embarrassing porn so is less likely to see any warnings.