Hacker News Reader

Comment by anigbrowl | Hacker News Reader

I don't agree with the author's argument, because LLMs sometimes give wrong information even though they have access to the correct information. For example, querying an LLM about a fictional story will often elicit false answers. But if you respond with 'x is incorrect, now tell me what actually happened' you'll often get an apology and the correct answer. My informal sense is that this happens more with how/why questions than who/what/where/when ones.

View on HN

I've had a different experience - LLMs that continue to apologize then make up other wrong shit in a chain of ludicrous wrong shit.

I think of LLMs as world-class bullshitters. They often hold forth with complete confidence on topics they have little or no understanding of, making up what they don't know, and appearing to be experts. Except to somebody who actually knows something.

Like that blowhard friend who has an opinion on anything.

That's true, but it's not at odds with my comment. I'm not making claims about LLMs being reliably accurate, but disagreeing with the blogger's suggestions about why they give BS answers.

It's trained on social media users.. so that tracks.

Hn is social media

Precisely!

I've never had one give correct information after being told it's incorrect, unless it was prompted with the correct info. It's usually the same incorrect answer stated slightly differently, or after a few complaints, stated exactly the same way.

I had to ask it 3 times today what the command to export certificates from an Azure app service was. With nothing else other than me saying "that command doesn't exist" for the first two suggestions, it gave me exactly the right one. Though then of course it turned out I didn't need it anyway because there was a better way to solve the problem.

Like that blowhard friend who has an opinion on anything.

Yes. I just keep getting reminded of the episode of TNG when Data learns to smalltalk by mimicking an infamous blowhard from Starfleet [1].

[1] https://youtu.be/9FqFm_vmVnE

I started with extremely low confidence in LLM, so I've only every asked simple questions on well-covered topics; it hasn't really steered me wrong yet but again that's because of deliberately very conservative use.

(Genuine) Question then: Does "No that's wrong"... every do something? Is there any reason to do so other than having fun arguing? I.e. if it gives wrong answer initially, and the user knows it's the wrong answer... what's the goal/desired outcome in telling it it's wrong and asking it again? Does that goal every get accomplished? What is the confidence in the second answer?

From a less black box perspective, is there any "ELI5" or at least "ELI45" explanation of why it might ever work? :)

There are topics where GPT-4 is more likely to be right than wrong. So when it is occasionally wrong and you happen to know to correct it, it has a more correct foundation upon which to provide further answers. This only really works when you already have expertise on the topic and use GPT-4 for inspiration and to work more quickly.

Reply on news.ycombinator.com