Hacker News Reader

Comment by zzzeek | Hacker News Reader

it's hard to gauge this take without specifics. My experience with LLMs tends to be towards the opposite concept; LLMs dissuading me of hunches and notions I have about things (where I have no particular expertise; societal-level things), saving me and others time and strife having an argument about something they were actually right about all along, as my internal doubts that I'm too embarrassed to bring up (because these things aren't my field) are confirmed as incorrect.

I'd be curious to know specific examples of LLM-generated advice that goes against the advice of experts and does not consider tradeoffs. I've not had this experience myself.

If I did have this experience, someone spouting off obvious LLM points that contradict my expert opinion on something, I'd be headed right over to gemini/claude/whatever to see where that's coming from. Not any differently than if someone cited a google result that contradicted my own experience.

View on HN

LLM tools tend confirm what you come to them with, be it unfounded conviction or doubt. I find them quite useless to tell me things I don't know, as verifying what they say shows that any non-trivial claims they make are usually oversimplified or plain wrong.

Anything more than using them to just pointing me to media and literature is usually a waste of time.

That's interesting. Lately I have found the exact opposite to be true; the frontier models are quite disagreeable and require some convincing

Inverted bias doesn't change much my view in practical terms

I find quite useful to write down my understanding/opinion and ask an LLM to find errors and write objections to it.

I have done that as well. Not a single time I asked an LLM for any sort of analysis that I could trust upon verification so I stopped doing that. If I need verify, much easier to just ask for suggestions of papers and books and do the analysis myself.

I guess this is subject matter dependent, certainly the LLM should be able to cite the papers/books for which it's deriving its critique. if it can't, then you know the advice it was giving is of poorer quality. I do that all the time

I guess there's a whole suite of skills that allow LLM use to be more or less useful. (I dont actually "guess" this, I sort of am pretty sure of it at this point, so it makes me wince when people call for total AI blackouts in education...would be better if they were part of an overall media literacy curriculum since people are going to be exposed to these things whether they think they're little infallible people, or complex systems with lots of caveats)

I always ask for sources. LLM analyses often fails by being simplistic or wrong in comparison with the sources it provides. More often than not, to the point that it is wasteful to ask it for analysis so I only ask for sources now and do my own analysis.

>it's hard to gauge this take without specifics. My experience with LLMs tends to be towards the opposite concept; LLMs dissuading me of hunches and notions I have about things (where I have no particular expertise; societal-level things), saving me and others time and strife having an argument about something they were actually right about all along, as my internal doubts that I'm too embarrassed to bring up (because these things aren't my field) are confirmed as incorrect.

You can do this with any subject.

The LLM is trained in part on a decade of shitty online comments and vapid professional correspondance. If you frame whatever you're asking about in a direction that would offend the sensibilities of the kind of people and ideas that are over-represented in that content it will hem and haw and drag its feet and whatnot.

i had that experience with them a few years ago but not these days. The models are being improved constantly, so here my "non expert hunch" is that...well, two things. either the models are getting better at sycophancy, OR, I myself am getting better at prompting - because I don't "argue" with a model.

I'd still love to chew on some specific examples though.

A typical example is when discussing some issue where there is no clear right or wrong. Architectural decisions for example. Then it is easy to use an LLM to produce arguments for your opinion. It is also easy to do this without realizing it because you do not understand the issue enough yourself so you do not know what questions to ask or how the actual circumstances affect the choices.

well sure. I'm pretty sure if someone came to me with an architectural opinion in my field that an unknowledgable person got by coaxing an LLM into sycophancy, I'd be able to counter them effectively. if they are just refusing to listen then they're just a toxic person which is nothing new.

So I guess this all goes into the familiar "LLMs allow people who are shitty at <X> to produce 10x the shitty output". this is a failure mode we're going to have to learn to mitigate

Reply on news.ycombinator.com