[abalashov]

Yeah, but this is like saying that one needn't focus so much on LLMs making mistakes because humans also mistakes.

They do, but the shape of the way LLMs will confidently mislead you is quite different to the way misinformed humans, or even the malevolent and mendacious humans, will mislead you.

[brookst]

There are plenty of authoritative reference books full of errors. Teachers in every subject can be wrong, convincingly.

[abalashov]

Yes, but human errors are based on genuine and elaborate misconceptions (or propagandistic intent), not squishy, facile "you're right to push back on that, I straight up made that up" type stuff.

[ChrisMarshallNY]

I just apply the "Wikipedia model" to things like LLMs (and StackOverflow, which can be just as bad -but without the admitting error part).

I look for citations and footnotes. In many cases, accuracy isn't something that I worry about, as bad info becomes apparent, almost immediately. In cases where it matters, though, I may try a couple of verifications.

One problem with humans, is that we can be quite insecure, and will fight to the death to defend a provably wrong position, because we can't bear to be seen as in error.

[jcgl]

> bad info becomes apparent, almost immediately

Can you elaborate on this? I suspect that you’re thinking mostly of cases in which you already have a fair bit of domain expertise. But in the general case, this seems to be very untrue. Which is why it’s so pernicious that LLMs can generate such quantities of syntactically-plausible-but-factually-untrue text.

[ChrisMarshallNY]

Valid point, but I don’t really just start learning whole new vocations, cold. I’m an engineer/software developer, with over 40 years’ experience, and my learning is generally some branch off that (like learning asynchronous programming, or a new UI framework).

Can’t really put it into precise terminology, but I get a “gut feeling,” that something is/is not plausible, and it happens pretty quickly; usually when I start some implementation. I tend to learn by do[0] (note the “sample playground,” in each essay), so the “reality filter” gets applied fairly rapidly.

In any case, I have been dealing with misinformation (sometimes, deliberate), for a long time, and I’m still working fairly effectively, so I guess it works.

Can’t argue with results.

[0] https://littlegreenviper.com/series/swiftwater/

[jcgl]

Sure, I take your point that the smell-test works reasonably well for domains related or adjacent to one's own.

But (not speaking about your use specifically here) many people (most, I'd wager) use LLMs for many things beyond their own expertise. And it's there that they're most likely to be ensnared without even knowing it.

I definitely agree with your notion that learning-by-doing is a helpful salve for LLM falsehoods. It's no panacea (working != correct (an incorrect solution can appear correct over a given interval)), but it's a good way of working in general that helps keep LLMs in check. And it's very natural to code or other things that can be immediately applied.

But learning-by-doing of course doesn't work with topics that aren't immediately applied. Which includes lots of topics that people use LLMs for (Wikipedia too, for that matter).

The set of unfamiliar-or-unapplied is practically a lot larger than the set of familiar-or-applied.

[abalashov]

Fair enough, but I'm not sure this is how the general population uses LLM chatbots, nor how the highly qualified always use LLMs.

On the contrary, I believe most people use them precisely to find out more about topics of which they know vanishingly little, much as they'd have used Google before LLMs.

[dpkirchner]

I wish humans that make up "facts" would admit to it as readily as LLMs do. Even though I know it's just a training trick to ground the response.

[varjag]

It's not anywhere close. Like, by four orders of magnitude.