Hacker News Reader

Comment by mstaoru | Hacker News Reader

Well it's not only this, or protection from LLMs training on LLM output. LLMs training on human output is also problematic.

I was traveling to an obscure small town, doing some "research" with LLMs beforehand. Every and each one told me enthusiastically to go to "Foobar square" (name changed) for the "best street food in XYZ town", some added a lot of colorful details.

There was no Foobar square in XYZ town. There was no Foobar square anywhere in the world. There was a SINGLE old Reddit comment, with no upvotes, to a unpopular post in an unpopular subreddit, where someone clearly badly misspelled the name of the square, and said something like "for street food go to Foobar square". Nothing about "the best" even.

It's all a lie.

View on HN

I think this was a common game on city/town subs. It happened here, there was a post asking for a good restaurant and someone just made up a name. It went viral and people started posting made-up menus for the place, reviews, and for a couple of months any time someone asked about a restaurant this fictional place would get mentioned.

It was all done as a joke to see if they could get Gemini or ChatGPT to start recommending it.

A friend did some vandalism on Wikipedia 20 years ago (!), and yet, LLMs quote his "original research".

Time to start some new restaurants matching those. Like Bubba Gump.

The Montréal subreddit has been doing this for ages before LLMs were a thing because every summer and fall there's endless threads from tourists and students asking the same questions that recommending a local gay bathhouse became the meme answer.

I've had gemini claiming code would compile and run while also outputting the same variable in the same sniplet with "fork" "frok" and "fokr" in the name. I'm not surprized it's trained on garbadge.

Your one example doesn't make all of LLMs a lie.

It's like someone reading a National Enquirer article about "Hillary Clinton being an alien from outer space" (a real headline topic from decades ago) and drawling the conclusion that all journalism is "a lie". The user has to understand media literacy and be at least a little skeptical of the claims that are made, then cross reference with another source.

It kind of does, mathematically. It doesn't label confidence. If 0.0001% of answers is a lie, without knowing which parts are a lie exactly, you cannot trust any of them.

If you need to independently verify every fact, why not just gather facts yourself in the first place.

Let's say, a mathematical concept of lie. I still use them every day, of course.

"why not just gather facts yourself" vs "i use them every day", the duality of man. But yes, I appreciate that LLM output is theoretically completely untrustworthy- but when in practice I observe that it's around 90% accurate, I have to rely on my own internal calibration for how useful it is (depends on type, nature of task ofc)

As they say, the less you know, the better LLM output is. :D

Okay, but the anecdote states that every model repeated the pseudo-factoid about Foobar square, not just the 4 GB open source model equivalent of a tabloid.

I think the key phrase here is, "an obscure small town." There may only be a single mention of this place, hence the only one on which a response can be based. This says more about the user's understanding of LLMs than it does about LLMs.

This says more about the user's understanding of LLMs than it does about LLMs.

"Tell me everything you know about (obscure small town), (state). Only what's unique to (town), not commonly-known facts" is an excellent way to test for hallucinatory tendencies in a new model, in my experience. Likely the best I've found.

Quality of results is almost linearly proportional to the size of the model in many cases. The largest models like K3 and GLM 5.3 will either confine their responses to known true facts about the town and its surroundings, or admit they don't have enough information to answer. Smaller ones will reliably make up hilarious or downright-strange things.

Another good test is https://whatever.scalzi.com/2025/12/13/ai-a-dedicated-fact-f... , which still works on the newest models. Of the open-weight models available, only Kimi K3 will consistently admit it has no idea who Scalzi's novel is dedicated to. The rest still make up random stuff and present it confidently.

TL,DR: progress is possible, and it has been made, but it's happening slower than many people think.

Reminds me of https://petergpt.github.io/bullshit-benchmark/viewer/index.v...

Without disclosing what you were prompting for, it's impossible to evaluate your claim.

Correct. We can either accept the claim or disregard it. The comment I replied to opted to accept it and then committed a fallacy, hence my response.

> Your one example doesn't make all of LLMs a lie.

It's not a lie though, because the truth isn't guaranteed by the mechanism that generates the answer.

by what mechanism do you believe LLMs verify truth? they are amoral token generators

LLMs have no sense of truth, they are just next token generators

Reply on news.ycombinator.com