Ask HN: Why LLM Is Lying? | Hacker News Reader

Ask HN: Why LLM Is Lying?

11 pointsdanver018 comments

Why LLM good at giving me answers that looked perfectly reasonable none of them work ?????

Because that's what LLMs do. They don't give "answers", they don't know what "works" and what doesn't. They create text based on a heuristic you provide and a token matching algorithm. That's it. That's the trick.

It isn't returning code, because it doesn't know what "code" is. It's returning language, essentially "code-shaped text." It only happens to work as well as it does when it does because the model is trained on examples of existing working code supplied by humans, therefore whatever it returns is likely to at least mostly be correct, at least for common cases where a high probability match is likely.

so "artificial intelligence" is a myth ?

If by "artificial intelligence" you mean that an LLM can simulate some aspect of what is considered intelligent behavior, no. Clearly LLMs can do that. But then so can a Markov chain.

If by "artificial intelligence" you mean something like the computer in Star Trek - essentially a sentient and self-aware being - then yes, that is a myth. That isn't what LLMs are. Although plenty of people believe otherwise, for whatever reason.

The problem is because LLMs can use and respond to natural language, we humans are hardwired to see them as the latter and anthropomorphize them. We imagine that if we give them a problem there's basically a little man inside the machine smart enough to understand that problem, searching through its data and trying to solve that problem the way a human would. But no, the only thing they're doing is constructing semantically correct output matching an input.

And it's wild that it works as well as it does, but most of that appearance of intelligence comes down to training on human effort and the result of human assumptions and bias.

It really goes to show how much humans associate intelligence with language — that if it sounds intelligent, then it must be intelligent.

It doesn't help that they've been trained to refer to themselves in ways that imply a sense of self-awareness ("as a large language model, I") or that they employ emotional language, which is far more effective at influencing people than rationality.

If people on HN can believe LLMs are sentient, sapient and intelligent beings (even more so than other humans I suspect) then there isn't much chance for average people getting caught in the intersection of LLM marketing, a hundred years of pop sci-fi cultural conditioning and a million years of primate evolution.

LLMs don’t know what Truth is. That’s up to the human. They are fancy search results. Right, wrong, true false, LLMs do not know this.

Because a bunch of AI boosters and snake-oil salesmen born every second pivoting to AI keep telling you it was “AGI” when in reality, these LLMs have no basis of reality or the truth as it confidently makes things up.

Every day, the AI boosters have a slot machine to sell you and you fell for it.

i hate every AI tech influencers for this reason

What gets me is that they get certain things wrong about Arknights, for instance none of them get it straight how Scavenger recovers DP. And I really want help with Arknights because the guides suck, though I guess LLMs are bad at Arknights questions because... the guides suck.

The thing is that LLMs are not moral subjects, they don't feel bad the way you feel or the way a dog or a horse feels when they let somebody down. I worked for a company developing prototypical foundation models circa 2018 and one of the reasons that I didn't invent ChatGPT is that I wouldn't have given a system credit for making lucky guesses.

I can't imagine that Arknights trivia is a part of many English-language training datasets.

Many of these things (Copilot and Google's AI Mode) go out and search.

I have no idea if I'd do better with Chinese sources, if there was a big FAQ or Wiki in Chinese I probably could load it into IntelliJ IDEA and ask Junie questions about it... Maybe I should! I guess it would be six months ahead in terms of events and might recommend using operators I can't get but I could live with that... And for that matter I don't like the quality of translations I have available for things like Investiture of the Gods

Do you need translations for Investiture of the Gods? Is it not enough to behold https://en.wikipedia.org/wiki/Daji#Daji_as_a_mythical_fox ?

It repeats mixes of what people said and did in it's training data. What goes in is what its model says has the highest probability of filling in the blanks. There's even many lies and inaccurate data in the training data of most models.

"Garbage In, Garbage Out"

"You get out of it what you put into it."

i have a fear that some Evil ppl will fake the history by training LLM with wrong data, so future generation will struggle knowing the truth

a large amount of truth we know we learned because instead of consulting other oracles, we asked the universe about it; why would that path not remain open to your future generations?

They do lie. Ingore the "technical" people explaining how they just 'get stuff wrong' based on a lack of information or bad training/data.

You can prove this by asking any of the super-smart LLMs that have access to 'current' data (searching, whatever), which US President has bombed the most countries in their two terms. They will claim it was Obama, or that they cannot determine the answer because it's "complicated". The truth is, the USG and it's technocrats instruct and train these bots to lie in support of the state agenda.

These bots will even claim the true answer is misinformation even after you force them to the factually correct answer. Just like the Wizard of Oz, it's just sad little man pulling the strings of a terrifying facade.

>They do lie.

'Lying', 'hallucinating' and other efforts to anthropomorphize a computer program only serves to reinforce the snake-oil being sold worldwide.

An LLM cannot lie to you. Lying would imply that it somehow knows the truth, and chooses to tell you something other than the truth. The LLM doesn't know anything. It's just providing you with answer-shaped responses.

Reply on news.ycombinator.com