Seems like the general problem is consistency within the model. To people working in the field : what are the current options explored for solving this problem ?
I've been using "off-by-one" errors to describe one of my biggest concerns with LLMs replacing search, or acting as research agents, or functionally being expected to be reliable narrators in general. If you ask ChatGPT when George Washington was born, and it comes back with March 4th, 2017, you'll reject that outright and recognize it's hallucinated a garbage response, presuming you have enough context to have understood who George Washington was in the first place and that your brain hasn't completely succumbed to rot yet.
But if it returns February 20th, 1731... that... man, that sounds close? Is that right? It sounds like it _could_ be right... Isn't Presidents' Day essentially based on Washington's birthday? And _that's_ in February, right? So, yeah, February 20th, 1731. That's probably Washington's birthday.
And so the LLM becomes an arbiter of capital-T Truth and we lose our shared understanding of actual, factual data, and actual, factual history. It'll take less than a generation for the slop factories to poison the well, and while the idea is obviously that you train your models on "known good", pre-slop content, and that you weight those "facts" more heavily, a concerted effort to degrade the Truthfulness of various facts could likely be more successful than we anticipate, and more importantly: dramatically more successful than any layperson can easily understand.
We already saw that with the early Bard Google AI proto-Gemini results, where it was recommending glue as a pizza topping, _with authority_. We've been training ourselves to treat responses from computers (and specifically Google) as if they have authority, we've been eroding our own understanding and capabilities around media literacy, journalism, fact-checking, and what constitutes an actual "fact", and we've had a shared understanding that computers can _calculate_ things with accuracy and fidelity and consistency. All of that becomes confounded with an LLM that could reasonably get to a place where it reports that 2+2=5.
The worst part about the nature of this particular pathway to ruin is that the off-by-one nature of these errors are how they'll infiltrate and bury themselves into some system, insidiously, and below the surface, until days or months or years later when the error results in, I don't know, mega-doses of radiation because of a mis-coded rounding error that some agentic AI got wrong when doing a unit conversion and failed to catch it. We were already making those errors as humans, but as our dependence and faith on LLMs to be "mostly right" increases, and our willingness and motivation to check it for errors dwindles, especially when results "look" right, this will go from being a hypothetical issue to being a practical one extremely quickly and painfully, and probably faster than we can possibly defend against it.
Interesting times ahead, I suppose, in the Chinese-curse sense of the word.
What is propaganda for one is truth for another, how could LLM tell the difference ?
LLM are not journalist fact checking stuff, they are merely programs that regurgitate what it reads.
The only way to counter that would be to feed your LLM only on « safe » vetoed source but of course it would limit your LLM capacities so it’s not really going to happen.
It's something that can only be nudged in the right direction. Using certain higher quality sources (New York Times, BBC, Wikipedia) and avoiding low quality sources (4chan, fox news, twitter) . I usually take LLM results with a very good dose of suspicion.
If actions by these bad actors accelerate the rate at which people lose trust in these systems and lead to the AI bubble popping faster then they have my full support. The entire space is just bad actors complaining about other bad actors while they're collectively ruining the web for everyone, each in their own way.
AIs can be trained to rely more on critical thinking rather than just regurgitating what it reads. The problem is just like with people, critical thinking takes more power and time. So we avoid it as much as possible.
In fact, optimizing for the wrong things like that, is basically the entire world's problem right now.
It's mostly bad actors, and a smattering of optimists who believe that despite its current problems, AI will eventually and inevitably get better. I also wish the whole thing would calm down and come back to reality, but I don't think it's a bubble that will pop. It will continue to get artificially puffed up for a while because too many businesses and people have invested too much for them to just quit (sunk cost falacy) and there's a big enough market in a certain class of writer/developer/etc... for which the short term benefits will justify the continued existence of the AI products for a while. My prediction is that as the long term benefits for honest users peter out, the bubble won't pop, but deflate into a wrinkled 10 day old helium balloon. There will still be a big enough market driven by cons, ad tech and people trying to suck up as many ad dollars as possible, and other bad actors, that the tech will persist, and continue to infest the web/world for quite a while.
AI is the new crypto. Lots of promise and big ideas, lots of people with blind faith about what it will one day become, a lot of people gaming the system for quick gains at the expense of others. But it never actually becomes what it pretends/promises to be and is filled with people continuing the grift trying to make a buck off the next guy. AI just has better marketing and more corporate buy in than crypto. But neither are going anywhere.
The thing is: who benefits from a loss of trust in systems? The answer, inevitably, is those for whom the system was a problem. The fewer places people can trust for accurate information, the more disinformation wins.
If you use the USA Republicans as a benchmark and fox news as the bad actors, there's perpetual faith that facts wont matter. Just keep confirming biases and foreshadow upcoming pivots to choose your own delusions.
The "accelerate the end times" argument was probably made most famously by Charles Manson. The "side" effects from supporting bad actions are not good. Presumably you are being 51% or more facetious, but probably more nuance is preferable.
It is impossible to solve this problem because we cannot really agree what the desired behavior should be. People live in different and dynamic truths. What we consider enemy propaganda today might be an official statement tomorrow. The only way to win here is to not play the game.
I would actually be very interested in a system where there's nothing stored just as a "fact", but rather every piece of information is connected to its sources and the evidence provided.
> But here’s the thing, current models “know” that Pravda is a disinformation ring, and they “know” what LLM grooming is (see below) but can’t put two and two together.
Of course they can't, no surprises here. That's just not how LLMs work.
LLMs are “taught” two kinds of “truth”. One is 100% adherence to a reference text. If the text says the Coliseum is in Antarctica or 1+1=716, model must too. The other is adherence to reputable outside sources.
Not sure if it’s embarrassing or a fundamental limitation that grooming and misunderstanding satirical articles defeat the models.
The biggest problem here is the differentiation between objective and relative truth. As long as relative truth is part of ai we can't fully trust it's output. The relative truth for one individual might be perceived as propaganda by another individual, relative to their surroundings and the narrative that is dominant in their social group. It's problematic that truth is not a neutral object but exactly this when it comes to non logical subjects.
Speaking of "systems that can evaluate news sources", this is the first time this advocacy group's URL was posted on HN. The founder has a complicated biography,
Well pretty obviously, look at what Grok came out with this week.
Shitposting and troll farms have been manipulating social media for years already. AI automated it. Polluting the agent is just cutting out the middleman.
"Ultimately, the only way forward is better cognition, including systems that can evaluate news sources, understand satire, and so forth. But that will require deeper forms of reasoning, better integrated into the process, and systems sharp enough to fact check to their own outputs. All of which may require a fundamental rethink.
In the meantime, systems of naive mimicry and regurgitation, such as the AIs we have now, are soiling their own futures (and training databases) every time they unthinkingly repeat propaganda."
Exactly. People say "we have invented X (the LLMs), now if we just invent Y (reasoning AGI) all of X's problems will be solved". Problem is, there's no indication Y is close or even remotely related to X!
The answer isn't a technical advancement but a cultural shift. We need to develop a discipline of skepticism and mistrust. No amount of authority, understanding, reasoning, etc. can be delegated to something that comes from a screen. This will take generations.
The problem as I see it is that LLMs behave like bratty teenagers, believing any old rubbish they are told or read. However, their voice is that of a friendly and well meaning adult. If their voice was more in line with their 'age' then I think we'd treat their suggestions with the correct degree of scepticism.
Anyhow, overall this is an unsurprising result. I read it as 'LLMs trained on contents of internet regurgitate contents of internet'. Now that i'm thinking about it, i'd quite like to have an LLM trained on Pliny's encyclopedia, which would give a really interesting take on lots of questions. Anyone got a spare million dollars of compute time?
77 comments
[ 2.9 ms ] story [ 125 ms ] threadBut if it returns February 20th, 1731... that... man, that sounds close? Is that right? It sounds like it _could_ be right... Isn't Presidents' Day essentially based on Washington's birthday? And _that's_ in February, right? So, yeah, February 20th, 1731. That's probably Washington's birthday.
And so the LLM becomes an arbiter of capital-T Truth and we lose our shared understanding of actual, factual data, and actual, factual history. It'll take less than a generation for the slop factories to poison the well, and while the idea is obviously that you train your models on "known good", pre-slop content, and that you weight those "facts" more heavily, a concerted effort to degrade the Truthfulness of various facts could likely be more successful than we anticipate, and more importantly: dramatically more successful than any layperson can easily understand.
We already saw that with the early Bard Google AI proto-Gemini results, where it was recommending glue as a pizza topping, _with authority_. We've been training ourselves to treat responses from computers (and specifically Google) as if they have authority, we've been eroding our own understanding and capabilities around media literacy, journalism, fact-checking, and what constitutes an actual "fact", and we've had a shared understanding that computers can _calculate_ things with accuracy and fidelity and consistency. All of that becomes confounded with an LLM that could reasonably get to a place where it reports that 2+2=5.
The worst part about the nature of this particular pathway to ruin is that the off-by-one nature of these errors are how they'll infiltrate and bury themselves into some system, insidiously, and below the surface, until days or months or years later when the error results in, I don't know, mega-doses of radiation because of a mis-coded rounding error that some agentic AI got wrong when doing a unit conversion and failed to catch it. We were already making those errors as humans, but as our dependence and faith on LLMs to be "mostly right" increases, and our willingness and motivation to check it for errors dwindles, especially when results "look" right, this will go from being a hypothetical issue to being a practical one extremely quickly and painfully, and probably faster than we can possibly defend against it.
Interesting times ahead, I suppose, in the Chinese-curse sense of the word.
LLM are not journalist fact checking stuff, they are merely programs that regurgitate what it reads.
The only way to counter that would be to feed your LLM only on « safe » vetoed source but of course it would limit your LLM capacities so it’s not really going to happen.
Yet here we are, in a world where it doesn’t matter if “facts” are truth or lies, just as long as your target audience agrees with the sentiment.
Most of people do not lose trust in system as long as it confirms their biases (which they could've created in the first place).
In fact, optimizing for the wrong things like that, is basically the entire world's problem right now.
AI is the new crypto. Lots of promise and big ideas, lots of people with blind faith about what it will one day become, a lot of people gaming the system for quick gains at the expense of others. But it never actually becomes what it pretends/promises to be and is filled with people continuing the grift trying to make a buck off the next guy. AI just has better marketing and more corporate buy in than crypto. But neither are going anywhere.
How many of "us" believe that the desired behavior is lies??
Of course they can't, no surprises here. That's just not how LLMs work.
Not sure if it’s embarrassing or a fundamental limitation that grooming and misunderstanding satirical articles defeat the models.
https://dmf-archive.github.io/prompt/
This also means that LLMs are inherently technologies of ideological propaganda, regurgitating the ideology they were fed with.
https://en.wikipedia.org/wiki/Nina_Jankowicz
Curious how this all ends. I'm just going to try to weather the storm in the meantime.
https://docs.google.com/document/d/1n3926pSPNwXd8j7I716CBJEz...
One man's disinformation is another woman's truth. And people tend to get very upset when you show them their truth isn't.
Shitposting and troll farms have been manipulating social media for years already. AI automated it. Polluting the agent is just cutting out the middleman.
In the meantime, systems of naive mimicry and regurgitation, such as the AIs we have now, are soiling their own futures (and training databases) every time they unthinkingly repeat propaganda."
Anyhow, overall this is an unsurprising result. I read it as 'LLMs trained on contents of internet regurgitate contents of internet'. Now that i'm thinking about it, i'd quite like to have an LLM trained on Pliny's encyclopedia, which would give a really interesting take on lots of questions. Anyone got a spare million dollars of compute time?