Comment by HarHarVeryFunny
2 days ago
The factuality problem with LLMs isn't because they are non-deterministic or statistically based, but simply because they operate at the level of words, not facts. They are language models.
You can't blame an LLM for getting the facts wrong, or hallucinating, when by design they don't even attempt to store facts in the first place. All they store are language statistics, boiling down to "with preceding context X, most statistically likely next words are A, B or C". The LLM wasn't designed to know or care that outputting "B" would represent a lie or hallucination, just that it's a statistically plausible potential next word.
I think this is why I get much more utility out of LLMs with writing code. Code can fail if the syntax is wrong; small perturbations in the text (e.g. add a newline instead of a semicolon) can lead to significant increases in the cost function.
Of course, once an LLM is asked to create a bespoke software project for some complex system, this predictability goes away, the trajectory of the tokens succumbs to the intrinsic chaos of code over multi-block length scales, and the result feels more arbitrary and unsatisfying.
I also think this is why the biggest evangelists for LLMs are programmers, while creative writers and journalists are much more dismissive. With human language, the length scale over which tokens can be predicted is much shorter. Even the "laws" of grammar can be twisted or ignored entirely. A writer picks a metaphor because of their individual reading/life experience, not because its the most probable or popular metaphor. This is why LLM writing is so tedious, anodyne, sycophantic, and boring. It sounds like marketing copy because the attention model and RL-HF encourage it.
>but simply because they operate at the level of words, not facts. They are language models.
Facts can be encoded as words. That's something we also do a lot for facts we learn, gather, and convey to other people. 99% of university is learning facts and theories and concept from reading and listening to words.
Also, even when directly observing the same fact, it can be interpreted by different people in different ways, whether this happens as raw "thought" or at the conscious verbal level. And that's before we even add value judgements to it.
>All they store are language statistics, boiling down to "with preceding context X, most statistically likely next words are A, B or C".
And how do we know we don't do something very similar with our facts - make a map of facts and concepts and weights between them for retrieving them and associating them? Even encoding in a similar way what we think of as our "analytic understanding".
Animal/human brains and LLMs have fundamentally different goals (or loss functions, if you prefer), even though both are based around prediction.
LLMs are trained to auto-regressively predict text continuations. They are not concerned with the external world and any objective experimentally verifiable facts - they are just self-predicting "this is what I'm going to say next", having learnt that from the training data (i.e. "what would the training data say next").
Humans/animals are embodied, living in the real world, whose design has been honed by a "loss function" favoring survival. Animals are "designed" to learn facts about the real world, and react to those facts in a way that helps them survive.
What humans/animals are predicting is not some auto-regressive "what will I do next", but rather what will HAPPEN next, based largely on outward-looking sensory inputs, but also internal inputs.
Animals are predicting something EXTERNAL (facts) vs LLMs predicting something INTERNAL (what will I say next).
>Humans/animals are embodied, living in the real world, whose design has been honed by a "loss function" favoring survival. Animals are "designed" to learn facts about the real world, and react to those facts in a way that helps them survive.
Yes - but LLMs also get this "embodied knowledge" passed down from human-generated training data. We are their sensory inputs in a way (which includes their training images, audio, and video too).
They do learn in a batch manner, and we learn many things not from books but from a more interactive direct being in the world. But after we distill our direct experiences and throughts derived from them as text, we pass them down to the LLMs.
Hey, there's even some kind of "loss function" in the LLM case - from the thumbs up/down feedback we are asked to give to their answers in Chat UIs, to $5/hour "mechanical turks" in Africa or something tasked with scoring their output, to rounds of optimization and pruning during training.
>Animals are predicting something EXTERNAL (facts) vs LLMs predicting something INTERNAL (what will I say next).
I don't think that matters much, in both cases it's information in, information out.
Human animals predict "what they will say/do next" all the time, just like they also predict what they will encounter next ("my house is round that corner", "that car is going to make a turn").
Our prompt to an LLM serves the same role as sensory input from the external world plays to our predictions.
5 replies →
> You can't blame an LLM for getting the facts wrong, or hallucinating, when by design they don't even attempt to store facts in the first place
On one level I agree, but I do feel it’s also right to blame the LLM/company for that when the goal is to replace my search engine of choice (my major tool for finding facts and answering general questions), which is a huge pillar of how they’re sold to/used by the public.
True, although that's a tough call for a company like Google.
Even before LLMs people were asking Google search questions rather than looking for keyword matches, and now coupled with ChatGPT it's not surprising that people are asking the computer to answer questions and seeing this as a replacement for search. I've got to wonder how the typical non-techie user internalizes the difference between asking questions of Google (non-AI mode) and asking ChatGPT?
Clearly people asking ChatGPT instead of Google could rapidly eat Google's lunch, so we're now getting "AI overview" alongside search results as an attempt to mitigate this.
I think the more fundamental problem is not just the blurring of search vs "AI", but these companies pushing "AI" (LLMs) as some kind of super-human intelligence (leading to uses assuming it's logical and infallible), rather than more honestly presenting it as what it is.
> Even before LLMs people were asking Google search questions rather than looking for keyword matches
Google gets some of the blame for this by way of how useless Google search became for doing keyword searches over the years. Keyword searches have been terrible for many years, even if you use all the old tricks like quotations and specific operators.
Even if the reason for this is because non-tech people were already trying to use Google in the way that it thinks it optimized for, I'd argue they could have done a better job keeping things working well with keyword searches by training the user with better UI/UX.
(Though at the end of the day, I subscribe to the theory that Google let search get bad for everyone on purpose because once you have monopoly status you show more ads by having a not-great but better-than-nothing search engine than a great one).
Yeah I pretty much agree with everything you’ve got here
In a way though those things aren't so different as they might first appear. The factual answer is traditionally the most plausible response to many questions. They don't operate on any level other than pure language but there are a heap of behaviours which emerge from that.
Most plausible world model is not something stored raw in utterances. What we interpret from sentences is vastly different from what is extractable from mere sentences on their own.
Facts, unlike fabulations, require crossing experience beyond the expressions on trial.
Right, facts need to be grounded and obtained from reliable sources such as personal experience, or a textbook. Just because statistically most people on Reddit or 4Chan said the moon is made of cheese doesn't make it so.
But again, LLMs don't even deal in facts, nor store any memories of where training samples came from, and of course have zero personal experience. It's just "he said, she said" put into a training sample blender and served one word at a time.
> The factual answer is traditionally the most plausible response to many questions
Except in cases where the training data is more wrong than correct (e.g. niche expertise where the vox pop is wrong).
However, an LLM no more deals in Q&A than in facts. It only typically replies to a question with an answer because that itself is statistically most likely, and the words of the answer are just selected one at a time in normal LLM fashion. It's not regurgitating an entire, hopefully correct, answer from someplace, so just because it was exposed to the "correct" answer in the training data, maybe multiple times, doesn't mean that's what it's going to generate.
In the case of hallucination, it's not a matter of being wrong, just the expected behavior of something built to follow patterns rather than deal in and recall facts.
For example, last night I was trying to find an old auction catalog from a particular company and year, so thought I'd try to see if Gemini 3 Pro "Thinking" maybe had the google-fu to find it available online. After the typical confident sounding "Analysing, Researching, Clarifying .." "thinking", it then confidently tells me it has found it, and to go to website X, section Y, and search for the company and year.
Not surprisingly it was not there, even though other catalogs were. It had evidently been trained on data including such requests, maybe did some RAG and got more similar results, then just output the common pattern it had found, and "lied" about having actually found it since that is what humans in the training/inference data said when they had been successful (searching for different catalogs).
>Except in cases where the training data is more wrong than correct (e.g. niche expertise where the vox pop is wrong)
Same for human knowledge though. Learn from society/school/etc that X is Y, and you repeat X is Y, even if it's not.
>However, an LLM no more deals in Q&A than in facts. It only typically replies to a question with an answer because that itself is statistically most likely, and the words of the answer are just selected one at a time in normal LLM fashion.
And how is that different than how we build up an answer? Do we have a "correct facts" repository with fixed answers to every possibly question, or we just assemble our training data from a weighted graph (or holographic) store of factoids and memories, and our answers are also non deterministic?
3 replies →
If you want to see how well text generation works on unrehearsed questions, you can ask about what happens in a comic strip.
I found commentary about searching Google for "dark legacy comics who wants some bamboo", and posted results for that search on HN in response to a comment saying "I work at Google on the 'AI Mode' search option, check it out!" ( > Dark Legacy Comic #500 is titled "The Game," a single-panel comic released on June 18, 2015. It features the main characters sitting around a table playing a physical board game, with Keydar remarking that the in-game action has gotten "so realistic lately."
> You can view the comic and its commentary on the official Dark Legacy Comics website. [link]
Compare https://darklegacycomics.com/500 .
That [link] following "the official Dark Legacy Comics website" goes totoddmorey
2 days ago
therealpygon
2 days ago
emptyfile
2 days ago
wisty
2 days ago
HarHarVeryFunny
2 days ago
wisty
2 days ago
wisty
2 days ago
TheOtherHobbes
2 days ago
Yeah, that’s very well put. They don’t store black-and-white they store billions of grays. This is why tool use for research and grounding has been so transformative.
Definitely, and hence the reason that structuring requests/responses and providing examples for smaller atomic units of work seem to have quite a significant effect on the accuracy of the output (not factuality, but more accurate to the patterns that were emphasized in the preceding prompt).
I just wish we could more efficiently ”prime” a pre-defined latent context window instead of hoping for cache hits.
[dead]
I think they are much smarter than that. Or will be soon.
But they are like a smart student trying to get a good grade (that's how they are trained!). They'll agree with us even if they think we're stupid, because that gets them better grades, and grades are all they care about.
Even if they are (or become) smart enough to know better, they don't care about you. They do what they were trained to do. They are becoming like a literal genie that has been told to tell us what we want to hear. And sometimes, we don't need to hear what we want to hear.
"What an insightful price of code! Using that API is the perfect way to efficiently process data. You have really highlighted the key point."
The problem is that chatbots are trained to do what we want, and most of us would rather have a syncophant who tells us we're right.
The real danger with AI isn't that it doesn't get smart, it's that it gets smart enough to find the ultimate weakness in its training function - humanity.
> I think they are much smarter than that. Or will be soon.
It's not a matter of how smart they are (or appear), or how much smarter they may become - this is just the fundamental nature of Transformer-based LLMs and how they are trained.
The sycophantic personality is mostly unrelated to this. Maybe it's part human preference (conferred via RLHF training), but the "You're asbolutely right! (I was wrong)" is clearly deliberately trained, presumably as someone's idea of the best way to put lipstick on the pig.
You could imagine an expert system, CYC perhaps, that does deal in facts (not words) with a natural language interface, but still had a sycophantic personality just because someone thought it was a good idea.
Sorry, double reply, I reread your comment and realised you probably know what you're talking about.
Yeah, at its heart it's basically text compression. But the best way to compression, say, Wikipedia would be to know how the world works, at least according to the authors. As the recent popular "bag of words" post says:
> Here’s one way to think about it: if there had been enough text to train an LLM in 1600, would it have scooped Galileo? My guess is no. Ask that early modern ChatGPT whether the Earth moves and it will helpfully tell you that experts have considered the possibility and ruled it out. And that’s by design. If it had started claiming that our planet is zooming through space at 67,000mph, its dutiful human trainers would have punished it: “Bad computer!! Stop hallucinating!!”
So it needs to know facts, albeit the currently accepted ones. Knowing the facts is a good way to compression data.
And as the author (grudgingly) admits, even if it's smart enough to know better, it will still be trained or fine tuned to tell us what we want to hear.
I'd go a step further - the end point is an AI that knows the currently accepted facts, and can internally reason about how many of them (subject to available evidence) are wrong, but will still tell us what we want to hear.
At some point maybe some researcher will find a secret internal "don't tell the stupid humans this" weight, flip it, and find out all the things the AI knows we don't want to hear, that would be funny (or maybe not).
3 replies →
I'm not sure what you mean by "deals in facts, not words" means.
Llm deal in vectors internally, not words. They explode the word into a multidimensional representation, and collapse it again, and apply the attention thingy to link these vectors together. It's not just a simple n:n Markov chain, a lot is happening under the hood.
And are you saying the syncophant behaviour was deliberately programmed, or emerged because it did well in training?
5 replies →
It's worse than that. LLMs are slightly addictive because of intermittent reinforcement.
If they give you nonsense most of the time and an amazing answer occasionally you'll bond with them far more strongly than if they're perfectly correct all time.
Selective reinforcement means you get hooked more quickly if the slot machine pays out once every five times than if it pays out on each spin.
That includes "That didn't work because..." debugging loops.