N=1, but I disagree strongly. I'm writing a hard-science science fiction story, and the physics of it is at (and frankly, beyond) my skillset. The story's plot has had to change over a dozen times as I realized errors in my application of physics in the story.
Throughout, I've been reviewing the physics with LLMs, mainly Gemini 3.1 Pro Preview, but also with Claude and OpenAI. Often I have the LLMs debate each other -- "My friend [another model] said XYZ about the physics, is that right or wrong?" In almost all cases, Gemini explains why the other models are wrong, and when I send its explanation to them, they concede it is right and they are wrong.
As I said, I did the above checks literally dozens of times as I wrote the story. And everything was dialed in: no further issues claimed by anyone, me or the LLMs.
Not with Fable. I managed to get it to review the story while it was running, and it listed out something like ten issues: some minor, some general knowledge-based, and two that were impressive:
1. It pointed out where Gemini (and I, and other LLMs) had missed a , resulting in values about 152 times larger than they should have been. I sent that to Gemini and it fully conceded that it had been wrong all along. 2. It pointed out a simple inconsistency in the application of special relativity (I thought I had that at least dialed in, but no :-/ ) that affected a very specific plot point. The story is novella-length, about 28,000 words long, and this is a point that was mentioned in the first two pages, and then not again until the very last page. And it's obvious, once you realize it. And I missed it. Gemini missed it. Claude and ChatGPT missed it.
Only Fable found it. Again, N=1, but that was a remarkable run I got out of it in the couple days it was available.
> Moving forward, the industry cannot continue to train bigger and bigger models since their intelligence not only plateaus but often will get worse
These are wild claims - why are we concluding that bigger models and more data = more hallucination? That’s actually the opposite of what’s been happening over the last couple years. Some models may still hallucinate more but they all hallucinate much less than the original 175B ChatGPT which was smaller and trained on (much) less data than anything current.
Edit: My mention of data comes from this quote:
> A shift is happening among major AI labs, who are becoming increasingly skeptical of endless parameter count and training data scaling
My take on the current situation: it seems clear that the industry has seen that there is still a lot left to squeeze out of sub-1T models. But for that you do need more, high-quality data in the distribution which you want to unlock capabilities for.
Here is something I would like people to chew on. Perhaps the smartest researchers in the world across multiple labs know more about this than we do? Perhaps they are aware of issues like the data wall and diminishing marginal returns. And perhaps they are being honest when they tell you there is no wall?
That’s not what your quotes said. They said bigger models = plateau in intelligence, nothing about more data or increased hallucinations
The relevant quote for what you’re talking about would be:
> It’s been proven that when a model is trained on large volumes of highly factual and non-theoretical data, it learns to always have an answer.
So there’s two separate claims: 1) bigger models have plateauing results 2) models trained on larger amounts of factual data have a higher hallucination rate
I’m pretty sure #1 is well known, I think OpenAI’s own research on scaling laws showed diminishing returns on parameter count and training data volume years ago. I don’t know what the support for #2 is besides for the actual post contents.
Yes, pretraining still exists. But for the past few years, pretraining by reading the internet is just the initial bootstrapping of LLM training. The RL training they get from bespoke training data, with very very different characteristics than what these armchair analyses claim, dominates these days.
There are a bounded number of (useful) derivations/combinations of Duff's device.
If Frontier Labs wish to reduce hallucinations on factual things, they will have to hire people (or the data providers will need to) to do fundamental research above and beyond what is available in extant literature and the web. IE if the LLMs want to lower precision error, they need to go out and actually find more expertise. If the wikipedia page for Pompey lacks data, where are they going to get it from? How would they even _identify_ that the page has holes?
Yes, they can digitize more books but that is untrustworthy data - if there were enough eyeballs on a particular work, it would be in the internet. If it's not, they'd need to hire the experts themselves. They need expert reviewers in virtually every interesting topic, which fundamentally is an intractable problem, especially since things change all the time. Maybe even uninteresting topics, too?
I dunno, it doesn't seem to me "more data" is the magic bullet here. Yeah, it will "help" but we're already on the flat part of the S shaped curve.
My take from trying to understand this stuff is some sort of algorithmic improvement is necessary to get another step change in how well LLMs perform in this area. I could be wrong!
There are journalists being hired to write Atlantic-worthy articles that exist only as LLM training data, because they're getting paid more than the Atlantic would pay them for it.
It's insane.
Yes, they are hiring the experts themselves. To create new knowledge above and beyond what's on the internet. To be locked away as LLM training data.
The largest characteristic of all of this new data is it is targeted at LLM's weak points.
It's not just more data, it's custom tutorials built for what LLMs struggle at.
1) Identify the gaps
2) Determine how to fix them
3) Implement a fix (especially if that fix is: identify and find experts)
4) And judge the result
How do they know [person] is an expert in [some field]? How do they find that person? How many experts are necessary to give the right information? How do we evaluate the results, especially if it's novel?
You can find a lot of people who disagree on many topics, and those turtles go all the way down.
I'm not in disagreement that your work will help reduce hallucinations and improve model performance! It is.
I predict (I hope I'm wrong!) that we're going to hit some asymptote that is not at 0% hallucinations (and I would even put a substantial nonzero probability that "overall" hallucination rate bottoms out at some minimum and then slowly grows because we just can't keep up with the new garbage we throw at it).
You just stumbled upon billion dollar businesses: Mercor, micro1, Scale AI, Surge AI, etc
They have a PhD from a top school, they are a licensed attorney, they are a licensed physician, a board certified cardiologist, etc.
They are constantly recruiting from these populations with well-paying side gigs.
> 4) And judge the result
That's what they pay the experts for. And to have experts review the other experts with peer review.
> You can find a lot of people who disagree on many topics, and those turtles go all the way down.
Which is why everything has to be well-calibrated and not just a hot take - a well reasoned opinion any expert would find fair.
Noone is really caring about hallucinations on point facts these days though, it is much more about complex reasoning tasks. Can they move the bar on the complexity of software LLMs do on their own? Can they get to a point where LLMs can begin to replace physicians? Financial advisors? Actuaries? etc.
The boundary is pretty thin there though. E.g., Gemini recently told me that a certain papers claims that two frameworks are mathematically equivalent, while the paper shows the opposite, and yesterday Google's AI overview told me that no World Cup matches were scheduled for that day despite their being several of them. The model probably used complex reasoning to arrive at both (incorrect) answers, but superficially they look like basic errors of fact.
You write the prompt, and then write rubrics to judge the responses, and you found something the model failed at. Congratulations, you just earned $500, now do it again.
But be careful: they are watching you and they don't want you giving away their secrets!
2. What criteria do such vendors typically require?
"As a side gig, I write novel software that solves problems no existing software does,"
and
"Yes, they are hiring the experts themselves. To create new knowledge above and beyond what's on the internet. To be locked away as LLM training data."
More likely you're joking and/or paranoid!8-))
This is actually really easy to do if you step out of web/gui/crud and into something where you won't find public code, most ever, because it's trade secret. For example, manufacturing.
Mercor, one of the larger vendors for contracting with experts to create bespoke data, says on their webpage they're paying $3M/day to their contractors for data.
So well into the billions of dollars a year for bespoke training data.
That's also ignoring the RLVR data labs can get from software - they can use the vibe coding sessions as training data as well without paying more.
They are just one of many.
A human asks a question, then writes rubrics to judge the LLMs response, so rather than evaluating a specific response, those rubrics can live on as the LLM evolves and gives different answers. There are more complex variants as well, but that's the basic principle.
Well known in a multiverse branch where Fable was a dud?
Here’s the paper from OpenAI where Dario himself was a co-author: https://arxiv.org/pdf/2001.08361
> We have observed consistent scalings of language model log-likelihood loss with non-embedding parameter count N, dataset size D, and optimized training computation Cmin, as encapsulated in Equations (1.5) and (1.6). Conversely, we find very weak dependence on many architectural and optimization hyperparameters. Since scalings with N,D,Cmin are power-laws, there are diminishing returns with increasing scale.
Right, what happened is everyone went to Fable and asked it to make the very best bicycle pelican SVG, no mistakes. And Fable's bicycle pelican SVGs were such timeless masterpieces, we all instantly got AI psychosis. Happily, you were immune to this.
I can’t prove it but I suspect there’s a bit of that going on.
Maybe GPT 5.5 is heavily nerfed due to lack of compute, memory, and energy?
I agree that it's farfetched to conclude that bigger models have pleateued.
I'm pretty sure it's mostly due to the training data quality. No idea, why this never gets mentioned in those discussions.
It was obvious right from the get go, that the scaling law just enabled some abilities, that were described by the underlying data and allowing the ANN to abstract it in the latent space.
> These are wild claims -
Indeed, it is not clear there was any actual intelligence at any point.
A lot of generated content sure, sometimes even useful, but not necessarily anything more.
If someone can "design a custom asyncio event loop policy in that overrides get_child_watcher()", I would call that person intelligent. Does that mean that person is not actually intelligent but a mere content creation machine?
Traditionally if you can create content, this shows you're intelligent. Created content is often called "intellectual" property. If a person can understand complex ideas and make connection between them, that is considered intellectual work. You have to be intelligent to do intellectual work. If a person can solve problems, this is also called intelligence. If the person can solve more complex problems, that person is said to have higher intelligence. This is often measured with a scale called IQ (Intelligence Quotient). There are other types of intelligence but they are basically the variations of the same ability. Most definitions of intelligence also involve an ability to adapt into the environment.
Since intelligence is such a broad concept what exactly is the difference between the actual intelligence and AI, other than one is natural and the other one is artificial?
I understand being anti-AI because of the very real societal concerns. But ignoring what is in front of you is not a solution.
Because that's what they measured in this case.
You can create contrived logic problems, but they often turn into language games because English is not formal logic.
And you can train on "monty hall" style problems, but those too are language games that are intriguing to humans but obvious when framed slightly differently.
In other words, model trainers are fighting against the overwhelming mediocrity of the training corpus (all of the recorded human output from history).
As models improve, the next phase will be models co-designed with humans to overcome these limits. The way we use language and the process we use to problem solve (we currently call this "orchestration") will evolve as part of this. Meatspace metaphors map badly when we have massive context and don't need the same limits. How different is hallucination from extrapolation, etc.
Much of the skepticism and confusion about LLMs is no different than a person of average intelligence hearing a highly intelligent person explain something and considering the explanation gibberish, then arrogantly accusing the intelligent person of being unhelpful.
Much like dogs were domesticated from wolves to have traits that make them good around humans, LLMs will evolve around our limits, around our arrogance, around our aesthetic biases and prejudices. Intelligence and rationality is fundamentally not what most humans want from an LLM.
There's an open question about whether this is theoretically possible, but it doesn't seem like it to me.
Human generated data is an effect of reasoning. Attempting to extract executive function from it is kind of like taking an anti-derivative of a function.
This has always seemed like the root of hallucinations to me. It sort of follows the parallels to lossy compression that a lot of people draw. You're extracting some characteristics by observing the relationship between tokens, and then trying to argue that those characteristics are equivalent to the thing that generated the original tokens.
Surely there's some sort of overlap there, but viewed that way, it seems obvious that more and more parameters and scaling won't solve the fundamental problem. There's only so much meaning you can extract from token relationships.
It's like trying to derive the shape of a flame from the smoke it produces.
The original intelligence that created those tokens was driven by a whole universe of inputs, from hormones to starlight to gravity, not to mention all of the strange things about consciousness and parapsychology that is so poorly understood.
The machines are definitely useful for a certain class of tasks - those that don't require much executive function, and the useful work mostly involves pattern matching.
The problem is, we seem to be mistaking effect for cause and imagining that these things have greater capabilities than they'll ever posess.
The investors that don't understand this are indeed going to learn a bitter lesson.
ofcourse you knew what you were doing but disappointing that this was top comment.
- A very parallel type of computation that is fast and generally accurate and integrates hundreds of variables. It’s sometimes labeled as intuition or system 1 thinking.
- A much slower, step by step, analytical type, commonly linked with your pre-frontal cortex (one of the newest parts of the brain). Sometimes called system 2 thinking.
Maybe the way the universe works is that all computation more or less is one of those two types. In which case, an LLM alone is only the first part, which is often right but its results also cannot ever be proven.
I'd also hesitate to attribute this difference in hallucination rates purely to model size. Yes, GLM-5.2 hallucinates much less frequently than DeepSeek-V4 Pro with twice as many parameters, but DeepSeek-V4 Flash is less than half the size of GLM-5.2 and tops the AA-Omniscience hallucination index. Opus 4.8, which is likely larger than DeepSeek-V4 Pro, has a 36% hallucination rate on the index, above GLM-5.2's 28%, but way below the DeepSeek numbers. Opus also has a 47% accuracy rate vs GLM-5.2's 25%. If you use these numbers to calculate the absolute hallucination rate (i.e., the number of hallucinated responses divided by the total number of responses), you get 19% for Opus and 21% for GLM-5.2.
So yes, all else equal larger models may be more prone to hallucination in scenarios where they don't know the answer, but there are a lot of other factors that affect hallucination rates, and it's not totally clear that this is the main metric that's worth tracking.
It obviously breaks down with humans too, given we so easily hallucinate and confuse things we "know". However i still suspect we're more reliable at probing information we've experienced vs not. Even if the case of poisoned knowledge, eg a crime scene accidentally implying information to a witness that the witness doesn't actually know, we still "know" that poisoned information via incorrect inference. Ie we "experienced" it.
Wonder what architecture would allow for this style of information/weight probing for an LLM.
If Opus gets all but the hardest questions right, it might have a higher hallucination rate because the questions it gets wrong are the questions where verification or hallucination detection are the most difficult
Something about the cost model of US near frontier has the cattle prod out whenever a model is uncertain but thrashes on whether to search. Search flinch is roughly all hallucination.
I don't even wait for the model's turn, if there's a man page or Hoogle hit, stuff the last prefix cache cut point. You come out ahead.
I’m not sure how to explain it, but the more I see LLM-written code the more I feel it’s bad code doing a good job of masquerading as good code. I think this take will become less-hot in the next year or two when we see enterprise greenfield projects that were created entirely with LLM “assistance” go to prod. I think we’ll find that the code is difficult for humans to read, understand, debug, and extend- and I think the larger the codebase the harder it will be for LLMs to maintain. More opportunity for hallucination, larger context windows needed, more tokens bought and spent for smaller and smaller code changes. I think the more code an LLM writes for an app, the worse that codebase becomes.
I strongly suspect most closed source code developed under commercial or internal pressure is pretty awful after a few years of development.
All LLM code has to do is suck less than existing code. And that's presuming the code quality doesn't improve as the models, the harnesses and our ways of working with them improve.
LLMs doesn't have this benefit. You forget to add the correct to the system prompt, and the LLM will repeat the same mistake over and over, and worse than that, their mistakes aren't based on their understanding, it's basically random guesses.
Humans, even bad coders, still seem to have some sort of architecture in mind, even if it's spaghetti, whereas LLMs (obviously) don't think more than a few steps, and never about the full scope of what they're contributing too, and on purpose too, because you want the context to be as small as possible when you work with LLMs.
With LLMs you need to thread carefully between "What does the LLM need to know?" and "Can I skip passing this to the LLM this time?" while a human you can more or less dump them everything you sit on, and let them shift it through, and they'll mostly make it out OK.
i dont see why software engineers are paid so well, and are so hard to hire?
just dump a bunch of requirements on a homeless person and itll just work out
Whilst I don't claim any true "understanding" as that is a very loaded term that doesn't mean it's just random guesses.
Anyone using recent LLM coding agents on a regular basis would probably agree that there's something going on that fits some non-athropomorphizing, non-sentience-assigning definition of "understanding"
As for the point about improvement - I think that's an orthogonal issue to the overall code quality. With regard to human codebases - there's plenty of scenarios that negate the improvement of individuals. We're comparing organizations with LLMs - not individuals with LLMs and that makes a significant difference.
But anyway, let the LLM verify the code to give advice on improvements but don't let it write code unverified. That's my opinion on it anyway.
But as soon as you do minimal reviews and high-level corrections, applications turn out just fine.
Can there be bugs? Sure. That's the price of not reading or understanding every line. It should depend on the criticality of your software how much of these you tolerate and how much you don't (reviewing, understanding, testing everything 100% like you were used to if you had written it yourself will kill most if not all of your gained speed)
But I never got the impression of unmaintainability or unfixable bugs.
Actually the other side around: A really good cleanup pass, architectural changes, or bugfixes are seldom more than a few prompts and 2 hours away, provided your overall base is decent and you actually gave a fuck from the start.
I've yet to come across a human developer who's output would meet this standard, despite writing every line.
In fact, having an LLM review our code is catching quite a few bugs before it reaches QA.
The humans may skip unit tests and need reminding; the AI always write unit tests once it's in AGENTS.md or whatever, but my experience* was that 5-10% of the time the LLM's attempt at a "test" would, instead of executing the code and examining the results, open the source code as a text file and run a regex to find/exclude certain substrings.
* At the start of this year, because Anthropic and OpenAI were both offering free trials. IDK how much things have changed since then, some things change fast in this domain, other things don't.
Riding the exponential means you have to update priors more often.
My observation is that they are equally bad and hard to maintain or even more so than the new ones.
One thing I’ve noticed is that the LLM assisted ones have a lot more comments which is nice but take more time to read.
And they do it faster than any human developer.
They clearly are only assistants for the moment, you can use them to do work ... but only if you could do the said work yourself alone in the first place.
I'm an experienced developer, but I don't count myself as a web dev or a python dev; I can review the web and python stuff I get out of the AI (sometimes I need to ask the AI follow-up questions so I can find official documentation for what it did), but I can't write it.
But the difference I allude to here is more like how "book reviewer" is a different job than "book author": yes, if you can review a book, you can also write one. Eventually.
On a more serious note, I think the problem will be the inability to handle/maintain the systems once they are too big and nobody has no idea what's inside of them or what they do.
Is this a good idea? Probably not—in the past we would only do that when the architecture was causing serious problems since it always has tons of behaviors that will accidentally not get carried forward, some of which are load bearing and will cause bugs.
Now we can do it in an afternoon and get the same long term bug behavior.
Do you have a cite for this?
If a human makes up some bullshit lie, I wouldn't accuse them of making it up only if they actually knew the correct answer. If you don't know, the only correct answer is I don't know. Any other answer is made up bullshit. Why is it only a hallucination if and only if the LLM contains the answer? If you make something up it's still wrong. It shouldn't matter if you could give the correct answer. You didn't, and instead invented some bullshit instead?
Follow up question, how can I apply this rule set to the next test I have to take? I'd love to be able to use "I didn't know" as the excuse for why I made something up.
edit:
> and it's not totally clear that this is the main metric that's worth tracking.
I don't know, the rate at which some model is willing to make up something feels useful. If the argument I see repeated on HN so much is that it's impossible to completely get rid of hallucinations; being able to choose a model that's less likely to invent some lie seems like a positive trait, no?
Either way, I'm happy to agree that a restrictive definition, where a lie doesn't count as a hallucination iff the model doesn't know the answer feels strictly, infinitely less useful than an exact error rate. What percentage of emitted tokens are misleading would be useful for me. Anyone know any group that's attempted to quantify the global error rate?
I don't understand the point of describing it as 3 possible outcomes. I'm objecting to it because the only reason I would do something like that would be to obscure the severity of the model defects.
I would expect any engineer to evaluate any model using accuracy, (error rate), and usefulness (definitive answer rate), as strictly independent metrics. Did it answer, and if it answered, did emit incorrect or misleading information and how many quantifiable bits of each.
The false negative rate (model confirmed to contain the requested output/information via other method but was unable to for the given test) is significant, but given a non-definitive answer is significantly different from a definitive and incorrect answer. Why would you want to group hallucinations?
I'm sure I'm missing something, but the reason I suspect that's how it's done, is to [intentionally] obscure the actual meaningful metrics.
Number/rate of useful answers (correct and incorrect) and error rate (given any answer how often will that answer be defective in some way).
To be clear, I'm differentiating hallucination rate from eagerness to answer, even though they're obviously linked because I believe presenting 20 correct answers, 20 incorrect answers, and 60 abstentions as a hallucination rate of 25% as obviously malicious. If I give you 40 answers, 20 correct and 20 incorrect. the error rate is 50% and if it refused an additional 60 times, it's usefulness rate would be 40%... arguably 20% depending on how strict you choose to be about the definition of useful.
And, just in case it's unclear, I'm using you exclusively rhetorically. I don't think you personally are being misleading, only that you're explaining how it's done... but that's the problem isn't it.
When pushed, I then start thinking and realise my mistake. System 1 vs 2?
When someone asks a question, if I don't know the answer; I say I don't know.
System 1 vs 2 doesn't really matter... I won't use an LLM that's willing to make up random shit. Equally I also won't work with a human who does that. Trust and confidence a system will function correctly is an important quality, in both humans and genai
In other words, you shouldn't choose the model that hallucinates the least without detailed prompting, since a well-crafted agents.md clause should go a long way to improving output, and almost certainly the top scoring order will be different. To the point that I don't find this type of raw comparison useful beyond maybe 'make sure you test that one with more explicit prompts'.
You're prompting it wrong is quickly becoming the new, you're holding it wrong.
It's wild how willing software engineers are to blame the user when the actual problem is their own defective design.
Ideally we all, as an industry, will stop accepting this as reasonable excuse for the demonstrated incompetence
Now granted, if the boat salesmen were pushing hard on the idea that the boat would fly and even put little wings on the side and I bought the boat I might get really angry when I found out that it didn't fly. And I might angrily storm into the salesroom yelling about how the design is defective. But if someone pointed out 'hey, it's a boat perhaps you should stick to sailing around in it and stop getting your undies in a bundle about it not flying' the correct response is probably to take a closer look, ignore the salesmen, and cruise around the lake. LLM's are quite handy at some things and have some weird limits. Learn the limits, enjoy your time at sea.
It's not that you're holding it wrong, you're just wrong for expecting it to work the way it's described (able to one shot most problems these days).
Sam Altman himself had a blog post about this a while ago that seemed to suggest this thought, so I guess it's obvious to everyone. But if that is so I assume it's just not as easy in practice.
AA-Omniscience is the only AI benchmark I know of where randomly guessing gets you a lower average score than answering all questions with "I don't know"
For your scenario the confident confident strategy will give average of -90. Saying I dont't know to all will give 0.
A lot of models have negative AA-Omniscience Index.
They also do have AA-Omniscience Accuracy and AA-Omniscience Hallucination Rate that handle "I don't knows" differently.
They are much better incentives. In real life a wrong answer is much more damaging than a don't know.
If your example had "Validate any details before sharing them with the user, with multiple sources" as the system prompt, it was using a model that is strong at following system prompts precisely and had access to some basic tools, then it'd spend maybe minutes more, but the answer would have been way more accurate.
But no, Google want "the new search results" (LLM hallucinations) to be on top, so we end up with "sounds plausible" answers instead "Collection of evidence from reliable/semi-reliable" or similar, which sucks. We could have quality, but it's too expensive/slow, so we get slop instead, just to maximize for speed and convenience.
Like when agent 1 says X, agent 2 verifies it as Y and the original question ends up being some weird amalgamation of Z with additional ”this is really true” statements sprinkled on top.
I agree Google responses hurt more than help, but I’ve also gotten identical outcomes of 40min self-reasoning Opus threads (it’s less common obviously).
Yeah, seems what grounds agents right now is quite literally human thoughts and text, so if you're doing something like that, you really need to pass the original user prompt through the entire way, for every "child" to keep in mind the final thing, otherwise it does seem to spiral out of control.
I don't know. Is it?
I don't think anyone is trying to add "a coherent worldview" by reducing hallucinations, not sure how that even realistically could be aim.
What people want, is for the models to stop giving confident answers that are clearly incorrect. Yes, it won't lead to "a coherent worldview", but it'll at least stop wasting people's time if the model said "You know what, what you said doesn't make sense / isn't clear, is what you mean .... ?" or even "I'm not sure" or "I don't know".
Currently, if you have the wrong starting point, ask the model to do something, they more often than not just go ahead and do that, misunderstandings or not. They seem optimized to never push back, unless you prompt for that, and most seem to favor "I'm just gonna assume X" rather than taking a step back and figuring out how to not assume. Again, unless you prompt against that behaviour/steering it into a different workflow.
Training an extra "don't know" token means you have to build a moat between every other token. Between "yes" and "no", you don't have a muddled noisy area where both "yes" and "no" have relatively high probabilities, you need a new peak where "don't know" is higher. Then you just have new muddled areas between "yes" and "don't know", and "don't know" and "no". That requires even more finesse to train another answer in between.
Instead, you could check whether multiple options are about equally likely. But then you have to check if they are actually synonyms, like are the top two choices "Genève" and "Geneva", which is a good sign that the model knows the answer? Or are the top two "yes" and "no"?
The task was simple, using the MS-MARCO[0] dataset which contains queries, search results, answers, I made a training set that has:
1. Questions paired with real results supporting them (mixed with some irrelevant results), and a correct answer
2. Questions paired only with irrelevant results, with the answer “No answer present”
The dataset was huge (close to 1M samples), and I trained using different techniques, from SFT (just mimicking the dataset) to DPO (good answer contrasted with a bad answer for the same user query) to GRPO (verifier that checks my annotations whether an answer was present or not)
Lo and behold, this didn’t reduce hallucination, rather made it much worse. Now the model started claiming “No answer present” even when it is, or even when the question didn’t need search results in the first place (simple stuff like what is X+Y).
Now you could argue that my training was basic compared to what frontier labs could do. Yet I think it hints at a more profound limitation. LLMs are finicky and don’t have a neat understand of things from first principles (list of search results, check relevance of result to user query, if answers are below a certain threshold of relevance then don’t consider them to answer …).
tl;dr: not as simple as one might think, perhaps not attainable at all.
You can definitely tune a model to say "I don't know" more often but it will cost you performance, the model will reject some questions that it could answer meaningfully. In the degenerate case the model could collapse predicting that sequence always or almost always.
But I guess my logic breaks down here a bit, because if there is such a thing as a validated answer, then the correct answer is in fact never uncertainty. The correct answer is to continue post training until the model gets it right. So perhaps the real answer is to create RLVR tasks where the valid answer is "I don't know" and nothing else, like this benchmark does. Or maybe that doesn't work either, no matter how many you create.
I feel as though there is some kind of philosophical lesson to be had from how hard hallucinations are to get rid of. Maybe, similarly to humans, successful models are often "arrogant" in a sense. Perhaps you just never solve an Erdös problem without some degree of self deception that it's possible for you to do so. In this line of thinking, greatness in humans is actually not related to humility, but just being so good that you actually get things right when you try. Expressing humility is of course something great people tend to do, but I'm referring to what happens under the hood.
If you squint a bit, that's kinda the trend with models. The useful ones are not that much less likely to hallucinate, they are just good enough that they tend to get it right. This comparison is of course probably not even remotely correct, but at least it's fun to anthropomorphize a bit.
1) Has a certain standard of evidence been met?
2) Are the related arguments free of logical inconsistencies?
We can train the LLMs to do 2, and maybe even 1 to some extent (exactly what quality of evidence a computer can practically gather is limited). But that isn't going to get rid of hallucinations, for the same reason courts are hit-and-miss or the conclusions of studies often aren't very reliable. These techniques help, but sometimes they still get people to say things that, on close inspection, turn out to be nonsense. And those best-effort approaches are too much to expect for most questions an LLM will be handed which are informal, low stakes and don't need strong supporting evidence or logical rigour.
I think it is underestimated how many LLM-style hallucinations people themselves have. It just isn't obvious because most humans have a strategy of only repeating what the herd says after it has been socially vetted, which makes their individual eccentricities less obvious.
TLDR; I don't think it looks like an easy problem for RLVR, it looks technically unsolvable. Even making progress requires a philosophical breakthrough on the nature of truth so that the objective function can be established.
But even in muddy fields of reality like medicine, there are objective facts to be found. When someone comes into an ER with chest pain, you often find a true, undeniable reason for why that is happening. If their lung has collapsed, a coronary artery is clogged or the aortic artery is dissecting, even if you don't find that out it tends to be clear in retrospect. The area of reality that becomes muddy is when use proxy signals to try to figure out who gets promoted to expensive/harmful examinations we can make final conclusions from, or the cases that don't fit cleanly into one bucket or the other. But very often, the gold standard truly is golden.
Of course, many realms of reality cannot be verified in this way. But I'd argue that there are quite a few that can.
Does mathematics count as not a hallucination though? Particularly in pure mathematics they take a certain pride coming up with wild concepts as unrooted as possible in anything relevant to human existence. The name of the game is purely about maintaining internal logical consistency - which is something an AI can do while hallucinating.
AI hallucinations in maths might be logically consistent or not be. But in that particular case it starts to get a bit iffy what we call it when someone imagines something that doesn't exist. This gets back to the thing where we can train AIs to be logically consistent, but we can't force that consistency to be grounded in any particular universe. Ie, it'll hallucinate but in a very well rationalised way - coincidentally mimicking how a number of mathematicians seem to approach life.
This is the central issue; there is a very real trade-off between facts and verifiablity. Mathematics is perfectly verifiable because it is fact free. We don't have a reliable general system to verify facts. We do have reliable systems for checking arguments (logic).
Yes, there are mathematical concepts that seem to exist purely in the realm of mathematics, but maths often touches reality in a consistent way that reflect experimental results. This seems to imply that there is more to mathematics than just internal consistency. And the parts that do not correspond to any observation right now, might just reach out and touch reality in the future. It is possible to create logically consistent systems that have nothing to do with reality, but this is not the mathematics that most mathematicians are thinking about.
Observation is the final arbiter of fact. Maybe we don't have a general system to verify ALL facts, but many facts are 100% verifiable, although not most of them. "Beyond reasonable doubt" is of course the highest level of fact as far as the scientific method is concerned, but some facts are so far beyond reasonable doubt that you might as well just call them true. In the average living human body, there is a particular clump of tissues that consistently corresponds a concept most experts would describe as a "heart", and it does in fact pump blood. True fact.
We can rank them based on how much they know and people will gravitate towards those that do know more.
It's a market after all.
Haven't heard about that law, but seems unlikely we can come up with ("discover") any sort of law that uses a concept ("truth") humans can't even agree what it means, and that's not for a lack of trying, we've been trying to figure it out for millenniums already with no end in sight.
Might be why we're already rarely seeing models output an "I don't know".
"Confidently incorrect" has negative value. At best, a human realizes the answer is wrong and At worst, the incorrect information makes is not identified and can cause untold damage. By having the potential to be so severely wrong, it lessens the value of correct answers because there is a lower confidence value on their output.
If someone sold you a "Solved all your problems" machine, and it suddenly doesn't solve all your problems, then probably no, you shouldn't pay.
But the way I'm being sold LLMs, is basically "A text generator that gives your plausible-sounding human text that sometimes hallucinates and gets things wrong, based on your input", then regardless of what the outcome is, I still made use of the "Input > Output" part, which is what I bought into, so I should still pay for that.
Now of course bunch of people will say they been sold the former, but the companies themselves seem to be selling the latter. That's my perspective from a person who doesn't follow "influencers" and what not though, which seem to be selling the public on the former rather than the latter.
My ask:
> In a couple sentences, explain to me the product I'm being sold with ChatGPT. What does it do for me?
The Reply from ChatGPT:
> ChatGPT is a conversational AI that helps you think, create, learn, analyze, and get things done faster. You can use it to answer questions, draft and edit writing, summarize information, brainstorm ideas, learn new topics, write code, plan projects, and increasingly act as an assistant that can search for information, work with documents, generate images, and help complete tasks.
> In simple terms: you're buying access to an AI that turns natural language into useful work—saving time, expanding your capabilities, and giving you an always-available collaborator for both everyday tasks and specialized knowledge work.
This sounds much more like the former, a "solve all your problems" machine.... not a plausible-sounding text generation machine.
Only two weeks ago Sam Altman said their new data center "could" be where cancer gets cured[0]. It is only the people who deeply understand AI who see it as a text generator of plausible-sounding text. That isn't what the marketing department, the CEO, or the product itself seem to be saying. I'm using OpenAI as the example here, but the others don't seem much different.
> Can I trust the output you give me?
And I assume it explains what to trust VS not.
I think in the bottom you should also see something like "Any text can contain mistakes" or similar too, which I know is a far cry from what some people push in the press in regards to capabilities, but I still don't see the platforms themselves as lying about this, while I do see a bunch of people constantly over-hyping the possibilities.
I'm not sure why "can I trust the output you give me?" would be a logical followup to the first response it gave me, seeing as it's response didn't say anything about hallucinations or mistakes. It said it could do "useful work" with all kinds of examples, including "specialized knowledge work".
The note under the text field, in gray as to not draw the user's attention, feels more like a CYA line from the lawyers, rather than an instruction they really want users to take to heart. That line also doesn't appear on the main home page. I only shows up after the first prompt is submitted and focus shifts to the conversation. I don't think a CYA line in gray fine print is enough to make users understand it's a plausible-sounding text generation machine instead of an answer machine. Even if I ask that point blank it gives a wordy... yes, but not really, it's being debated by philosophers... response.
> If you can dream it, Claude can help you do it. Claude can process large amounts of information, brainstorm ideas, generate text and code, help you understand subjects, coach you through difficult situations, simplify your busywork so you can focus on what matters most, and so much more.
What marketing copy have you read for LLMs that is like you mentioned?
> But the way I'm being sold LLMs, is basically "A text generator that gives your plausible-sounding human text that sometimes hallucinates and gets things wrong, based on your input"
so, thats all.
Curiously, this post and article is the only submission and interaction the OP has made, and these claims support the product he's intending to release.
Wow! I already knew from previous research shared here that hallucinations are a fundamental problem for LLMs and likely to be unfixable, just like prompt injection, but I didn't realize the hallucination rates were so bad!
Everyone has been acting like the best models only hallucinate in edge cases, but even the best performing one mentioned here - GLM-5.2 - has a hallucination rate of 28% when it doesn't "know" the answer to something.
That said, I think the title on the blog - "Bigger models are not the way" is probably more fitting and touches on what should be even bigger news. If bigger models and bigger training sets have already stopped producing proportional returns, then it seems likely we are already near the top of the S-curve. That's huge news, considering the valuation of companies like OpenAI and xAI is largely based around the (absurd) idea of ever increasing scaling from these models.
The question-tokens define the answer-tokens. That's it. The art relies in clustering the relevant weights together.
Circuits which emerge in the layers during training are much more complicated than a simple Bayesian relation.
There can be, you don't know if the closed source models aren't using something like DeepSeek's Engram.
While DeepSeek describe this as "knowledge lookup", what Engram is really trying to do is separate dynamic reasoning from static pattern recall, with the static patterns just being word-level n-gram statistics, not declarative facts/knowledge.
Just because 2-3 words often appear together in a sequence doesn't mean they represent a fact or truth (or falsehood) - it is just an n-gram statistical regularity.
If Engram helps reduce LLM GPU memory and FLOP requirements then that is great, but it's not a solution for Hallucination.
This implies that bigger models are more likely to hallucinate? That doesn't match my experience.
Such a weird thing to start with. The legal status of Fable does not mean that it's not intelligent. If anything, the problem is the opposite, someone thinks it's too intelligent (and/or that Anthropic wouldn't share its last gen intelligent models on the terms the government demanded).
I'm already hallucinating about how this could work and it involves catapults
Hallucinations all the way down...
Being an LLM is easy!
In addition, I think that during HFRL, the labs has a bias for interesting answers that admit a solution and under represent the "bad" questions that admit no good answer. In addition they probably do less effort to HFRL on questions the model should admit it doesn't know.
As humans we have been trained all our lives, in the real world, to be confronted with questions we don't know the response right away and we learned to very quickly assess that we don't know or that we are not sure about the answer.
Another thing we have and LLM have not is fear. We have an amygdala in our brain, separated from the logic thinking part, that can raise a signal of fear so that we get much more carefully about what we say. On the other LLM has no fear organ like the amygdala and just learn to respond based on the patterns in it's training corpus. It never "fears" looking bad or being fired because it gave a wrong answer so it can merrily give perfectly wrong answers.
So, we see hallucination rates can be improved with training but currently the lab are not optimizing for that because there is an high stake race to get the most intelligent and capable model.
Alternatively I can see creating a separate amygdala-like organ for an LLM and that organ may asynchronously fires signal, based on the user prompt and the LLM thinking trace, to inject into the LLM reasoning a fear signal so that it can steer it's answer to something more safe.
However the fear has to arise in the first place, to raise the alert.
And, of course, it was burning 10 times more tokens for this output.
Idk if it was the harness (OpenCode), my AGENT or my prompts, but I was getting exactly what I wanted, and quickly.
With GPT-5.5 it tries to play smart, takes much more times and is often stuck on basic stuff that DeepSeek solves oneshot.
You have any session logs or similar that shows this thing? Never once, since I started using the codex TUI when it became available, has GPT models gotten stuck on something another model breeze through, I quite literally run every prompt I do through multiple providers, this would be very visible very quickly for me.
I remember trying every -codex variant of the models and could never get them to be productive for tasks taking longer than 5-10 minutes, compared to GPT 5.5 which quite literally worked through the night day (with the /goal feature), and actually had something valuable and useful in the end this morning that wasn't exploding in LOC and complexity. I don't think any of the -codex variants would have been able to do this at all, based on how they worked when I last used them.
They're basing this all on public benchmarks which stopped being a reliable indicator of anything the last 2-3 years. Of course it'll be filled with more terrible lines of thoughts.
People really need to stop placing such importance on public benchmarks. They're valuable for comparing very close models, useful to evaluate if quantization and similar have negative impact, but you're not gonna be able to tell if one model is better than the other based on one scoring a few percentage points higher than a completely different model.
GLM 5.2 tends to stray way more than and 5.1. It also hallucinates you things subtly: morphs requirements, makes unfounded conclusions. This output is not something I experienced in any model I seen so far.
In coding it's especially annoying because it steers whole request. E.g. I give instruction: "make we a Rust-WASM-Canvas app" and GLM 5.2 goes like "Oh user surely doesn't mean that. I'll better build Dioxus app instead".
I've had more success with creating a plan first and then implementing it in (short-lived) sub-agents.
Ironically good software architecture patterns (small functions, single responsibility) heavily impact the performance of these models as well. They do surprisingly well in well architectured codebases.
They do very poorly in anything that's a mess where Opus and GPT 5.5 still get reasonable performance.
the oss models are impressive but it's pretty clear how quickly they fall off when you try to use them outside of a narrow set of problems they benchmarked well on when compared to opus/5.5
The article uses the example of GLM being smaller than DeepSeek, yet better on hallucinations as "smaller can be good too"
But the GLM family itself is scaling up fast: GLM-5.x family is 754B, double the previous generation of GLM-4.x
> comes within just 4 points of GPT-5.5 and 9 points of Fable 5
9 percentage points IS a big difference
From how they measure it, a model that simply answers "I don't know." to any prompt would be the one hallucinates the least. So it's not surprising at all that a smaller model can perform better.
a key method to help with hallucinations is to provide good sources when asking questions (context engineering / knowledge base)
What about using two models, with a smaller model used for this kind of negative reasoning?
We really don't know what the actual reason is given the politics at play. I would bet more on the Trump administration looking for any excuse to punish Anthropic
Maybe the pre-2024 users do, but I've seen plenty of those exact "frontier models never hallucinate" comments on HN as well
> Why is everyone expecting LLMs to be like the Star Trek computer?
Because they are often marketed as magic AIs, not as mere language models.
[0] https://bpspsychub.onlinelibrary.wiley.com/doi/10.1111/bjso....
"they say u hallucinate 3x more than GLM 5.2, whats your comeback to this? do i need to dump u? $article"