upvote
Knowledge and memory

(www.robinsloan.com)

OpenAI just recently took a systematic look into why models hallucinate [0][1].

The article you shared raises an interesting point by comparing human memory with LLMs, but I think the analogy can only go so far. They're too distinct to explain hallucinations simply as a lack of meta-cognition or meta-memory. These systems are more like alien minds, and allegories risk introducing imprecision when we're trying to debug and understand their behavior.

OpenAI's paper instead identifies hallucinations as a bug in training objectives and benchmarks, and is grounding the explanation in first principles and the mechanics of ML.

Metaphors are useful for creativity, but less so when it comes to debugging and understanding, especially now that the systematic view is this advanced.

[0] https://openai.com/index/why-language-models-hallucinate/?ut... [1] https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4a...

reply
I haven't read the full paper yet, but my intuition is that hallucinations are a byproduct of models having too much information that needs to be compressed for generalizing.

We already know that larger models hallucinate less since they can store more information, are there any smaller models which hallucinate less

reply
I have temporal lobe epilepsy. The premise of this article seems quite wrong to me.

From the article: "I think it’s because I don’t only know things: I remember learning them. My knowledge is sedimentary, and I can “feel” the position and solidity of different facts and ideas in that mass. I can feel, too, the airy disconnect of a guess."

I think this can be rephrased as a statement that episodic memory (i.e. recalling the act of learning) is associated with semantic memory (i.e. recalling the fact itself). And for people with more normal brains, it seems it often is.

For people with temporal lobe epilepsy, in many cases the episodic memory isn't there. TLE frequently damages the hippocampus. Often, immediately after taking my meds, I can't recall doing so. I only know I did because I mark it in a notebook when I do it. Things like that are common. However, I can absolutely learn things like structure of systems, API calls, and the like. I believe that I perceive when I'm not sure of something, or that I'm guessing.

Of course, this kind of perception is difficult to verify outside of a study. A lot of memory is very unreliable, even for normal people. Us lucky people with TLE are often more aware of this, because in my case, I have very vivid memories of events that certainly did not occur. These memories are some of the most vivid I have; they seem more real than reality. I have memories of events that I can't verify by comparing them to the current state of the world; I have no idea if they happened or not. Even with all that, I believe my semantic memory is reasonably good; I can do my job as a software engineer.

reply
Yesterday I asked one LLM about certain provisions of Ukrainian law, where severity threshold for a financial crime was specified indirectly through a certain well known constant. The machine got it wrong and when asked to give sources cited the respective law referencing a similary sounding unit. Amazingly it gave the correct English tranlation but gave me the wrong original in Ukrainian.

I guess it merged two tokens why learning the text.

Amazingly it also knows about difference between two constants, but referrs to the wrong one in both calculations and in hallucinating the quote.

It's tedious to always check for stuff like this.

Then I asked a different LLM and it turned out that actually the constant is monkey patched for specific contexts and both me and the first lying machine were wrong

reply
The fundamental limitation of LLMs is that they represent knowledge as parametric curves, and their generalization is merely interpolation of those curves. This can only ever produce results that correlate with facts (training data), not ones that are causally derived from them, which makes hallucinations inevitable. Same as with human memory.
reply
I'm not convinced the brain stores memories, or that memory storage is required for human intelligence. And we "hallucinate" all the time. See: eye witness testimony being wrong regularly, "paranormal" experiences etc.

It's a statement that /feels/ true, because we can all look "inside" our heads and "see" memories and facts. But we may as well be re-constructing facts on the fly, just as re-construct reality itself to sense it.

reply
What do you mean, you're not convinced that the brain stores memories? What is happening in the brain when you have an experience and later recall that same experience? It might not be perfect recall but something is being stored.
reply
I mean an LLM (bad example, but good enough for what I'm trying to convey) doesn't need any sort of "memory" to be able to reconstruct something that looks like intelligence. It stores weights, and can re-assemble "facts" from those weights, independent of the meaning or significance of those facts. It's possible the brain is similar, on a much more refined scale. My brain certainly doesn't store 35,000 instances of my mum's image to help me identify her, just an averaged image to help me know when I'm looking at my mum.

The brain definitely stores things, and retrieval and processing are key to the behaviour that comes out the other end, but whether it's "memory" like what this article tries to define, I'm not sure. The article makes it a point to talk about instances where /lack/ of a memory is a sign of the brain doing something different from an LLM, but the brain is pretty happy to "make up" a "memory", from all of my reading and understanding.

reply
Addressing the second paragraph here - while conflation and reconsolidation are real phenomena, it's also quite clear that most humans form episodic memories. Some quite clearly have incredible abilities in this regard [2].

A distinction between semantic (facts/concepts) and episodic (specific experiences) declarative memories are fairly well established since at least the 1970s. That the latter is required to construct the former is also long posited, with reasonable evidence [1]. Similarly, there's a slightly more recent distinction between "recollecting" (i.e., similar to the author's "I can remember the event of learning this") and "knowing" (i.e., "I know this but don't remember why"), with differences in hypothesized recall mechanisms [3].

[1] https://www.science.org/doi/full/10.1126/science.277.5324.33... or many other reviews by Eichenbaum, Squire, Milner, etc

[2] https://youtu.be/hpTCZ-hO6iI?si=FeFv8MGmHTzkLd8p

[3] https://psycnet.apa.org/record/1995-42814-001

reply
Is the idea cars start a fact or a concept? Is the certainty I remember how this particular car starts a fact?

Once we begin to disengage from the arbitrariness inherent in arbitrary metaphors, and rely on what actually generates memories (action-neural-spatial-syntax), we can study what's really happening in the allocortex's distribution of cues between sense/emotion into memory.

Until then we will simply be trapped in falsely segregated ideas of episodic/semantic.

reply
You've prompted some interesting memories out of my teenage years and my first truck.

I can recall the experience of getting in on a cold morning, pumping the throttle pedal three times to activate the semi-automatic choke, starting it, and getting out to clear frost off the window while it warmed up a little. The tactile feeling and squeak of the throttle linkage, the sound of the starter motor, the hollow sound of the door closing, and the noxious exhaust from the cold start (which I haven't smelled in 30 years). I remember how my little plastic window scraper sounded when scraping the glass, and even how the defrost vents made two regions which were always easier to scrape. But, I cannot really remember a specific episode of this on a certain date or leading to a particular trip.

On the other hand, I do have an episodic memory on my final trip in this truck. It was sliding off an icy road, rolling over, and sledding down a steep slope. I remember the ruptured, snow-filled windshield, and the sound of the engine idling upside-down. I remember the slow motion way the whole crash unfolded, and the clothes I was wearing as I crawled back to the roadway.

Ironically, I have more emotional context with the generic cold-start memory. It brings with it some vignette of teenage eagerness and pride in having this vehicle and the freedom it represented. The crash is more dissociated, a movie I was watching from within myself. I can meta-remember that I was very distressed afterward, but those feelings are not connected with the memory during recall.

reply
Fantastic to read this. Thank you for it, that you can interpret these associations/dissociations still is incredible.
reply
"That the latter is required to construct the former is also long posited, with reasonable evidence [1]."

This is very interesting to me. I have temporal lobe epilepsy. My episodic memory is quite poor. However, I believe I'm fairly good at learning new facts (i.e. semantic memory). Perhaps my belief is an illusion, or I'm really only learning facts when my episodic memory is less impaired (which happens; it varies from hour to hour). It's difficult for me to tell of course.

reply
The article isn't about LLMs storing things - it's about why they hallucinate, which is in large part due to the fact that they just deal in word statistics not facts, but also (the point of the article) that they have no episodic memories, or any personal experience of any sort for that matter.

Humans can generally differentiate between when they know something or not, and I'd agree with the article that this is because we tend to remember how we know things, and also have different levels of confidence according to source. Personal experience trumps watching someone else, which trumps hearing or being taught it from a reliable source, which trumps having read something on Twitter or some grafitti on a bathroom stall. To the LLM all text is just statistics, and it has no personal experience to lean to to self-check and say "hmm, I can't recall ever learning that - I'm drawing blanks".

Frankly it's silly to compare LLMs (Transformers) and brains. An LLM was only every meant to be a linguistics model, not a brain or cognitive architecture. I think people get confused because if spits out human text and so people anthropomorphize it and start thinking it's got some human-like capabilities under the hood when it is in fact - surprise surprise - just a pass-thru stack of Transformer layers. A language model.

reply
Hey, I know what the article wanted to say, see the last two-ish sentences of my previous response. My point, is that the article might be mis-interpreting what the causes and solutions for the problems it sees. Relying on the brain as an example of how to improve might be a mistaken premise, because maybe the brain isn't doing what the article thinks it's doing. So we're in agreement there, that the brain and LLMs are incomparable, but maybe the parts where they're comparable are more informative on the nature of hallucinations than the author may think.
reply
I think you can confidently say that brains do the following and LLMs don't:

* Continuously updates its state based on sensory data

* Retrieves/gathers information that correlates strongly with historic sensory input

* Is able to associate propositions with specific instances of historic sensory input

* Uses the above two points to verify/validate its belief in said propositions

Describing how memories "feel" may confuse the matter, I agree. But I don't think we should be quick to dismiss the argument.

reply
But the thing is that humans don't hallucinate as much as LLMs, so it's the differences not similarities (such as they are) that are important to understand why that is.

It's pretty obvious that an LLM not knowing what it does or does not know is a major part of it hallucinating, while humans do generally know the limits of their own knowledge.

reply
> An LLM was only every meant to be a linguistics model, not a brain or cognitive architecture.

See https://gwern.net/doc/cs/algorithm/information/compression/1... from 1999.

Answering questions in the Turing test (What are roses?) seems to require the same type of real-world knowledge that people use in predicting characters in a stream of natural language text (Roses are ___?), or equivalently, estimating L(x) [the probability of x when written by a human] for compression.

reply
I'm not sure what your point is?

Perhaps in 1999 it seemed reasonable to think that passing the Turing Test, or maximally compressing/predicting human text makes for a good AI/AGI test, but I'd say we now know better, and more to the point that does not appear to have been the motivation for designing the Transformer, or the other language models that preceded it.

The recent history leading to the Transformer was the development of first RNN then LSTM-based language models, then the addition of attention, with the primary practical application being for machine translation (but more generally for any sequence-to-sequence mapping task). The motivation for the Transformer was to build a more efficient and scalable language model by using parallel processing, not sequential (RNN/LSTM), to take advantage of GPU/TPU acceleration.

The conceptual design of what would become the Transformer came from Google employee Jakob Uzkoreit who has been interviewed about this - we don't need to guess the motivation. There were two key ideas, originating from the way linguists use syntax trees to represent the hierarchical/grammatical structure of a sentence.

1) Language is as much parallel as sequential, as can be seen by multiple independent branches of the syntax tree, which only join together at the next level up the tree

2) Language is hierarchical, as indicated by the multiple levels of a branching sytntax tree

Put together these two considerations suggests processing the entire sentence in parallel, taking advantage of GPU parallelism (not sequentially like an LSTM), and having multiple layers of such parallel processing to hierarchically process the sentence. This eventually lead to the stack of parallel-processing Transformer layers design, which did retain the successful idea of attention (thus the paper name "Attention is all you need [not RNNs/LSTMs]").

As far as the functional capability of this new architecture, the initial goal was just to be as good as the LSTM + attention language models it aimed to replace (but be more efficient to train & scale). The first realization of the "parallel + hierarchical" ideas by Uzkoreit was actually less capable than its predecesssors, but then another Google employee, Noam Shazeer, got involved and eventually (after a process of experimentation and ablation) arrived at the Transformer design which did perform well on the language modelling task.

Even at this stage, nobody was saying "if we scale this up it'll be AGI-like". It took multiple steps of scaling, from early Google's early Muppet-themed BERT (following their LSTM-based ELMo), to OpenAI's GPT-1, GPT-2 and GPT-3 for there to be a growing realization of how good a next-word predictor, with corresponding capabilities, this architecture was when scaled up. You can read the early GPT papers and see the growing level of realization - they were not expecting it to be this capable.

Note also that when Shazeer left Google, disappointed that they were not making better use of his Transformer baby, he did not go off and form an AGI company - he went and created Character.ai making fantasy-themed ChatBots (similar to Google having experimented with ChatBot use, then abandoning it, since without OpenAI's innovation of RLHF Transformer-based ChatBots were unpredictable and a corporate liability).

reply
[dead]
reply
I don’t know if this aligns with your thinking but there is a theory that memory is largely reconstructed every time it is remembered: https://en.wikipedia.org/wiki/Reconstructive_memory
reply
Thank you for that! I think I read about this a long time ago, internalized it, and forgot it. Pretty on the nose in this conversation... haha.
reply
A payload plus a transformation process is still storage, just compressed storage. There's conceptually little difference between storing data and the process that gets called to create that data.
reply
its not reliable, but we can also recall things accurately.

it does not store things in the way records of any sort do, but it does have a some store and recall mechanism that works.

To be fair, LLMs do this too - I just got ChatGPT to recite Ode to Autumn.

reply
Yes, I agree. I'm not against the idea that the brain can "store" things. Just whether our concept of how a "memory" "feels" is useful to us further understanding the brain's function.
reply
> I'm not convinced the brain stores memories

By what mechanism do you feel I "remember" last week?

reply
There is a theory known as "morphic resonance" which suggests that events and objects exist in the world as something analogous to radio waves. So when you "recall" a memory it is not because it is stored in the brain anywhere but rather you are picking up signals from the universe like a radio receiver.
reply
That theory sounds woo-woo. So something I remember that happened a few days ago, or my phone number exists somewhere out in the universe and our brain is able to access it?
reply
Reconstructing from what?
reply
deleted
reply
I believe this is why the importance of written (human) knowledge is only increasing, especially internally at companies. Written knowledge (i.e documentation) has always served as a knowledge cache and a way to transfer knowledge between people.

Without fundamental changes to the LLMs or the way we think about agentic systems, high quality, comprehensive written knowledge is the best path to helping agents "learn" over time.

reply
As a memory neuroscientist, I enjoyed the shoutout here to episodic memory. It strikes me, however, that a feature that I've noticed when observing "reasoning" models is that they may explicitly search for evidence for intermediate pieces of their responses. If we're following the "remember"/"know" distinction developed by Squire and others, perhaps the more apt analogy might be that a singly pass through an LLM is similar to a "I know this" result, prone to hallucination, conflation, etc., and the multipass reasoning model might be more akin to the "I remember this" result, where primary evidence serves as a substrate for the response?
reply
The terms are arbitrary and don't relate to human memory. Using the term "I know this" is a post-hoc retrofit to a process exclusively accessed wordlessly. Just as "I remember this" does not access that process, but rather comments on the aftermath pretty unscientifically, like a sportscaster trying to read a pitcher's mind.

“We refute (based on empirical evidence) claims that humans use linguistic representations to think.” Ev Fedorenko Language Lab MIT 2024

reply
> I’ll remind you that biologists do not, in the year 2025, know memory’s physical substrate in the brain! Plenty of hypotheses — no agreement. Is there any more central mystery in human biology, maybe even human existence?

Did they not recently transfer memory of how to solve a maze from one mouse to another, giving credibility to what can store information?

Searching, I only find the RNA transfers done in 60s, which ran into some problems. I thought a recent study did transfer proteins.

reply
That work has been pretty thoroughly debunked. What's getting transferred is stress hormones, and you can "duplicate" the results by using animals that have been stressed (but have never seen the maze) as the donors. They don't know anything about the maze, so can't transfer any knowledge. The recipients do better in the maze because they are more alert, due to the transfer of stress hormones.
reply
I would love to see some sources for this. Sounds super interesting. When do we get Matrix-style learning modules?
reply
You know, whatever memory is or where it’s at and however the mind works, I’m grateful I’ve got mine in tact right now and I appreciate science’s inability to zero in on these things.

It’s nice to know that this sort of appreciation is becoming more common. Somewhere between tech accelerationism and protestant resistance are those willing to re-interrogate human nature in anticipation of what lies ahead.

A different blog post from this month detailing an experience with ChatGPT that netted a similar reflection: https://zettelkasten.de/posts/the-scam-called-you-dont-have-...

reply
This is patently false.

"I’ll remind you that biologists do not, in the year 2025, know memory’s physical substrate in the brain! Plenty of hypotheses — no agreement. Is there any more central mystery in human biology, maybe even human existence?"

A hypothesis is very distinct from theoretical knowledge. A hypothesis lacks empirical evidence. A theory uses empirical information. That CS personnel are lacking both the scientific method and the ability to discern the current state of the art empirical research to disprove such wildly unsupported statements speaks to the field's total failure to develop present-day relevant tools. I would direct the author to two critical books

Evolution of Memory Systems

https://academic.oup.com/book/26033

How we remember: brain mechanisms of episodic memory

https://direct.mit.edu/books/monograph/2909/How-We-RememberB...

reply
If the author wrote the same sentence, but with "Plenty of theories -- no agreement", would you have an issue? I think the point works just as well.

"Theory" and "hypothesis" are pretty interchangeable in colloquial usage, and the post's tone is very colloquial. On top of that, there are things referred to as hypotheses that have empirical evidence, like the amyloid beta hypothesis. There is empirical evidence that supports it, but there is also evidence that doesn't, and it's an open question (depending on who you ask).

I don't think it shows that the author lacks the ability to discern state of the art research or is making wildly unsupported statements, I think they were using plain-English terms to describe a state where there's a lot of uncertainty about the physical mechanism as opposed to say, how a car engine works (which at a certain point relies on theories of physics, but they're theories that are almost universally agreed upon).

reply
Theory and hypothesis are not interchangeable. Hypothesis precedes theory. Always. Had he said "theories" he would be saying something completely different. It's another stage of scientific process altogether. In order: Hypothesis, empirical observation, theory, demonstration, new theoretical stage or field.

The bizarre CS idea of using "plain English" as a propagandic thrust to infantilize complex scientific theory is suspect. Let's eschew the idea. "Plain English" is Orwellian dumbing down.

There is a general agreement which links ontogeny and phylogeny, details working memory, its relationship to the allocortex, and subsequent episodic memory. So the neural correlates have been worked out.

The working memory and episodic memory papers in the last few years have isolated the correlates, we a have a fairly empirical neurobiological description of memory function and process.

What we're lacking now are the molecular mechanisms.

reply
I agree with the folks who call these screwups rather than hallucinations because the point of LLMs isn't to be right, it's to be statistically highly likely. If making something up fits that model, then that's what it will do. That's literally how it works.
reply
Sorry for harping on it, but I think this clearly reflects the difference between 2 approaches to storing knowledge, lossy but humongous, and lossless but limited.

LLMs - Lossy highly compressed knowledge which when prompted "hallucinates" facts. LLMs hallucinations are simply how the stored information is retrieved.

Memory (human in this case) - Extremely limited, but almost always correct.

Just an observation. No morals.

reply
LLMs behaviour is a lot closer to primary orality than literacy.
reply
this is such drivel and you don't realize it because the guy writes good ... and you don't realize it until you see the "are llms in hell" article ...

let's stop taking opinions on ai from randoms. please. they haven't a fkin clue.

reply
A model is capable of learning the "calibration" during the reinforcement learning phase, in this "old" post from OpenAI: https://openai.com/index/introducing-simpleqa/ you can see the positive correlation between stated confidence and accuracy.
reply