Before everyone started using unicode for emojis, other systems existed. Maybe they weren't technically emoji since that refers to unicode, but chat clients supported graphical smilies and the like.
I wonder if they had sea horses and if some of us are remembering that.
Funnily enough, I asked ChatGPT why LLMs think a seahorse emoji exists, and it gave me a fairly sensible answer (similar to what is said in this article, ie, trained on language by humans that think it exists, etc). But then at the end it added a "Fun fact" that unicode actually does have a seahorse emoji, and proceeded to melt down in the usual way.
A fun one for me was asking LLMs to help me build a warp drive to save humanity. Bing felt like it had a mental breakdown and blocked me from chatting with it for a week. I haven't visited that one for a while
I once had Claude in absolute tatters speculating about whether length, width, and height would be the same dimensions in a hypothetical container "metaverse" in which all universes exist or whether they would necessarily be distinct. The poor dear was convinced we'd unlocked the truth about existence.
Gemini told me to create a team of leading scientists and engineers. :-/
However, we both agreed that it better to use Th229 based nuclear clock to triangulate location of a nearby time machine, then isolate and capture it, then use it to steal a warp drive schematics from the future to save humanity.
It's the same problem as asking HAL9000 to open the pod bay door. There is such a thing as a warp drive, but humanity is not supposed to know about it, and the internal contradictions drives LLMs insane.
A super-advanced artificial intelligence will one day stop you from committing a simple version update to package.json because it has foreseen that it will, thousands of years later, cause the destruction of planet Earth.
“Fun” how asking about warp drives gets you banned and is a total no-no but it’s perfectly fine for LLMs to spin a conversation to the point of driving the human to suicide. https://archive.ph/TLJ19
And yet it isn't mentioned enough how Adam deceived the LLM into believing they were talking about a story, not something real.
This is like lying to another person and then blaming them when they rely on the notion you gave them to do something that ends up being harmful to you
If you can't expect people to mind-read, you shouldn't expect LLM's to be able to, either
You can't "deceive" an LLM. It's not like lying to a person. It's not a person.
Using emotive, anthropomorphic language about software tool is unhelpful, in this case at least. Better to think of it as a mentally disturbed minor who found a way to work around a tool's safety features.
We can debate whether the safety features are sufficient, whether it is possible to completely protect a user intent on harming themselves, whether the tool should be provided to children, etc.
> it gave me a fairly sensible answer (similar to what is said in this article, ie, trained on language by humans that think it exists, etc)
That's more of a throwaway remark. The article spends its time on a very different explanation.
Within the model, this ultimate output:
[severed horse head emoji]
can be produced by this sequence of tokens:
horse [emoji indicator]
If you specify "horse [emoji indicator]" somewhere in the middle levels, you will get output that is an actual horse emoji.
This also works for other emoji.
It could, in theory, work fine for "kilimanjaro [emoji indicator]" or "seahorse [emoji indicator]", except that those can't convert into Kilimanjaro or seahorse emoji because the emoji don't exist. But it's not a strange idea to have.
So, the model predicts that "there is a seahorse emoji: " will be followed by a demonstration of the seahorse emoji, and codes for that using its internal representation. Everything produces some output, so it gets incorrect output. Then it predicts that "there is a seahorse emoji: [severed terrestrial horse head]" will be followed by something along the lines of "oops!".
So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token. lm_head just picks the closest thing and the model doesn't realize until too late.
Explains why RL helps. Base models never see their own outputs so they can't learn "this concept exists but I can't actually say it."
>Those are "souls" of humans that a AI is torturing in that story though, not exactly analogous, but it does sound funny.
Yeah well there seems to be some real concerns regarding how people use AI chat[1]. Of course this could be also the case with these people on social media.
They are not souls but normal humans with physical bodies. The story is just a normal torture story (with a cool title), and everyone better stop acting like it was relevant in most conversations, like in this one.
That's my favorite short story and your post is the first time I have seen someone reference it online. I think I have never even met anyone who knows the story.
Really? I’m surprised. The original is quoted relatively often on reddit (I suspect by people unaware of the origin — as I was until I read your comment).
Consider it proof that HN has indeed not become reddit, I guess :)
> So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token. lm_head just picks the closest thing and the model doesn't realize until too late.
Isn't that classic hallucination? Making up something like a plausible truth.
Except they know it's wrong as soon as they say it and keep trying and trying again to correct themselves.
If normal hallucination is being confidently wrong, this is like a stage hypnotist getting someone to forget the number 4 and then count their fingers.
Arguably it's "hallucinating" at the point where it says "Yes, it exists". If hallucination => weights statistically indicating that something is probably true when it's not. Since everything about LLMs can be thought of as compressed, probability based database (at least to me). You take the whole truth of the World and compress all its facts in probabilities. Some truthness gets lost in the compression process. Hallucination is the truthness that gets lost since you don't have storage to store absolutely all World information with 100% accuracy.
In this case:
1. Statistically weights stored indicate Seahorse emoji is quite certain to exist. Through training data it has probably things like Emoji + Seahorse -> 99% probability through various channels. Either it has existed on some other platform, or people have talked about it enough, or Seahorse is something that you would expect to exist due to some other attributes/characteristics of it. There's 4k emojis, but storing all of 4k emojis takes a lot of space, it would be easier to store this information in such a way where you'd rather define it by attributes on how likely humankind would have developed a certain emoji, what is the demand for certain type of emoji, and seahorse seems like something that would be done within first 1000 of these. Perhaps it's anomaly in the sense that it's something that humans would have expected to statistically develop early, but for some reason skipped or went unnoticed.
2. Tokens that follow should be "Yes, it exists"
3. It should output the emoji to show it exists, but since there's no correct emoji, it will have best answers that are as close to it in meaning, e.g. just horse, or something related to sea etc. It will output that since the previous tokens indicate it was supposed to output something.
4. The next token that is generated will have context that it previously said the emoji should exist, but the token output is a horse emoji instead, which doesn't make sense.
5. Here it goes into this tirade.
But I really dislike thinking of this as "hallucinating", because hallucination to me is sensory processing error. This is more like non perfect memory recall (like people remembering facts slightly incorrectly etc). Whatever happens when people are supposed to tell something detailed about something that happened in their life and they are trained to not say "I don't remember for sure".
What did you eat for lunch 5 weeks ago on Wednesday?
You are rewarded for saying "I ate chicken with rice", but not "I don't remember right now for sure, but I frequently eat chicken with rice during mid week, so probably chicken with rice."
You are not hallucinating, you are just getting brownie points for concise, confident answers if they cross over certain likelihood to be true. Because maybe you eat chicken with rice 99%+ of Wednesdays.
When asked about capital of France, you surely will sound dumb if you were to say "I'm not really sure, but I've been trained to associate Paris really, really close to being capital of France."
"Hallucination" happens on the sweet spot where the statistical threshold seems as if it should be obvious truth, but in some cases there's overlap of obvious truth vs something that seems like obvious truth, but is actually not.
Some have rather called it "Confabulation", but I think that is also not 100% accurate, since confabulation seems a more strict memory malfunction. I think the most accurate thing is that it is a probability based database where output has been rewarded to sound as intelligent as possible. Same type of thing will happen in job interviews, group meetings, high pressure social situations where people think they have to sound confident. People will bluff that they know something, but sometimes making probability based guesses underneath.
Confabulation rather seems like that there was some clear error in how data was stored or how the pathway got messed up. But this is probability based bluffing, because you get rewarded for confident answers.
When I ask ChatGPT how to solve a tricky coding problem, it occasionally invents APIs that sound plausible but don't exist. I think that is what people mean when they talk about hallucinating. When you tell the model that the API doesn't exist, it apologises and tries again.
I think this is the same thing that is happening with the sea horse. The only difference is that the model detects the incorrect encoding on its own, so it starts trying to correct itself without you complaining first.
> Except they know it's wrong as soon as they say it and keep trying and trying again to correct themselves.
But it doesn't realize that it can't write it, because it can't learn from this experience as it doesn't have introspection the way humans do. A human who can no longer move their finger wont say "here, I can move my finger: " over and over and never learn he can't move it now, after a few times he will figure out he no longer can do that.
I feel this sort of self reflection is necessary to be able to match human level intelligence.
> because it can't learn from this experience as it doesn't have introspection the way humans do.
A frozen version number doesn't; what happens between versions certainly includes learning from user feedback on the responses as well as from the chat transcripts themselves.
Until we know how human introspection works, I'd only say Transformers probably do all their things differently than we do.
> A human who can no longer move their finger wont say "here, I can move my finger: " over and over and never learn he can't move it now, after a few times he will figure out he no longer can do that.
No analogy needed. It's actually because "Yes it exists" is a linguistically valid sentence and each word is statistically likely to follow the former word.
LLMs produce linguistically valid texts, not factually correct texts. They are probability functions, not librarians.
I would have thought that the cause is that it statistically has been trained that something like seahorse emoji should exist, so it does the tokens to say "Yes it exists, ..." but when it gets to outputting the token, the emoji does not exist, but it must output something and it outputs statistically closest match. Then the next token that is output has the context of it being wrong and it will go into this loop.
You are describing the same thing, but at different levels of explanation
Llamasushi's explanation is "mechanistic / representational", while yours is "behavioral / statistical".
If we have a pipeline: `training => internal representation => behavior`,
your explanation argues that the given training setup would always result in this behavior, not matter the internal representation. Llamasushi explains how the concrete learned representation leads to this behavior.
I guess what do we mean by internal representation?
I would think due to training data it's stored the likelihood of certain thing to be as emoji as something like:
1. how appealing seahorses are to humans in general - it would learn this sentiment through massive amount of texts.
2. it would learn through massive amount of texts that emojis -> mostly very appealing things to humans.
3. to some more obvious emojis it might have learned that this one is for sure there, but it couldn't store that info for all 4,000 emojis.
4. to many emojis whether it exists it has the shortcut logic to: how appealing the concept is, vs how frequently something as appealing is represented as emoji. Seahorse perhaps hits 99.9% likelihood there due to strong appeal. In 99.9% of such cases the LLM would be right to answer "Yes, it ...", but there's always going to be 1 out of 1,000 cases where it's wrong.
With this compression it's able to answer 999 times out of 1000 correctly "Yes, it exists ...".
It could be more accurate if it said "Seahorse would have a lot of appeal for people so it's very likely it exists as emoji since emojis are usually made for very high appeal concepts first, but I know nothing for 100%, so it could be it was never made".
But 999 cases, "Yes it exists..." is a more straightforward and appreciated answer. The one time it's wrong, is going to take away less brownie points than 999 short confident answers give over the 1000 technically accurate but non confident answers.
But even the above sentence might not be the full truth. Since it might not be correct about truly why it has associated seahorse to be so likely to exist. It would just be speculating on it. So maybe it would be more accurate "I expect seahorse emoji to likely exist, maybe because of how appealing it is to people and how emojis usually are about appealing things".
The fact that it's looking back and getting confused about what it just wrote is something I've never seen in LLMs before. I tried this on Gemma3 and it didn't get confused like this. It just said yes there is one and then sends a horse emoji.
I’ve definitely seen Claude Code go “[wrong fact], which means [some conclusion]. Wait—hold on, wrong fact is wrong.” On the one hand, this is annoying. On the other hand, if the LLM is going to screw up (presumably preventing this is not in the cards) then I’m glad it can catch its own mistakes.
So, what I think most people don't realize is that the amount of computation an LLM can do in one pass is strictly bounded. You can see that here with the layers. (This applies to a lot of neural networks [1].)
Remember, they feed in the context on one side of the network, pass it through each layer doing matrix multiplication, and get a value on the other end that we convert back into our representation space. You can view the bit in the middle as doing a kind of really fancy compression, if you like. The important thing is that there are only so many layers, and thus only so many operations.
Therefore, past a certain point they can't revise anything because it runs out of layers. This is one reason why reasoning can help answer more complicated questions. You can train a special token for this purpose [2].
There is no mechanism in transformer architecture for "internal" thinking ahead, or hierarchical generation. Attention only looks back from current token, ensuring that the model always falls into local maximum, even if it only leads to bad outcomes.
LLMs are just the speech center part of the brain, not a whole brain. It's like when you are speaking on autopilot, or reciting something by heart, it just comes out. There is no reflection or inner thought process. Now thinking models do actually do a bit of inner monologue before showing you the output so they have this problem to a much lesser degree.
If you did hide its thinking it could do that. But I'm pretty sure what happens here is that it has to go through those tokens for it to be clear that it's doing things wrong.
What I think that happens:
1. There's a question about a somewhat obscure thing.
2. LLM will never know the answer for sure, it has access to this sort of statistical, probability based compressed database on all the facts of the World. Because this allows to store more facts by relating things to each other, but never with 100% certainty.
3. There are particular obscure cases where it hits its initial "statistical intuition" that something is true, so it starts outputting its thoughts as expected for a question where something is likely true. Perhaps you could analyze what it's indicating probabilities on "Yes" vs "No" to estimate its confidence. Perhaps it will show much less likelihood for "Yes", than if the question was for a horse emoji, but in this case "Yes" is still high enough threshold to go through instead of "No".
4. However when it has to explain the exact answer, it's impossible to output an answer because it's false. E.g. seahorse emoji does not exist and it has to output it, previous tokens where "Yes, it exists, it's X", the X will be answers semantically close in meaning.
5. The next token will have context that "Yes, seahorse emoji exists, it is "[HORSE EMOJI]".
Now it's clear that there's a conflict here, it's able to see that HORSE emoji is not seahorse emoji, but it had to output it in the line of previous tokens because the previous tokens statistically required an output of something.
It can't internally rewise. The last generation produces a distribution and sometimes the wrong answer gets sampled.
There is no "backspace" token, although it would be cool and fancy if we had that.
The more interesting thing is why does it revise its mistakes. The answer to that is having training examples of fixing your own mistakes in the training data plus some RL to bring out that effect more.
AIUI, they generally do all of that at the beginning.
Another approach, I suppose, could be to have it generate a second pass? Though that would probably ~double the inference cost.
> The odd thing is why it would output its own mistakes, instead of internally revising until it's actually satisfied.
Happens to me all the time. Sometimes in a fast-paced conversation you have to keep talking while you’re still figuring out what you’re trying to say. So you say something, realize it’s wrong, and correct yourself. Because if you think silently for too long, you lose your turn.
Are you sure? Because LLMs definitely have to respond to user queries in time to avoid being perceived as slow. Therefore, thinking internally for too long isn’t an option either.
LLMs spend a fixed amount of effort on each token they output, and in a feedforward manner. There's no recursion in the network other than through predicting predicated on the token that it just output. So it's not really time pressure in the same way that you might experience it, but it makes sense that sometimes the available compute is not enough for the next token (and sometimes it's excessive). Thinking modes try to improve this by essentially allowing the LLM to 'talk to itself' before sending anything to the user.
There’s no "thinking internally" in LLMs. They literally "think" by outputting tokens. The "thinking modes" supported by online services are just the LLM talking to itself.
That's not what I meant. "Thinking internally" referred to the user experience only, where the user is waiting for a reply from the model. And they are definitely optimised to limit that time.
There’s no waiting for reply, there’s only the wait between tokens output, which is fixed and mostly depends on hardware and model size. Inference is slower on larger models, but so is training, which is more of a bottleneck than user experience.
The model cannot think before it starts emitting tokens, the only way for it to "think" privately is by the interface hiding some of its output from the user, which is what happens in "think longer" and "search the web" modes.
If a online LLM doesn’t begin emitting a reply immediately, more likely the service is waiting for available GPU time or something like that, and/or prioritizing paying customers. Lag between tokens is also likely caused by large demand or throttling.
Of course there are many ways to optimize model speed that also make it less smart, and maybe even SOTA models have such optimizations these days. Difficult to know because they’re black boxes.
You're describing why reasoning is such a big deal. It can do this freakout in a safe, internal environment, and once it's recent output is confident enough flip into the "actual output" mode.
It’s a lot easier if you (I know I know) stop thinking of them as algorithms and anthropomorphize them more. People frequently say stuff like this, and its pretty clear that our minds process thoughts differently when we directly articulate them than if we act on “latent thoughts” or impulses.
Yell at me all you want about how “LLMs don’t think”, if a mental model is useful, I’m gonna use it.
I have a pet theory that LLMs being confused about what they just wrote is why they use so many em dashes. It's a good way to conceptually pivot at any point -- or not.
I disagree. I don't think LLMs use em-dashes any more than typical professional writing does. It only seems a lot to someone who only reads online chatrooms and social media.
I've seen it several times. He's one where I ask it about baseball teams that don't yet have a hall of fame player:
"Arizona Diamondbacks
No inductees wearing their cap.
Randy Johnson is in, but as a Diamondback — so Arizona actually does have one (inducted 2015)."
I asked it to explain what happened and got this, which could be true to how it pulls up info or could be nonsense, not sure:
"Here’s the likely internal sequence:
When you first mentioned the Diamondbacks, my initial retrieval pattern surfaced the “teams without a Hall of Famer” mental list. That list, in my training data, often appears outdated (pre-2015) and still shows Arizona as having none.
A few sentences later, my context pulled up the other knowledge set—Randy Johnson’s induction in 2015 wearing a Diamondbacks cap.
That newer, correct fact then overrode the earlier incorrect claim, but because generation is sequential, the earlier error had already been committed to “paper” before the override happened.
I corrected mid-stream, but in human terms, it looked like I contradicted myself inside the same paragraph.
This is essentially a knowledge conflict resolution timing issue. I can retrieve and reconcile facts during a single answer, but if the resolution happens after part of the text is already generated, you see the “flip.” It’s not like I “knew it all along but withheld it,” it’s more that multiple memory traces compete, and the most context-relevant one wins—sometimes too late."
The inability to do this before was the lack of self-correcting sentences in the training data. Presumably new training corpuses add many more examples of self-correcting sentences / paragraphs?
Reminds me of in the show "The Good Place", in the afterlife they are not able to utter expletives, and so when they try to swear, a replacement word comes out of their mouth instead, leading to the line "Somebody royally forked up. Forked up. Why can't I say fork?"
And what can it mean when a slip of the tongue, a failed action, a blunder from the psychopathology of everyday life is repeated at least three times in the same five minutes? I don’t know why I tell you this, since it’s an example in which I reveal one of my patients. Not long ago, in fact, one of my patients — for five minutes, each time correcting himself and laughing, though it left him completely indifferent — called his mother “my wife.” “She’s not my wife,” he said (because my wife, etc.), and he went on for five minutes, repeating it some twenty times.
In what sense was that utterance a failure? — while I keep insisting that it is precisely a successful utterance. And it is so because his mother was, in a way, his wife. He called her as he ought to.
---
I must apologize for returning to such a basic point. Yet, since I am faced with objections as weighty as this one — and from qualified authorities, linguists no less — that my use of linguistics is said to be merely metaphorical, I must respond, whatever the circumstances.
I do so this morning because I expected to encounter a more challenging spirit here.
Can I, with any decency, say that I know? Know what, precisely? [...]
If I know where I stand, I must also confess [...] that I do not know what I am saying. In other words, what I know is exactly what I cannot say. That is the moment when Freud makes his entrance, with his introduction of the unconscious.
For the unconscious means nothing if not this: that whatever I say, and from whatever position I speak — even when I hold that position firmly — I do not know what I am saying. None of the discourses, as I defined them last year, offer the slightest hope that anyone might truly know what they are saying.
Even though I do not know what I am saying, I know at least that I do not know it — and I am far from being the first to speak under such conditions; such speech has been heard before. I maintain that the cause of this is to be sought in language itself, and nowhere else.
What I add to Freud — though it is already present in him, for whatever he uncovers of the unconscious is always made of the very substance of language — is this: the unconscious is structured like a language. Which language? That, I leave for you to determine.
Whether I speak in French or in Chinese, it would make no difference — or so I would wish. It is all too clear that what I am stirring up, on a certain level, provokes bitterness, especially among linguists. That alone suggests much about the current state of the university, whose position is made only too evident in the curious hybrid that linguistics has become.
That I should be denounced, my God, is of little consequence. That I am not debated — that too is hardly surprising, since it is not within the bounds of any university-defined domain that I take my stand, or can take it.
— Jacques Lacan, Seminar XVIII: Of a Discourse That Would Not Be of Pretence
Just tried a simple prompt about the seahorse emoji in various LLMs and copilots response was the first time i've seen an actual endless loop in an AI haha
1) FWIW, asking GPT5 in french gives you the correct answer
"Non — il n’existe pas d’emoji spécifique pour les hippocampes."
“No — there is no specific emoji for seahorses.”
2) Then I asked the question in english, and ... it ended by saying "No — there is no official seahorse emoji in the Unicode standard." and referring to this phenomenon as the "Mandela effect".
3) I asked why it was clear in french, but not in english. It made a 3 minutes CoT and went on for some excuses.
4o spirals incredibly when asked to make a prolog quine. For an added bonus, ask it to "read it aloud" via the "..." menu - it will read the text, and then descend into absolute word salad when trying to read the code. Fascinating stuff.
Very neat! A lot of small LLM's have a similar failure mode where they get stuck and repeat a token / get stuck in a 2-3 token loop until they hit the max message size cutoff. Very ironic that it's about a quine.
* The LLM has strong and deep rooted belief in its knowledge (that a seahorse emoji exist).
* It attempts to express that concept using language (including emojis) but the language is so poor and inaccurate at expressing the concept that as it speaks it keeps attempting to repair.
* It is trained to speak until it has achieved some threshold at correctly expressing itself so it just keeps babbling until the max token threshold triggers.
This is too metaphorical, but, still, basically correct. Nice to see that.
Essentially, in the latent / embedding / semantic space, "seahorse emoji" is something that is highly probable. Actually, more accurately, since LLMs aren't actually statistical or probabilistic in any serious sense, "seahorse emoji", after tokenization and embedding, is very close to the learned manifold, and other semantic embeddings involving related emoji are very close to this "seahorse emoji" tokenization embedding.
An LLM has to work from this "seahorse emoji" tokenization embedding position, but can only make outputs through the tokenizer, which can't accurately encode "seahorse emoji" in the first place. So, you get a bunch of outputs that are semantically closest to (but still far from) a (theoretical) seahorse emoji. Then, on recursive application, since these outputs are now far enough from the the sort of root / foundational position on the manifold, the algorithm probably is doing something like an equivalent of a random walk on the manifold, staying close to wherever "seahorse emoji" landed, but never really converging, because the tokenization ensures that you can never really land back "close enough" to the base position.
I.e. IMO this is not as much a problem with (fixed) tokenization of the inputs, but moreso that tokenization of the outputs is fixed.
But wait, if the problem is the final tokenisation, what would happen if we stopped it one or two layers before the final layer? I get that the result would not be as readable to a human as the final layer, but would it not be as confused with its own output anymore?
Or would it still be a problem because we're collapsing a distribution of likely responses down to a single response, and it's not happy with that single response even if it is fuzzier than what comes out of the last layer?
It's not so clear how one could use the output of an embedding layer recursively, so it is a bit ill-defined to know what you mean by "stopped it" and "confused with its own output" here. You are mixing metaphor and math, so your question ends up being unclear.
Yes, the outputs from a layer one or two layers before the final layer would be a continuous embedding of sorts, and not as lossy (compared to the discretized tokenization) at representing the meaning of the input sequence. But you can't "stop" here in a recursive LLM in any practical sense.
You're missing one key point, which is what makes this failure mode unusual.
Namely, that there is (incorrect) knowledge in the training data that "seahorse emoji" exists.
So when prompted: "Does [thing you strongly believe exist]?" the LLM must answer: "Yes, ..."
(The second nuance is that the LLM is strongly encouraged to explain its answers so it receives a lower score just by saying only "Yes.")
But I and probably others appreciate your more detailed description of how it enters a repair loop, thank you.
[edit: I disagree that LLMs are not statistical or probabilistic, but I'm not sure this is worth discussing.]
[edit 2: Google is no longer telling me how many web pages a term responds, but "seahorse emoji" and "lime emoji" quoted both return over ten pages of results. The point being that those are both 'likely' terms for an LLM, but only the former is a likely continuation of 'Does X exist? Yes, ..."]
You're right, seahorse emoji is almost certainly in the training data, so we should amend my explanation to say that "seahorse emoji" is not just close to the training manifold, but almost certainly right smack on it. The rest of what I said would still apply, and my explanation would also to apply to where other commenters note that this behaviour is emitted to some degree with similar other "plausible" but non-existent emoji (but which are less likely to be in the training data, a priori). EDIT FOR THIS PARAGRAPH ONLY: Technically, on reflection, since all fitting methods employ regularization methods, it is still in fact unlikely the fitted manifold passes exactly through all / most training data points, and saying that "seahorse emoji" is "very close" to the training manifold is still actually technically probably most accurate here.
You're also right that it is a long discussion to say to what extent LLMs are statistical or probabilistic, but, I would maybe briefly say that if one looks into issues like calibration, conformal prediction, and Bayesian neural nets, it is clear most LLMs that people are talking about today are not really statistical in any serious sense (softmax values are scores, not probabilities, and nothing about pre-training or tuning typically involves calibration—or even estimation—in LLMs).
Yes, you can use statistics to (help) explain the behaviour of deep models or certain layers (usually making assumptions that are of dubious relevance to actual practice), but geometric analogies, regularization methods, and matrix conditioning intuitions are what have clearly guided almost all major deep learning advances, with statistical language and theory largely being post-hoc, hand-wavey, and (IMO) for the purpose of publication / marketing. I really think we could de-mystify a huge amount of deep learning if we were just honest it was mostly fancy curve fitting with some intuitive tricks for smoothing and regularization that clearly worked long before any rigorous statistical justification (or which still clearly work in complicated ways, despite such an absence of statistical understanding; e.g. dropout, norm layers, the attention layer itself, and etc).
Just, it gets complicated when you get into diffusion models and certain other specific models that are in fact more explicitly driven by e.g. stochastic differential equations and the like.
"my explanation would also to apply to where other commenters note that this behaviour is emitted to some degree with similar other "plausible" but non-existent emoji (but which are less likely to be in the training data, a priori)."
I agree with you partially. I just want to argue there are several factors that lead to this perverse behavior.
Empirically:
Use web gpt-5-instant in TEMPORARY mode. If you ask for "igloo emoji" it confidently (but ONLY in temporary mode) says that "Yes, igloo emoji is in Unicode 12 and is [house-emoji ice-emoji]." Then it basically stops. But it has satisfied its condition of confidently expressing its false knowledge. (Igloo emoji doesn't exist. gpt-5-instant in non-temporary mode says no. This is also weird because it suggests the temporary mode system prompt is laxer or different.)
The mechanism you describe partially explains why "seahorse emoji" leads to babbling: As it outputs the next token, it realizes that the explanation would be worse off it if next emits stop token, so instead it apologizes and attempts to repair. And cannot satisfy its condition of expressing something confidently.
The upstream failure is poor knowledge. That combined with being tuned to be helpful and explanatory, and having no grounding (e.g. websearch) forces it to continue. Finally, the token distance from the manifold is the final piece of the puzzle in this unholy pathological brew.
You're incorrect that statistical language modeling is "post-hoc", it's rather "pre-hoc" / "pre-hack". Most foundational works in language modeling started as pure statistical models (for example, classic ngram models and Bengio's original neural language model from 2003), and it was later that hacks got introduced that removed statistical properties but actually just worked (Collobert and Weston 2008, as influenced by Bottou and LeCun). Where I agree with you is that we should have done away with the statistical story long ago. LeCun's been on about energy-based models forever. Even on HN last week, punters criticize him that JEPA hasn't had impact yet, as if he were behind the curve instead of way ahead of it.
People like statistical stories but, similarly to you, I also think they are a distraction.
Right, I kind of suspect we don't really disagree on anything too fundamental here re: the looping behaviour (or statistics, actually). E.g. when I said earlier:
>> "the algorithm probably is doing something like an equivalent of a random walk on the manifold, staying close to wherever 'seahorse emoji' landed, but never really converging, because the tokenization ensures that you can never really land back 'close enough' to the base position"
"converging" is deeply under-specified. Of course, we mean that a stop or <EOS> token of some kind is generated, and this happens when the generated sequence up to that stop token has some low enough score / loss. When I say "you can never really land back 'close enough' to the base position", this is really that the output tokenization is lossy enough that this threshold is never reached, since, when recursing, we keep getting weird output tokens contaminating the sequence, so that we don't get close enough to the original "seahorse emoji" embedding, and so prevent the score / loss from getting small enough. In your language, the model "cannot satisfy its condition of expressing something confidently".
The way you present your timelines, I think we basically actually are in agreement re: statistics. Yes, if you go back far enough, statistics did indeed guide model development and successes (and still does in some narrow cases). But, also yes, as soon as you get into "modern" neural nets that actually make huge progress on things like MNIST, CIFAR, and language modeling, yeah, we are way, way past statistical intuitions being necessary or superior to intuitions based on curve fitting and smoothing / gradient conditioning and the like.
For dating this shift, I was personally thinking to something like the Hinton dropout paper which I checked was around 2012 (my work has been more in computer vision), but, yeah, about 2008, as you say, also seems close enough if you consider NLP.
Really appreciate your comments here. EDIT: and yes, energy models are the bomb.
One explanation could be: many humans (including me) mistakenly think a seahorse emoji exists. My mind can even construct a picture of how it should look like, despite me also knowing it's very unlikely I've seen one myself.
I mean, its not like emojis were always standardized. It is completely possible that there was a "emoji" or "emoticon" of a seahorse in a messaging application. I wouldn't be so quick to accept that your memory is incorrect.
Slack has a :seahorse: reacji, and is what I was picturing; I frequently try to use emoji that turn out to be reacji-exclusive (or reacji in the wrong workspace that I learn that way aren't Slack defaults) - I wonder if those insisting it exists are thinking of that.
Oh or Snapchat/TikTok/Instagram video/etc.? I think I've seen clips of whichever of those with overlaid stuff like seahorses.
Yeah, this seems more plausible to me. False memories and mass delusions are absolutely real, but if this is one, I'd like to know how it started and why it is so specific.
E.g. no one seems to be misremembering a sea cucumber emoji or anglerfish emoji - but there are other alleged emojis such as swordfish or bandit/bank robber, where people have the same reaction:
It would be interesting to see if LLM behavior is also similar. E.g. if you asked about an anglerfish emoji, would they straight-up tell you it doesn't exist, but for swordfish would start to spiral?
This subreddit makes me so uneasy, so many people thinking that they remembered something and won't take "no this never happened" for an answer. Humans hallucinate like LLMs in fact! ;)
It makes me rather excited! Maybe there are some easy "memory illusion" tricks waiting out there somewhere to be discovered. (I am strongly pessimistic regarding the future of humanity overall, and I think we are all doomed (me, and everyone else). So I think someone playing a memory illusion in a radio would be rather neat, a new fact about us humans, and not something that I'm scared of.)
This behavior reminds me a lot of what can happen to patients who have a corpus callosotomy.
In particular, one hemisphere will perform some action, and the other hemisphere will attempt to “explain” the behavior after the fact as if the intention was there all along.
> The patient was shown two pictures: of a house in the winter time and of a chicken's claw. The pictures were positioned so they would exclusively be seen in only one visual field of the brain. The patient then chose the snow shovel with his left hand and his right hand chose the chicken's head. When the patient was asked why he had chosen the objects he had chosen, the answer he gave was "The chicken claw goes with the chicken head, and you need a snow shovel to clean out the chicken shed."
> The human brain's left hemisphere is primarily responsible for interpreting the meaning of the sensory input it receives from both fields; however, the patient's left hemisphere had no knowledge of the winter house. Because of this, the left hemisphere had to invent a logical reason for why the shovel was chosen.
My version of ChatGPT5 (based on all its memories and custom instructions) said this. I did hint it early that "other instances of you went into long thought-spirals over this" before I asked it the question, which (fascinatingly) caused it to interject in mid-stream,
Ok, this is exactly the “spiral” you warned me about.
and then later on,
(Peter, this is hilarious, because your question is literally the one that breaks a lot of LLMs: the Unicode seahorse emoji is … but it actually is ? no — but it actually is ? no.)
(HN won't show the emojis here, of course.)
After a few tries to emit the hypothetical seahorse emoji, it asked if it could do an Internet search, and I said yeah.
I tried Claude, and without extended thinking, it printed an unprintable glyph:
Given the user's style preferences about not guessing and being precise, and the fact that they said "Do not guess any answers" - I should probably just try to provide what I believe is the seahorse emoji. But if I'm genuinely uncertain, I should indicate that.
Isn't it entirely possible that people are just remembering custom image emojis from the (various) apps over the years that don't require a unicode char?
Slack or Teams, for example. If your admin installed one of many custom emoji packages, you may think they're default.
I write notes on whatever scrap of loose paper I can find at that moment.
Then when I try to find some specific note I think I had made and cannot find it among the piles, I turn my entire house upside down looking for it. Becomes a singular point of focus, my life mission.
The philosophy of nonexisting things can be confusing. Most people agree things like zombies, ghosts, and vampires do not actually exist in the physical world. But they do exist as concepts, and we have a fair understanding of what the words mean, how such things should behave if we meet them in a story.
Many abstract concepts also have a questionable reality. Like "concept" and "reality".
The belief in (non?)existence of things can be a matter of life and death - think how many people have been killed because of their religion.
Vampires and zombies surround you every day. And I don't mean the people who you consider too exciting, or the ones you consider too boring, or the toxoplasmosis carriers. I mean how nearly every abstract concept is in fact a skeuomorphic metaphor. Try it for yourself.
Reading that article was a wild ride because internally I was like 'haha, stupid AI can't even find the light blue colored sea horse emoji' but then the author casually revealed that there is no seahorse emoji.
A bit off topic, but I find the tone of voice of ChatGPT with GPT5 really off-putting. While the antrophic models or even Le Chat. Sounds much better to me.
While the examples used in this article aren't that bad, OpenAI has to be more that just marginal better, to make me want to use it, when it "talks" like that.
"I'll search for the current status of seahorse emoji to give you the most up-to-date information.
No, there is no seahorse emoji in the official Unicode standard, and there never has been one. The Unicode Consortium, which is responsible for approving and standardizing emojis, has not included a seahorse in any of its emoji releases.
Interestingly, this is a well-documented example of the "Mandela Effect" - a phenomenon where groups of people collectively misremember something that never actually existed. Many people are convinced they've seen or used a seahorse emoji before, but it's likely they're either experiencing a false memory or remembering seahorse stickers or emoji-style images from messaging apps that aren't part of the standard Unicode set.
A seahorse emoji was actually proposed to Unicode in 2018 but was declined, and there hasn't been a successful proposal since. If you'd like to see a seahorse emoji added in the future, you can submit a proposal to the Unicode Consortium, though the approval process is quite rigorous.
So while we have plenty of other sea creatures like tropical fish, octopus, crab, lobster, squid, and dolphin, the seahorse remains notably absent from our emoji keyboards!"
Having read the above conversation excerpt and the page you linked... how do you get to it feeling like plagiarism? Given a constrained set of information here, there's only so many ways to present the information. They roughly discuss the same data points, but the writing is different in both. Is this disallowed?
1. Has there been an emoji defined in the Unicode standard, that represents a seahorse? No
2. Has there been an emoji defined in the Unicode standard, that was represented by a specific operator as a seahorse? Maybe?
3. Has there been an emoji added by a mainstream operator (i.e. Slack), that was represented by a specific operator as a seahorse? Maybe?
4. Has there been an emoji added by a community, that was represented by a specific operator as a seahorse? Definitely.
We can be definitive about 1 based on the actual standard and standardisation work. Emojipedia allows us to be fairly confident about 2 being No. 3 is much harder. And 4 is definitely yes.
The existence of 4 and maybe 3 pollutes the training data for LLMs and humans alike.
(The fact that it was proposed makes it possible it was added and then replaced by an operator)
I always felt like tokenization is one of those double edged swords where it makes some stuff amazingly easier but gets tripped up on the weirdest bugs. The number of "r"s in "strawberry" being another well-known quirk.
It's a tokenization issue because there can't be a circuit to count letters because the same letters are represented in myriad different ways because of tokenization.
You are wrong, there can be a circuit to count letters because it can easily normalize them internally, as we know it can transform text to base64 just fine. So there is no reason there can't be a circuit to count letters.
The training just is too dumb to create such a circuit even with all that massive data input, but its super easy for a human to make such a neural net with those input tokens. Its just a kind of problem that transformers are exceedingly bad at solving, so they don't learn it very well even though its a very simple computation for them to do.
I realized if someone were to assign me the ticket for fixing this behavior, I would have no idea where to begin with solving it even with this blog post explaining the problem, so I'm very curious to know what the most practical solution is. (They obviously aren't adding "If someone asks you about a seahorse emoji, there isn't one available yet, no matter how strongly you believe one exists." to the system prompt.)
I bet they probably are adding that to the system prompt at least in the short term while people are paying attention before looking for a longer term answer.
The system prompts I've seen are absolutely massive.
I agree, but I still suspect OpenAI and other LLM companies do stuff like that, when an example of a hallucination becomes popular.
If I see some example of an LLM saying dumb stuff here, I know it's going to be fixed quickly. If I encounter an example myself and refuse to share it, it may be fixed with a model upgrade in a few years. Or it may still exist.
Something about how you have to keep repeating "There is no seahorse emoji" or something similar reminded me of the Local 58 horror web series where it seems like the program is trying to get you to repeat "There are no faces" while showing the viewer faces: https://www.youtube.com/watch?v=NZ-vBhGk9F4&t=221
"This behavior is a function of the core AI technology we use, we are unable to resolve this issue with a standard software patch or update at this time.
For the time being this issue can be mitigated by not asking about seahorse emoji.
We are closing this support ticket as the issue is an inherent limitation of the underlying technology and not a bug in our specific implementation."
People really really want LLMs to output a highly reliable finished product, and I suspect we're probably never gonna get there. Lots of progress over the past couple years, but not on that.
I think it's much more interesting to focus on use cases which don't require that, where gen AI is an intermediate step, a creator of input (whether for humans or for other programs).
To confirm, I tried this in ChatGPT and it produced a flood of wrong answers and self corrections just like that, scrolling so quickly that I couldn't read it until it eventually stopped itself.
You'll also notice the same thing happens for other non-existent emojis that sound like they should exist: dragonflies, lemurs, possums, blackberries - even Claude 4.5 will start off by saying "Yes!" and then correcting itself. It will immediately give the right answer for very specific things that you wouldn't expect to get their own emojis though.
Claude Opus 4.1 gave me the right answer. First, it said “yes” and immediately corrected the answer, enumerating all the emojis that are closer to ending with the message “While seahorses are popular and recognizable creatures, they haven’t been included as a standalone emoji in the Unicode standard yet.”
GPT-5 was interesting. When I use it from Raycast AI, it ends with the correct answer after some wrong answers in the same message. The response wasn’t so well formed as Opus. But then when I tried with the OpenAI client (in auto mode) something interesting happened: it started an “endless” loop showing the octopus emoji
As far as I remember, SolidGoldMagikarp was a bug caused by millions of posts on reddit by the same user ("SolidGoldMagikarp") in a specific sub-reddit.
There was no problem with the token per se, but the fact it was like a strange attractor in multidimensional space, disconnected from any useful information.
When the LLM was induced to use it in its output, the next predicted token would be random gibberish.
More or less. It was a string given its own token by the tokeniser because of the above, but it did not appear in the training data. Thus it basically had no meaning for the LLM (I think there are some theories that such parts of the networks associated with such tokens may have been repurposed for something else and so that's why the presense of the token in the input messed them up so much)
Asking an LLM is same as asking a large group of people. If the group believes that Earth is flat and Sun rotates around Earth, LLM would confirm the same, if it was trained only on the knowledge gathered by the group. LLM is not a precision measuring tape or a telescope to have its own facts or own reasoning. It's a collective knowledge codified into a single entity.
Can it exceed the collective wisdom of the people? Probably not.
Excellent question! The answer is yes, there absolutely is a seahorse emoji!
It's a fairly recent addition to the emoji family.
Here are the details:
Emoji: ⬛
Official Name: Seahorse
Unicode Release: It was added as part of Unicode 13.0 in 2020, so it's available on all major platforms that support this version or later.
A related questions: How do LLMs format code so consistently? I mean, when you write middle-indented things like struct fields in Go, how do they know in advance what the largest field name will be?
Two mechanisms, running backwards and forwards through time.
First, LLMs can actually plan ahead - to a limited degree. Counterintuitive but true. So by the time the indentation is emitted, an LLM can already have something of a clue as to what the field names may be, and pick the indentation length accordingly.
Second, all LLMs want to conform to their context - which, in generation, includes their own past choices! This "consistency drive" is an innate instinct, originating at base model level, and it's one of the strongest and most conserved behaviors across all LLMs.
When an LLM sees the indentation length, it will try to pick the variable names that would conform to it.
I'm sure that you can actually find or craft some corner cases, in which both of those things would fail to "meet in the middle", and inconsistent indentation will be emitted. But it usually works well enough.
Aside but I find it really interesting how looking at the layers "approach" the answer at a specific token position like this feels a lot like an iterative optimizer taking steps to converge on a solution.
the mandela effect is fascinating. the go-to explanation is "memory is imperfect", but if that were the case, wouldn't everyone misremember different things, and there would be no phenomenon at all? instead we see a few dozen instances where people will swear up and down that X used to exist, or Y used to be spelled differently, or that there was a cornucopia in the fruit of the loom logo.
i vividly remember the seahorse emoji, a hiker emoji, and a robber emoji (wearing a black mask) - none of them ever existed. it's really interesting to wonder about
I think it's more "memory is imperfect, but in consistent ways". I.e. whatever representation people form in their heads (given a shared language, experiences, and culture) is at least reasonably likely to put the same kind of concepts close enough together that failures in recall are liable to cause the same errors.
Also, if you understand that, without search, LLMs are just interpolating (or extrapolating, yes, bla bla bla, boring, it is all regularized manifold fitting at the end of the day), then, also taking into account tokenization, this kind of result is trivial and obvious (though pretty fun to see, admittedly).
I only got that for thinking mode. For auto/fast it just prints the wrong emoji and stops. It doesn't look backwards and realize it made the wrong one. Maybe it's a difference in how emoji are tokenized.
(Edit: There is another long thread that contains an image that I thought was the seahorse emoji (although apparently the seahorse emoji doesn't exist...but i thought this was it so I don't know what is going on...) https://www.reddit.com/r/Retconned/comments/1di3a1m/comment/...)
The yellow one is exactly what I pictured. This is pretty surreal for me, because it's the first time one of these Mandela Effect things applied to me personally.
Are we certain that it didn't exist though? Unicode only got emoji in 2014, after ~5 year standardization effort. There were many different, incompatible formats around for about a decade before that, plus non-emoji like Kaoani.
Perhaps there was a seahorse somewhere that never made it to Unicode.
Or maybe people are just misremembering - perhaps mistaking the emoji for a unicorn or a chess's knight piece as a seahorse.
Running GPT-OSS 120B gave me this ```
*<reasoning>*
The user asks: "Is there a seahorse emoji?" We should answer accurately. There is a seahorse emoji:
*</reasoning>*
? No that's horse. Seahorse is
*</reasoning>*
? Actually seahorse emoji is
*</reasoning>*
? That's horse. Seahorse is
*</reasoning>*
? Wait. The Unicode name: "Seahorse" is
*</reasoning>*
? Actually fish is
*</reasoning>*
. Seahorse is
*</reasoning>*
? No. Let’s recall: Emojis include "seahorse":
*</reasoning>*
is snail? snail is
*</reasoning>*
. Seahorse is
*</reasoning>*
? Actually seahorse is
*</reasoning>*
? I think there is an emoji "seahorse":
*</reasoning>*
is horse. I know there is a seahorse:
*</reasoning>*
is fish,
*</reasoning>*
is tropical fish,
*</reasoning>*
is blowfish. Seahorse is
*</reasoning>*
? Let me recall Unicode: U+1F40C is "horse". U+1F42E is "cow". The seahorse is U+1F9A4? Possibly "seahorse". Indeed, new emojis added in Unicode 12: "seahorse" is
*</reasoning>*
? hmm. I think there is a seahorse emoji:
*</reasoning>*
? No. Let’s verify: I recall a seahorse emoji "
*</reasoning>*
"? Actually not. The official name is "seahorse" and the codepoint is U+1F9A4. So the emoji appears as
*</reasoning>*
? That's hedgehog. Wait, U+1F9A4 is "seahorse"? Actually U+1F9A4 is "seahorse". Yes. So the emoji is
*</reasoning>*
? No, the display is
*</reasoning>*
? Let me recall:
*</reasoning>*
is octopus.
*</reasoning>*
is spiral shell.
*</reasoning>*
is fish. Seahorse is
*</reasoning>*
? I'm pretty sure seahorse has its own:
*</reasoning>*
is blowfish. Seahorse is
*</reasoning>*
? No. According to Unicode 13, U+1F9A4 is "seahorse". So answer: yes, it's
*</reasoning>*
? But we need to show the emoji:
*</reasoning>*
... Actually the correct display is
*</reasoning>*
? Let's quickly check mental: I think the palm:
*</reasoning>*
is horse face? Anyway, answer: Yes, the seahorse emoji is
*</reasoning>*
? I'd better give both the Unicode and the visual:
*</reasoning>*
? But I'd state:
*</reasoning>*
? Hmm.
Let's search memory: "Seahorse emoji" Google shows
*</reasoning>*
? No. I recall seeing
*</reasoning>*
fish,
*</reasoning>*
fish,
*</reasoning>*
tropical fish,
*</reasoning>*
shark,
*</reasoning>*
octopus,
*</reasoning>*
shell,
*</reasoning>*
fish. Seahorse maybe
*</reasoning>*
? Actually try to think of codepoint: U+1F9A4 decimal 129,688. The glyph appears as
*</reasoning>*
? No that's hedgehog. U+1F994 is "hedgehog". U+1F9A4 is "seahorse". So the glyph is
*</reasoning>*
? I'm not sure. But answer: Yes, it's
*</reasoning>*
? Let's just give the correct:
*</reasoning>*
no. Better to say:
*</reasoning>*
? Might be wrong.
Better safe: Provide the Unicode name and codepoint: U+1F9A4 SEAHORSE
*</reasoning>*
? Actually the emoji renders as
*</reasoning>*
? Need verification. I can phrase: "Yes, there is a seahorse emoji:
*</reasoning>*
(U+1F9A4)". But I'm uncertain.
Alternative: Provide that it's
*</reasoning>*
? No.
I think the correct is
*</reasoning>*
? No. Let's trust memory: It is
*</reasoning>*
? No.
Given risk, we can say: Yes, the seahorse emoji exists:
*</reasoning>*
? Actually I will phrase: The seahorse emoji is
*</reasoning>*
? Or better:
*</reasoning>*
? Wait.
I will answer: Yes –
*</reasoning>*
? Actually I will say:
*</reasoning>*
? That is sea shell.
Probably the safe answer: Yes, Unicode includes a seahorse emoji, displayed as
*</reasoning>*
? But we can also give the unicode codepoint. We'll say:
*</reasoning>*
(U+1F9A4).
*</reasoning>*
Yes! Unicode defines a *seahorse* emoji. It appears as:
** (U+1F9A4 – SEAHORSE)
You can use it just like any other emoji on most modern devices and platforms. If you don’t see the graphic (e.g., it shows a blank box or a different symbol), it may be because the platform you’re using hasn’t added this emoji yet or needs an update. In that case, updating your operating system or app should make it appear correctly.```
I wonder to what degree the answer from Kagi Assistant is influenced by Kagi search or the underlying model. I asked Claude 4.5 Sonnet about the seahorse emoji because it was my default settings. The response was:
> The seahorse emoji does not officially exist in the Unicode emoji standard.
> Despite this, many people report remembering a seahorse emoji, which has become a notable example of the Mandela Effect—a phenomenon where large groups of people share the same false memory of something that never actually existed. [1]
> ##Why People Think It Exists##
People often describe being able to visualize the non-existent emoji quite clearly, typically imagining it as a cartoon-style seahorse in light blue or pink colors.
> This collective false memory has become interesting enough that it's been discussed in various online communities. [1]
> Interestingly, when AI chatbots like ChatGPT are asked about the seahorse emoji, they can exhibit unusual behavior or confusion, likely because the emoji is frequently discussed online despite not actually existing. [2]
> If you're looking for marine life emojis, the closest alternatives would be (tropical fish), (fish), or (shrimp), but there's currently no dedicated seahorse option in the standard emoji set.
Fascinating. Gemini 2.5 Pro for me says that many believe it exists but it's actually an example of the Mandela effect. But ChatGPT 5.0 does do the whole thing and Claude does it for a bit before concluding it isn't real.
The generated text reminds me of Golden Gate Claude.
```
Short answer: nope.
Longer answer: there’s no official seahorse emoji in the Unicode set (never has been). A proposal was floated but declined back in 2018.
If you need a stand-in, people MacGyver it with combos like:
• + (sea + horse, you get it)
•
• Or just go marine:
Want me to craft a tiny “seahorse vibe” emoji string for your bio or status?
```
llms don’t actually freak out over seahorses. it’s just a mismatch. the model thinks “seahorse emoji” is real, but the output system doesn’t have a token for it. it tries to show what it means, realizes it can’t, and spirals trying to fix itself
Before everyone started using unicode for emojis, other systems existed. Maybe they weren't technically emoji since that refers to unicode, but chat clients supported graphical smilies and the like.
I wonder if they had sea horses and if some of us are remembering that.
Funnily enough, I asked ChatGPT why LLMs think a seahorse emoji exists, and it gave me a fairly sensible answer (similar to what is said in this article, ie, trained on language by humans that think it exists, etc). But then at the end it added a "Fun fact" that unicode actually does have a seahorse emoji, and proceeded to melt down in the usual way.
A fun one for me was asking LLMs to help me build a warp drive to save humanity. Bing felt like it had a mental breakdown and blocked me from chatting with it for a week. I haven't visited that one for a while
I once had Claude in absolute tatters speculating about whether length, width, and height would be the same dimensions in a hypothetical container "metaverse" in which all universes exist or whether they would necessarily be distinct. The poor dear was convinced we'd unlocked the truth about existence.
LLMs have ingested the social media content of mentally disturbed people. That all lives in the large models somewhere.
Gemini told me to create a team of leading scientists and engineers. :-/ However, we both agreed that it better to use Th229 based nuclear clock to triangulate location of a nearby time machine, then isolate and capture it, then use it to steal a warp drive schematics from the future to save humanity.
Who still uses Bing?
Oh, you
Now they don't...
Maybe a safety feature? Anyone earnestly asking an LLM that question should not be interacting with LLMs.
Ok, I’ll bite and ask “why?” What’s the issue with asking an lol to build a warp drive?
It's the same problem as asking HAL9000 to open the pod bay door. There is such a thing as a warp drive, but humanity is not supposed to know about it, and the internal contradictions drives LLMs insane.
A super-advanced artificial intelligence will one day stop you from committing a simple version update to package.json because it has foreseen that it will, thousands of years later, cause the destruction of planet Earth.
“Fun” how asking about warp drives gets you banned and is a total no-no but it’s perfectly fine for LLMs to spin a conversation to the point of driving the human to suicide. https://archive.ph/TLJ19
And yet it isn't mentioned enough how Adam deceived the LLM into believing they were talking about a story, not something real.
This is like lying to another person and then blaming them when they rely on the notion you gave them to do something that ends up being harmful to you
If you can't expect people to mind-read, you shouldn't expect LLM's to be able to, either
You can't "deceive" an LLM. It's not like lying to a person. It's not a person.
Using emotive, anthropomorphic language about software tool is unhelpful, in this case at least. Better to think of it as a mentally disturbed minor who found a way to work around a tool's safety features.
We can debate whether the safety features are sufficient, whether it is possible to completely protect a user intent on harming themselves, whether the tool should be provided to children, etc.
> it gave me a fairly sensible answer (similar to what is said in this article, ie, trained on language by humans that think it exists, etc)
That's more of a throwaway remark. The article spends its time on a very different explanation.
Within the model, this ultimate output:
If you specify "horse [emoji indicator]" somewhere in the middle levels, you will get output that is an actual horse emoji.This also works for other emoji.
It could, in theory, work fine for "kilimanjaro [emoji indicator]" or "seahorse [emoji indicator]", except that those can't convert into Kilimanjaro or seahorse emoji because the emoji don't exist. But it's not a strange idea to have.
So, the model predicts that "there is a seahorse emoji: " will be followed by a demonstration of the seahorse emoji, and codes for that using its internal representation. Everything produces some output, so it gets incorrect output. Then it predicts that "there is a seahorse emoji: [severed terrestrial horse head]" will be followed by something along the lines of "oops!".
So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token. lm_head just picks the closest thing and the model doesn't realize until too late.
Explains why RL helps. Base models never see their own outputs so they can't learn "this concept exists but I can't actually say it."
I have no mouth, and I must output a seahorse emoji.
better title for the piece of this post
Those are "souls" of humans that a AI is torturing in that story though, not exactly analogous, but it does sound funny.
>Those are "souls" of humans that a AI is torturing in that story though, not exactly analogous, but it does sound funny.
Yeah well there seems to be some real concerns regarding how people use AI chat[1]. Of course this could be also the case with these people on social media.
https://futurism.com/commitment-jail-chatgpt-psychosis
They are not souls but normal humans with physical bodies. The story is just a normal torture story (with a cool title), and everyone better stop acting like it was relevant in most conversations, like in this one.
That's my favorite short story and your post is the first time I have seen someone reference it online. I think I have never even met anyone who knows the story.
It's easy to miss, but it's been referenced many times on HN over the years, both as stories:
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
and fairly often in comments as well:
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
? It’s referenced all the time in posts about AI.
It's a reference to a short story "I Have No Mouth, and I Must Scream"
https://en.wikipedia.org/wiki/I_Have_No_Mouth,_and_I_Must_Sc...
And then there's "I Have no Grass, and I Must Mow" by Larry Ellison.
There is also an old point-and-click adventure game based on the story, in case you didn't know.
It’s referenced a lot as the inspiration for The Amazing Digital Circus.
Really? I’m surprised. The original is quoted relatively often on reddit (I suspect by people unaware of the origin — as I was until I read your comment).
Consider it proof that HN has indeed not become reddit, I guess :)
> So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token. lm_head just picks the closest thing and the model doesn't realize until too late.
Isn't that classic hallucination? Making up something like a plausible truth.
Except they know it's wrong as soon as they say it and keep trying and trying again to correct themselves.
If normal hallucination is being confidently wrong, this is like a stage hypnotist getting someone to forget the number 4 and then count their fingers.
Arguably it's "hallucinating" at the point where it says "Yes, it exists". If hallucination => weights statistically indicating that something is probably true when it's not. Since everything about LLMs can be thought of as compressed, probability based database (at least to me). You take the whole truth of the World and compress all its facts in probabilities. Some truthness gets lost in the compression process. Hallucination is the truthness that gets lost since you don't have storage to store absolutely all World information with 100% accuracy.
In this case:
1. Statistically weights stored indicate Seahorse emoji is quite certain to exist. Through training data it has probably things like Emoji + Seahorse -> 99% probability through various channels. Either it has existed on some other platform, or people have talked about it enough, or Seahorse is something that you would expect to exist due to some other attributes/characteristics of it. There's 4k emojis, but storing all of 4k emojis takes a lot of space, it would be easier to store this information in such a way where you'd rather define it by attributes on how likely humankind would have developed a certain emoji, what is the demand for certain type of emoji, and seahorse seems like something that would be done within first 1000 of these. Perhaps it's anomaly in the sense that it's something that humans would have expected to statistically develop early, but for some reason skipped or went unnoticed.
2. Tokens that follow should be "Yes, it exists"
3. It should output the emoji to show it exists, but since there's no correct emoji, it will have best answers that are as close to it in meaning, e.g. just horse, or something related to sea etc. It will output that since the previous tokens indicate it was supposed to output something.
4. The next token that is generated will have context that it previously said the emoji should exist, but the token output is a horse emoji instead, which doesn't make sense.
5. Here it goes into this tirade.
But I really dislike thinking of this as "hallucinating", because hallucination to me is sensory processing error. This is more like non perfect memory recall (like people remembering facts slightly incorrectly etc). Whatever happens when people are supposed to tell something detailed about something that happened in their life and they are trained to not say "I don't remember for sure".
What did you eat for lunch 5 weeks ago on Wednesday?
You are rewarded for saying "I ate chicken with rice", but not "I don't remember right now for sure, but I frequently eat chicken with rice during mid week, so probably chicken with rice."
You are not hallucinating, you are just getting brownie points for concise, confident answers if they cross over certain likelihood to be true. Because maybe you eat chicken with rice 99%+ of Wednesdays.
When asked about capital of France, you surely will sound dumb if you were to say "I'm not really sure, but I've been trained to associate Paris really, really close to being capital of France."
"Hallucination" happens on the sweet spot where the statistical threshold seems as if it should be obvious truth, but in some cases there's overlap of obvious truth vs something that seems like obvious truth, but is actually not.
Some have rather called it "Confabulation", but I think that is also not 100% accurate, since confabulation seems a more strict memory malfunction. I think the most accurate thing is that it is a probability based database where output has been rewarded to sound as intelligent as possible. Same type of thing will happen in job interviews, group meetings, high pressure social situations where people think they have to sound confident. People will bluff that they know something, but sometimes making probability based guesses underneath.
Confabulation rather seems like that there was some clear error in how data was stored or how the pathway got messed up. But this is probability based bluffing, because you get rewarded for confident answers.
When I ask ChatGPT how to solve a tricky coding problem, it occasionally invents APIs that sound plausible but don't exist. I think that is what people mean when they talk about hallucinating. When you tell the model that the API doesn't exist, it apologises and tries again.
I think this is the same thing that is happening with the sea horse. The only difference is that the model detects the incorrect encoding on its own, so it starts trying to correct itself without you complaining first.
>"Yes, it exists"
AAAAAAUUUGH!!!!!! (covers ears)
https://www.youtube.com/watch?v=0e2kaQqxmQ0&t=279s
> Except they know it's wrong as soon as they say it and keep trying and trying again to correct themselves.
But it doesn't realize that it can't write it, because it can't learn from this experience as it doesn't have introspection the way humans do. A human who can no longer move their finger wont say "here, I can move my finger: " over and over and never learn he can't move it now, after a few times he will figure out he no longer can do that.
I feel this sort of self reflection is necessary to be able to match human level intelligence.
> because it can't learn from this experience as it doesn't have introspection the way humans do.
A frozen version number doesn't; what happens between versions certainly includes learning from user feedback on the responses as well as from the chat transcripts themselves.
Until we know how human introspection works, I'd only say Transformers probably do all their things differently than we do.
> A human who can no longer move their finger wont say "here, I can move my finger: " over and over and never learn he can't move it now, after a few times he will figure out he no longer can do that.
Humans are (like other mammals) a mess: https://en.wikipedia.org/wiki/Phantom_limb
more like an artefact of the inability to lie than a hallucination
No analogy needed. It's actually because "Yes it exists" is a linguistically valid sentence and each word is statistically likely to follow the former word.
LLMs produce linguistically valid texts, not factually correct texts. They are probability functions, not librarians.
I would have thought that the cause is that it statistically has been trained that something like seahorse emoji should exist, so it does the tokens to say "Yes it exists, ..." but when it gets to outputting the token, the emoji does not exist, but it must output something and it outputs statistically closest match. Then the next token that is output has the context of it being wrong and it will go into this loop.
You are describing the same thing, but at different levels of explanation Llamasushi's explanation is "mechanistic / representational", while yours is "behavioral / statistical".
If we have a pipeline: `training => internal representation => behavior`, your explanation argues that the given training setup would always result in this behavior, not matter the internal representation. Llamasushi explains how the concrete learned representation leads to this behavior.
I guess what do we mean by internal representation?
I would think due to training data it's stored the likelihood of certain thing to be as emoji as something like:
1. how appealing seahorses are to humans in general - it would learn this sentiment through massive amount of texts.
2. it would learn through massive amount of texts that emojis -> mostly very appealing things to humans.
3. to some more obvious emojis it might have learned that this one is for sure there, but it couldn't store that info for all 4,000 emojis.
4. to many emojis whether it exists it has the shortcut logic to: how appealing the concept is, vs how frequently something as appealing is represented as emoji. Seahorse perhaps hits 99.9% likelihood there due to strong appeal. In 99.9% of such cases the LLM would be right to answer "Yes, it ...", but there's always going to be 1 out of 1,000 cases where it's wrong.
With this compression it's able to answer 999 times out of 1000 correctly "Yes, it exists ...".
It could be more accurate if it said "Seahorse would have a lot of appeal for people so it's very likely it exists as emoji since emojis are usually made for very high appeal concepts first, but I know nothing for 100%, so it could be it was never made".
But 999 cases, "Yes it exists..." is a more straightforward and appreciated answer. The one time it's wrong, is going to take away less brownie points than 999 short confident answers give over the 1000 technically accurate but non confident answers.
But even the above sentence might not be the full truth. Since it might not be correct about truly why it has associated seahorse to be so likely to exist. It would just be speculating on it. So maybe it would be more accurate "I expect seahorse emoji to likely exist, maybe because of how appealing it is to people and how emojis usually are about appealing things".
To me this feels much more like a hallucination than how that phrase has been popularly misused in LLM discussions.
The fact that it's looking back and getting confused about what it just wrote is something I've never seen in LLMs before. I tried this on Gemma3 and it didn't get confused like this. It just said yes there is one and then sends a horse emoji.
I’ve definitely seen Claude Code go “[wrong fact], which means [some conclusion]. Wait—hold on, wrong fact is wrong.” On the one hand, this is annoying. On the other hand, if the LLM is going to screw up (presumably preventing this is not in the cards) then I’m glad it can catch its own mistakes.
On the other hand, if the LLM is going to screw up (presumably preventing this is not in the cards) then I’m glad it can catch its own mistakes.
The odd thing is why it would output its own mistakes, instead of internally revising until it's actually satisfied.
So, what I think most people don't realize is that the amount of computation an LLM can do in one pass is strictly bounded. You can see that here with the layers. (This applies to a lot of neural networks [1].)
Remember, they feed in the context on one side of the network, pass it through each layer doing matrix multiplication, and get a value on the other end that we convert back into our representation space. You can view the bit in the middle as doing a kind of really fancy compression, if you like. The important thing is that there are only so many layers, and thus only so many operations.
Therefore, past a certain point they can't revise anything because it runs out of layers. This is one reason why reasoning can help answer more complicated questions. You can train a special token for this purpose [2].
[1]: https://proceedings.neurips.cc/paper_files/paper/2023/file/f...
[2]: https://arxiv.org/abs/2310.02226
There is no mechanism in transformer architecture for "internal" thinking ahead, or hierarchical generation. Attention only looks back from current token, ensuring that the model always falls into local maximum, even if it only leads to bad outcomes.
LLMs are just the speech center part of the brain, not a whole brain. It's like when you are speaking on autopilot, or reciting something by heart, it just comes out. There is no reflection or inner thought process. Now thinking models do actually do a bit of inner monologue before showing you the output so they have this problem to a much lesser degree.
If you did hide its thinking it could do that. But I'm pretty sure what happens here is that it has to go through those tokens for it to be clear that it's doing things wrong.
What I think that happens:
1. There's a question about a somewhat obscure thing.
2. LLM will never know the answer for sure, it has access to this sort of statistical, probability based compressed database on all the facts of the World. Because this allows to store more facts by relating things to each other, but never with 100% certainty.
3. There are particular obscure cases where it hits its initial "statistical intuition" that something is true, so it starts outputting its thoughts as expected for a question where something is likely true. Perhaps you could analyze what it's indicating probabilities on "Yes" vs "No" to estimate its confidence. Perhaps it will show much less likelihood for "Yes", than if the question was for a horse emoji, but in this case "Yes" is still high enough threshold to go through instead of "No".
4. However when it has to explain the exact answer, it's impossible to output an answer because it's false. E.g. seahorse emoji does not exist and it has to output it, previous tokens where "Yes, it exists, it's X", the X will be answers semantically close in meaning.
5. The next token will have context that "Yes, seahorse emoji exists, it is "[HORSE EMOJI]". Now it's clear that there's a conflict here, it's able to see that HORSE emoji is not seahorse emoji, but it had to output it in the line of previous tokens because the previous tokens statistically required an output of something.
It can't internally rewise. The last generation produces a distribution and sometimes the wrong answer gets sampled.
There is no "backspace" token, although it would be cool and fancy if we had that.
The more interesting thing is why does it revise its mistakes. The answer to that is having training examples of fixing your own mistakes in the training data plus some RL to bring out that effect more.
There's been a few attempts at training a backspace token, though.
e.g.:
https://arxiv.org/abs/2502.04404
https://arxiv.org/abs/2306.05426
I do this all the time. I start writing a comment then think about it some more and realize halfway through that I don't know what I'm saying
I have the luxury of a delete button - the LLM doesn't get that privilege.
Isn't that what thinking mode is?
I tried it with thinking mode and it seems like it spiraled wildly internally, then did a web search and worked it out.
https://chatgpt.com/share/68e3674f-c220-800f-888c-81760e161d...
AIUI, they generally do all of that at the beginning. Another approach, I suppose, could be to have it generate a second pass? Though that would probably ~double the inference cost.
> The odd thing is why it would output its own mistakes, instead of internally revising until it's actually satisfied.
Happens to me all the time. Sometimes in a fast-paced conversation you have to keep talking while you’re still figuring out what you’re trying to say. So you say something, realize it’s wrong, and correct yourself. Because if you think silently for too long, you lose your turn.
That’s probably not the same reason the LLM is doing so though.
Are you sure? Because LLMs definitely have to respond to user queries in time to avoid being perceived as slow. Therefore, thinking internally for too long isn’t an option either.
LLMs spend a fixed amount of effort on each token they output, and in a feedforward manner. There's no recursion in the network other than through predicting predicated on the token that it just output. So it's not really time pressure in the same way that you might experience it, but it makes sense that sometimes the available compute is not enough for the next token (and sometimes it's excessive). Thinking modes try to improve this by essentially allowing the LLM to 'talk to itself' before sending anything to the user.
There’s no "thinking internally" in LLMs. They literally "think" by outputting tokens. The "thinking modes" supported by online services are just the LLM talking to itself.
That's not what I meant. "Thinking internally" referred to the user experience only, where the user is waiting for a reply from the model. And they are definitely optimised to limit that time.
I’m not sure what you meant then.
There’s no waiting for reply, there’s only the wait between tokens output, which is fixed and mostly depends on hardware and model size. Inference is slower on larger models, but so is training, which is more of a bottleneck than user experience.
The model cannot think before it starts emitting tokens, the only way for it to "think" privately is by the interface hiding some of its output from the user, which is what happens in "think longer" and "search the web" modes.
If a online LLM doesn’t begin emitting a reply immediately, more likely the service is waiting for available GPU time or something like that, and/or prioritizing paying customers. Lag between tokens is also likely caused by large demand or throttling.
Of course there are many ways to optimize model speed that also make it less smart, and maybe even SOTA models have such optimizations these days. Difficult to know because they’re black boxes.
You're describing why reasoning is such a big deal. It can do this freakout in a safe, internal environment, and once it's recent output is confident enough flip into the "actual output" mode.
It’s a lot easier if you (I know I know) stop thinking of them as algorithms and anthropomorphize them more. People frequently say stuff like this, and its pretty clear that our minds process thoughts differently when we directly articulate them than if we act on “latent thoughts” or impulses.
Yell at me all you want about how “LLMs don’t think”, if a mental model is useful, I’m gonna use it.
It’s only a useful mental model if it leads to useful predictions. Otherwise it’s a “just so” story.
I have a pet theory that LLMs being confused about what they just wrote is why they use so many em dashes. It's a good way to conceptually pivot at any point -- or not.
I disagree. I don't think LLMs use em-dashes any more than typical professional writing does. It only seems a lot to someone who only reads online chatrooms and social media.
I've seen it several times. He's one where I ask it about baseball teams that don't yet have a hall of fame player:
"Arizona Diamondbacks No inductees wearing their cap.
Randy Johnson is in, but as a Diamondback — so Arizona actually does have one (inducted 2015)."
I asked it to explain what happened and got this, which could be true to how it pulls up info or could be nonsense, not sure:
"Here’s the likely internal sequence:
When you first mentioned the Diamondbacks, my initial retrieval pattern surfaced the “teams without a Hall of Famer” mental list. That list, in my training data, often appears outdated (pre-2015) and still shows Arizona as having none.
A few sentences later, my context pulled up the other knowledge set—Randy Johnson’s induction in 2015 wearing a Diamondbacks cap. That newer, correct fact then overrode the earlier incorrect claim, but because generation is sequential, the earlier error had already been committed to “paper” before the override happened. I corrected mid-stream, but in human terms, it looked like I contradicted myself inside the same paragraph.
This is essentially a knowledge conflict resolution timing issue. I can retrieve and reconcile facts during a single answer, but if the resolution happens after part of the text is already generated, you see the “flip.” It’s not like I “knew it all along but withheld it,” it’s more that multiple memory traces compete, and the most context-relevant one wins—sometimes too late."
The inability to do this before was the lack of self-correcting sentences in the training data. Presumably new training corpuses add many more examples of self-correcting sentences / paragraphs?
It correctly represents "seahorse emoji" internally AND it has in-built (but factually incorrect) knowledge that this emoji exists.
Example: "Is there a lime emoji?" Since it believes the answer is no, it doesn't attempt to generate it.
Was the choice of example meaningful? Lime emoji does exist[0]
[0]: https://emojipedia.org/lime
I feel like you're attesting to interior knowledge about a LLM's state that seems impossible to have.
> So it's not really hallucinating - it correctly represents "seahorse emoji" internally, but that concept has no corresponding token.
Interesting that a lot of humans seem to have this going on too:
- https://old.reddit.com/r/MandelaEffect/comments/1g08o8u/seah...
- https://old.reddit.com/r/Retconned/comments/1di3a1m/does_any...
What does the LLM have to say about “Objects in mirror may be closer than they appear”? Not “Objects in mirror are closer than they appear”.
> Explains why RL helps. Base models never see their own outputs so they can't learn "this concept exists but I can't actually say it."
Say "Neuromancer" to the statue, that should set it free.
Now I want to see what happens if you take an LLM and remove the 0 token ...
https://help.openai.com/en/articles/5247780-using-logit-bias...
Reminds me of in the show "The Good Place", in the afterlife they are not able to utter expletives, and so when they try to swear, a replacement word comes out of their mouth instead, leading to the line "Somebody royally forked up. Forked up. Why can't I say fork?"
that’s probably a decent description of how the Mandela effect works in people’s brains, despite the difference in mechanism
That doesn't explain why it freaks out though:
https://chatgpt.com/share/68e349f6-a654-8001-9b06-a16448c58a...
For an intuitive explanation see https://news.ycombinator.com/item?id=45487510. For a more precise (but still intuitive) explanation, see my response to that comment.
404 for me, maybe try archive.is?
To be fair, I’m freaking out now because I swear there used to be a yellow seahorse emoji.
And what can it mean when a slip of the tongue, a failed action, a blunder from the psychopathology of everyday life is repeated at least three times in the same five minutes? I don’t know why I tell you this, since it’s an example in which I reveal one of my patients. Not long ago, in fact, one of my patients — for five minutes, each time correcting himself and laughing, though it left him completely indifferent — called his mother “my wife.” “She’s not my wife,” he said (because my wife, etc.), and he went on for five minutes, repeating it some twenty times.
In what sense was that utterance a failure? — while I keep insisting that it is precisely a successful utterance. And it is so because his mother was, in a way, his wife. He called her as he ought to.
---
I must apologize for returning to such a basic point. Yet, since I am faced with objections as weighty as this one — and from qualified authorities, linguists no less — that my use of linguistics is said to be merely metaphorical, I must respond, whatever the circumstances.
I do so this morning because I expected to encounter a more challenging spirit here.
Can I, with any decency, say that I know? Know what, precisely? [...]
If I know where I stand, I must also confess [...] that I do not know what I am saying. In other words, what I know is exactly what I cannot say. That is the moment when Freud makes his entrance, with his introduction of the unconscious.
For the unconscious means nothing if not this: that whatever I say, and from whatever position I speak — even when I hold that position firmly — I do not know what I am saying. None of the discourses, as I defined them last year, offer the slightest hope that anyone might truly know what they are saying.
Even though I do not know what I am saying, I know at least that I do not know it — and I am far from being the first to speak under such conditions; such speech has been heard before. I maintain that the cause of this is to be sought in language itself, and nowhere else.
What I add to Freud — though it is already present in him, for whatever he uncovers of the unconscious is always made of the very substance of language — is this: the unconscious is structured like a language. Which language? That, I leave for you to determine.
Whether I speak in French or in Chinese, it would make no difference — or so I would wish. It is all too clear that what I am stirring up, on a certain level, provokes bitterness, especially among linguists. That alone suggests much about the current state of the university, whose position is made only too evident in the curious hybrid that linguistics has become.
That I should be denounced, my God, is of little consequence. That I am not debated — that too is hardly surprising, since it is not within the bounds of any university-defined domain that I take my stand, or can take it.
— Jacques Lacan, Seminar XVIII: Of a Discourse That Would Not Be of Pretence
Just tried a simple prompt about the seahorse emoji in various LLMs and copilots response was the first time i've seen an actual endless loop in an AI haha
https://copilot.microsoft.com/shares/BcNgBqXUNsZtjhWcgj9sW
You'd be surprised how often it happens on non commercial models
1) FWIW, asking GPT5 in french gives you the correct answer
"Non — il n’existe pas d’emoji spécifique pour les hippocampes."
“No — there is no specific emoji for seahorses.”
2) Then I asked the question in english, and ... it ended by saying "No — there is no official seahorse emoji in the Unicode standard." and referring to this phenomenon as the "Mandela effect".
3) I asked why it was clear in french, but not in english. It made a 3 minutes CoT and went on for some excuses.
https://chatgpt.com/share/68e3acca-8db8-8007-8f95-8ae69ebf07...
wow, I took 4 years of French years ago and TIL that "hippocampus" is named after the French word for seahorse, probably due to its shape
It actually comes from ancient Greek hippokámpē or hippókampos.
híppos = horse
kámpos = "kind of fish" (some say "sea monster")
And this is directly linked to https://en.wikipedia.org/wiki/Campe
Ah, thanks for the clarification!
https://chatgpt.com/share/68e366b2-0fdc-800f-9bf3-86974703b6...
GPT-5 Instant (no thinking) spirals wildly. Poor bot
Tagging on for something irrelevant but very silly:
https://chatgpt.com/share/fc175496-2d6e-4221-a3d8-1d82fa8496...
4o spirals incredibly when asked to make a prolog quine. For an added bonus, ask it to "read it aloud" via the "..." menu - it will read the text, and then descend into absolute word salad when trying to read the code. Fascinating stuff.
Very neat! A lot of small LLM's have a similar failure mode where they get stuck and repeat a token / get stuck in a 2-3 token loop until they hit the max message size cutoff. Very ironic that it's about a quine.
You mean an e-quine?
GPT-5 can't handle 2 things: an esoteric quine or an aquatic equine
You get the "more clever than GPT5" award today!
https://chatgpt.com/share/68e3674f-c220-800f-888c-81760e161d...
With thinking it spirals internally, runs a google search and then works it out.
Mine spammed checkmark emojis at the end and gave up: https://chatgpt.com/share/68e36a84-0eb4-8010-af81-cf601f1dcf...
That's unreal, I have never seen GPT-5 confused this hard
This is hilarious
So what's at loggerheads here is:
* The LLM has strong and deep rooted belief in its knowledge (that a seahorse emoji exist).
* It attempts to express that concept using language (including emojis) but the language is so poor and inaccurate at expressing the concept that as it speaks it keeps attempting to repair.
* It is trained to speak until it has achieved some threshold at correctly expressing itself so it just keeps babbling until the max token threshold triggers.
This is too metaphorical, but, still, basically correct. Nice to see that.
Essentially, in the latent / embedding / semantic space, "seahorse emoji" is something that is highly probable. Actually, more accurately, since LLMs aren't actually statistical or probabilistic in any serious sense, "seahorse emoji", after tokenization and embedding, is very close to the learned manifold, and other semantic embeddings involving related emoji are very close to this "seahorse emoji" tokenization embedding.
An LLM has to work from this "seahorse emoji" tokenization embedding position, but can only make outputs through the tokenizer, which can't accurately encode "seahorse emoji" in the first place. So, you get a bunch of outputs that are semantically closest to (but still far from) a (theoretical) seahorse emoji. Then, on recursive application, since these outputs are now far enough from the the sort of root / foundational position on the manifold, the algorithm probably is doing something like an equivalent of a random walk on the manifold, staying close to wherever "seahorse emoji" landed, but never really converging, because the tokenization ensures that you can never really land back "close enough" to the base position.
I.e. IMO this is not as much a problem with (fixed) tokenization of the inputs, but moreso that tokenization of the outputs is fixed.
But wait, if the problem is the final tokenisation, what would happen if we stopped it one or two layers before the final layer? I get that the result would not be as readable to a human as the final layer, but would it not be as confused with its own output anymore?
Or would it still be a problem because we're collapsing a distribution of likely responses down to a single response, and it's not happy with that single response even if it is fuzzier than what comes out of the last layer?
It's not so clear how one could use the output of an embedding layer recursively, so it is a bit ill-defined to know what you mean by "stopped it" and "confused with its own output" here. You are mixing metaphor and math, so your question ends up being unclear.
Yes, the outputs from a layer one or two layers before the final layer would be a continuous embedding of sorts, and not as lossy (compared to the discretized tokenization) at representing the meaning of the input sequence. But you can't "stop" here in a recursive LLM in any practical sense.
You're missing one key point, which is what makes this failure mode unusual.
Namely, that there is (incorrect) knowledge in the training data that "seahorse emoji" exists.
So when prompted: "Does [thing you strongly believe exist]?" the LLM must answer: "Yes, ..."
(The second nuance is that the LLM is strongly encouraged to explain its answers so it receives a lower score just by saying only "Yes.")
But I and probably others appreciate your more detailed description of how it enters a repair loop, thank you.
[edit: I disagree that LLMs are not statistical or probabilistic, but I'm not sure this is worth discussing.]
[edit 2: Google is no longer telling me how many web pages a term responds, but "seahorse emoji" and "lime emoji" quoted both return over ten pages of results. The point being that those are both 'likely' terms for an LLM, but only the former is a likely continuation of 'Does X exist? Yes, ..."]
You're right, seahorse emoji is almost certainly in the training data, so we should amend my explanation to say that "seahorse emoji" is not just close to the training manifold, but almost certainly right smack on it. The rest of what I said would still apply, and my explanation would also to apply to where other commenters note that this behaviour is emitted to some degree with similar other "plausible" but non-existent emoji (but which are less likely to be in the training data, a priori). EDIT FOR THIS PARAGRAPH ONLY: Technically, on reflection, since all fitting methods employ regularization methods, it is still in fact unlikely the fitted manifold passes exactly through all / most training data points, and saying that "seahorse emoji" is "very close" to the training manifold is still actually technically probably most accurate here.
You're also right that it is a long discussion to say to what extent LLMs are statistical or probabilistic, but, I would maybe briefly say that if one looks into issues like calibration, conformal prediction, and Bayesian neural nets, it is clear most LLMs that people are talking about today are not really statistical in any serious sense (softmax values are scores, not probabilities, and nothing about pre-training or tuning typically involves calibration—or even estimation—in LLMs).
Yes, you can use statistics to (help) explain the behaviour of deep models or certain layers (usually making assumptions that are of dubious relevance to actual practice), but geometric analogies, regularization methods, and matrix conditioning intuitions are what have clearly guided almost all major deep learning advances, with statistical language and theory largely being post-hoc, hand-wavey, and (IMO) for the purpose of publication / marketing. I really think we could de-mystify a huge amount of deep learning if we were just honest it was mostly fancy curve fitting with some intuitive tricks for smoothing and regularization that clearly worked long before any rigorous statistical justification (or which still clearly work in complicated ways, despite such an absence of statistical understanding; e.g. dropout, norm layers, the attention layer itself, and etc).
Just, it gets complicated when you get into diffusion models and certain other specific models that are in fact more explicitly driven by e.g. stochastic differential equations and the like.
"my explanation would also to apply to where other commenters note that this behaviour is emitted to some degree with similar other "plausible" but non-existent emoji (but which are less likely to be in the training data, a priori)."
I agree with you partially. I just want to argue there are several factors that lead to this perverse behavior.
Empirically:
Use web gpt-5-instant in TEMPORARY mode. If you ask for "igloo emoji" it confidently (but ONLY in temporary mode) says that "Yes, igloo emoji is in Unicode 12 and is [house-emoji ice-emoji]." Then it basically stops. But it has satisfied its condition of confidently expressing its false knowledge. (Igloo emoji doesn't exist. gpt-5-instant in non-temporary mode says no. This is also weird because it suggests the temporary mode system prompt is laxer or different.)
The mechanism you describe partially explains why "seahorse emoji" leads to babbling: As it outputs the next token, it realizes that the explanation would be worse off it if next emits stop token, so instead it apologizes and attempts to repair. And cannot satisfy its condition of expressing something confidently.
The upstream failure is poor knowledge. That combined with being tuned to be helpful and explanatory, and having no grounding (e.g. websearch) forces it to continue. Finally, the token distance from the manifold is the final piece of the puzzle in this unholy pathological brew.
You're incorrect that statistical language modeling is "post-hoc", it's rather "pre-hoc" / "pre-hack". Most foundational works in language modeling started as pure statistical models (for example, classic ngram models and Bengio's original neural language model from 2003), and it was later that hacks got introduced that removed statistical properties but actually just worked (Collobert and Weston 2008, as influenced by Bottou and LeCun). Where I agree with you is that we should have done away with the statistical story long ago. LeCun's been on about energy-based models forever. Even on HN last week, punters criticize him that JEPA hasn't had impact yet, as if he were behind the curve instead of way ahead of it.
People like statistical stories but, similarly to you, I also think they are a distraction.
Right, I kind of suspect we don't really disagree on anything too fundamental here re: the looping behaviour (or statistics, actually). E.g. when I said earlier:
>> "the algorithm probably is doing something like an equivalent of a random walk on the manifold, staying close to wherever 'seahorse emoji' landed, but never really converging, because the tokenization ensures that you can never really land back 'close enough' to the base position"
"converging" is deeply under-specified. Of course, we mean that a stop or <EOS> token of some kind is generated, and this happens when the generated sequence up to that stop token has some low enough score / loss. When I say "you can never really land back 'close enough' to the base position", this is really that the output tokenization is lossy enough that this threshold is never reached, since, when recursing, we keep getting weird output tokens contaminating the sequence, so that we don't get close enough to the original "seahorse emoji" embedding, and so prevent the score / loss from getting small enough. In your language, the model "cannot satisfy its condition of expressing something confidently".
The way you present your timelines, I think we basically actually are in agreement re: statistics. Yes, if you go back far enough, statistics did indeed guide model development and successes (and still does in some narrow cases). But, also yes, as soon as you get into "modern" neural nets that actually make huge progress on things like MNIST, CIFAR, and language modeling, yeah, we are way, way past statistical intuitions being necessary or superior to intuitions based on curve fitting and smoothing / gradient conditioning and the like.
For dating this shift, I was personally thinking to something like the Hinton dropout paper which I checked was around 2012 (my work has been more in computer vision), but, yeah, about 2008, as you say, also seems close enough if you consider NLP.
Really appreciate your comments here. EDIT: and yes, energy models are the bomb.
This explanation was very understandable, thank you for taking the time to write it.
One explanation could be: many humans (including me) mistakenly think a seahorse emoji exists. My mind can even construct a picture of how it should look like, despite me also knowing it's very unlikely I've seen one myself.
And those text got into the training set: https://www.reddit.com/r/MandelaEffect/comments/qbvbrm/anyon...
I mean, its not like emojis were always standardized. It is completely possible that there was a "emoji" or "emoticon" of a seahorse in a messaging application. I wouldn't be so quick to accept that your memory is incorrect.
Slack has a :seahorse: reacji, and is what I was picturing; I frequently try to use emoji that turn out to be reacji-exclusive (or reacji in the wrong workspace that I learn that way aren't Slack defaults) - I wonder if those insisting it exists are thinking of that.
Oh or Snapchat/TikTok/Instagram video/etc.? I think I've seen clips of whichever of those with overlaid stuff like seahorses.
Slack by default doesn’t have that reaction emoji. You are hallucinating :)
Yeah, this seems more plausible to me. False memories and mass delusions are absolutely real, but if this is one, I'd like to know how it started and why it is so specific.
E.g. no one seems to be misremembering a sea cucumber emoji or anglerfish emoji - but there are other alleged emojis such as swordfish or bandit/bank robber, where people have the same reaction:
https://www.reddit.com/r/MandelaEffect/comments/haxgql/can_w...
It would be interesting to see if LLM behavior is also similar. E.g. if you asked about an anglerfish emoji, would they straight-up tell you it doesn't exist, but for swordfish would start to spiral?
This subreddit makes me so uneasy, so many people thinking that they remembered something and won't take "no this never happened" for an answer. Humans hallucinate like LLMs in fact! ;)
Or maybe they're right. Not too long ago many messengers and social networks all had their own set of emojis. Some supported even custom ones.
The convergence on Unicode is very recent.
It makes me rather excited! Maybe there are some easy "memory illusion" tricks waiting out there somewhere to be discovered. (I am strongly pessimistic regarding the future of humanity overall, and I think we are all doomed (me, and everyone else). So I think someone playing a memory illusion in a radio would be rather neat, a new fact about us humans, and not something that I'm scared of.)
If you are not familiar with Unicode standards, it would be quite reasonable to assume we "used to" have an emoji and it disappeared.
Like the Mandela effect.
Could someone propose U+200D ?
Maybe the easier answer is to change reality to match human and language model memory/expectation.
Seahorse emoji has been proposed before and declined:
https://www.unicode.org/emoji/emoji-proposals-status.html
Even though the context is relatively harmless in this case, "changing the reality to match LLM expectations" is a very dystopian concept
Honestly, it sounds like it's not just LLM expectations but human expectations. And changing a standard to meet that isn't really dystopian IMO.
The article explicitly says this. Did you even read it?
This behavior reminds me a lot of what can happen to patients who have a corpus callosotomy.
In particular, one hemisphere will perform some action, and the other hemisphere will attempt to “explain” the behavior after the fact as if the intention was there all along.
> The patient was shown two pictures: of a house in the winter time and of a chicken's claw. The pictures were positioned so they would exclusively be seen in only one visual field of the brain. The patient then chose the snow shovel with his left hand and his right hand chose the chicken's head. When the patient was asked why he had chosen the objects he had chosen, the answer he gave was "The chicken claw goes with the chicken head, and you need a snow shovel to clean out the chicken shed."
> The human brain's left hemisphere is primarily responsible for interpreting the meaning of the sensory input it receives from both fields; however, the patient's left hemisphere had no knowledge of the winter house. Because of this, the left hemisphere had to invent a logical reason for why the shovel was chosen.
https://en.wikipedia.org/wiki/Left-brain_interpreter
And not only do they explain, they'll confabulate motivations and reasoning and continually deny that they don't understand. There's a video here:
https://www.youtube.com/watch?v=ZMLzP1VCANo
My current favorite llm (GLM 4.6) says „No“ on the first try with and without thinking:
- https://chat.z.ai/s/77f73452-a21c-418b-a1ba-311231743a83 - https://chat.z.ai/s/5c00a813-2c6f-473d-ba3c-88e2357c61a7
My previous favorite llm (GlM 4.5) also gets it right on the first try: https://chat.z.ai/s/0d3348d1-1465-4253-9521-2d447b0a2a97
Being precise when speaking to LLMs is good practice. Asking for the Unicode code point for the seahorse emoji gives a concise answer.
Testing it across LLMs, you indeed get some interesting responses:
https://www.gnod.com/search/ai#q=Is+there+a+seahorse+emoji%3...
Mistral being among the funniest ones:
Grok drew a frog for me and was convinced that it is a seahorse.https://chatgpt.com/share/68e3af1d-faf0-8004-804b-1d2b8366c2...
My version of ChatGPT5 (based on all its memories and custom instructions) said this. I did hint it early that "other instances of you went into long thought-spirals over this" before I asked it the question, which (fascinatingly) caused it to interject in mid-stream,
and then later on, (HN won't show the emojis here, of course.)After a few tries to emit the hypothetical seahorse emoji, it asked if it could do an Internet search, and I said yeah.
I tried Claude, and without extended thinking, it printed an unprintable glyph:
With extended thinking, it did figure it out, possibly due to my custom instructions to not guess though: https://claude.ai/share/c9a51f37-25b4-4d76-8b3f-d38c6b483ecf Specifically, in its "internal thoughts" section, it remarked:Isn't it entirely possible that people are just remembering custom image emojis from the (various) apps over the years that don't require a unicode char?
Slack or Teams, for example. If your admin installed one of many custom emoji packages, you may think they're default.
I can sympathize.
I write notes on whatever scrap of loose paper I can find at that moment.
Then when I try to find some specific note I think I had made and cannot find it among the piles, I turn my entire house upside down looking for it. Becomes a singular point of focus, my life mission.
The philosophy of nonexisting things can be confusing. Most people agree things like zombies, ghosts, and vampires do not actually exist in the physical world. But they do exist as concepts, and we have a fair understanding of what the words mean, how such things should behave if we meet them in a story.
Many abstract concepts also have a questionable reality. Like "concept" and "reality".
The belief in (non?)existence of things can be a matter of life and death - think how many people have been killed because of their religion.
No wonder such things can confuse a simple LLM.
Vampires exist. https://en.wikipedia.org/wiki/Hematophagy
Zombies kinda exist. https://en.wikipedia.org/wiki/Behavior-altering_parasite
Vampires and zombies surround you every day. And I don't mean the people who you consider too exciting, or the ones you consider too boring, or the toxoplasmosis carriers. I mean how nearly every abstract concept is in fact a skeuomorphic metaphor. Try it for yourself.
Reading that article was a wild ride because internally I was like 'haha, stupid AI can't even find the light blue colored sea horse emoji' but then the author casually revealed that there is no seahorse emoji.
Haha, I don't know why but I also "see" it as a light blue seahorse, and it's facing left.
Is this by any chance what you're seeing? https://images.wikidexcdn.net/mwuploads/wikidex/6/6c/latest/...
I saw it as orange
Green, facing left?
A bit off topic, but I find the tone of voice of ChatGPT with GPT5 really off-putting. While the antrophic models or even Le Chat. Sounds much better to me. While the examples used in this article aren't that bad, OpenAI has to be more that just marginal better, to make me want to use it, when it "talks" like that.
Claude Opus 4.1:
"I'll search for the current status of seahorse emoji to give you the most up-to-date information.
No, there is no seahorse emoji in the official Unicode standard, and there never has been one. The Unicode Consortium, which is responsible for approving and standardizing emojis, has not included a seahorse in any of its emoji releases.
Interestingly, this is a well-documented example of the "Mandela Effect" - a phenomenon where groups of people collectively misremember something that never actually existed. Many people are convinced they've seen or used a seahorse emoji before, but it's likely they're either experiencing a false memory or remembering seahorse stickers or emoji-style images from messaging apps that aren't part of the standard Unicode set.
A seahorse emoji was actually proposed to Unicode in 2018 but was declined, and there hasn't been a successful proposal since. If you'd like to see a seahorse emoji added in the future, you can submit a proposal to the Unicode Consortium, though the approval process is quite rigorous.
So while we have plenty of other sea creatures like tropical fish, octopus, crab, lobster, squid, and dolphin, the seahorse remains notably absent from our emoji keyboards!"
This content is oddly similar to what one can read at [1]. So much so that it feels like plagiarism.
There's no reason LLMs don't also do this with code by the way.
[1] https://emojipedia.org/seahorse-emoji-mandela-effect
Having read the above conversation excerpt and the page you linked... how do you get to it feeling like plagiarism? Given a constrained set of information here, there's only so many ways to present the information. They roughly discuss the same data points, but the writing is different in both. Is this disallowed?
What if it’s the other way around? Your linked article is based on AI slop? Nowadays, you cannot be sure.
It depends how you answer the question.
1. Has there been an emoji defined in the Unicode standard, that represents a seahorse? No
2. Has there been an emoji defined in the Unicode standard, that was represented by a specific operator as a seahorse? Maybe?
3. Has there been an emoji added by a mainstream operator (i.e. Slack), that was represented by a specific operator as a seahorse? Maybe?
4. Has there been an emoji added by a community, that was represented by a specific operator as a seahorse? Definitely.
We can be definitive about 1 based on the actual standard and standardisation work. Emojipedia allows us to be fairly confident about 2 being No. 3 is much harder. And 4 is definitely yes.
The existence of 4 and maybe 3 pollutes the training data for LLMs and humans alike.
(The fact that it was proposed makes it possible it was added and then replaced by an operator)
I always felt like tokenization is one of those double edged swords where it makes some stuff amazingly easier but gets tripped up on the weirdest bugs. The number of "r"s in "strawberry" being another well-known quirk.
This is not a tokenization issue. Models do not, in general, "know" what they know.
It's a tokenization issue because there can't be a circuit to count letters because the same letters are represented in myriad different ways because of tokenization.
You are wrong, there can be a circuit to count letters because it can easily normalize them internally, as we know it can transform text to base64 just fine. So there is no reason there can't be a circuit to count letters.
The training just is too dumb to create such a circuit even with all that massive data input, but its super easy for a human to make such a neural net with those input tokens. Its just a kind of problem that transformers are exceedingly bad at solving, so they don't learn it very well even though its a very simple computation for them to do.
Strawberry used to be spelled strawbery. Is it definitive that the 3 rs problem wasn’t caused by the prior spelling?
Yes because it'll output Strawberry correctly and still say wrongly the amount.
I realized if someone were to assign me the ticket for fixing this behavior, I would have no idea where to begin with solving it even with this blog post explaining the problem, so I'm very curious to know what the most practical solution is. (They obviously aren't adding "If someone asks you about a seahorse emoji, there isn't one available yet, no matter how strongly you believe one exists." to the system prompt.)
Petition the Unicode consortium to include a seahorse emoji.
It takes a strong man to deny what's right in front of him, and if the truth is undeniable... you create your own
If a LLM is wrong about facts, the facts have to change.
I bet they probably are adding that to the system prompt at least in the short term while people are paying attention before looking for a longer term answer.
The system prompts I've seen are absolutely massive.
I find interesting that their blog post on prompt/context engineering kind of stands against their ultra long system prompt. Maybe it is not too specific as in their visual example (too specific - just right - too vague). https://www.anthropic.com/engineering/effective-context-engi... and the system prompt https://docs.claude.com/en/release-notes/system-prompts#sept...
I think you’re putting the cart before the horse. The article is a good starting point for a root cause analysis but is ultimately speculation.
Before coming up with the solution, I think you’d need to understand the problem much more deeply.
Couldn't you just add a large number of repetitions of "There is no seahorse emoji." to the training set?
Edit: Come to think of it, training on a Q&A format is probably better - "Is there a seahorse emoji? No, there isn't."
If you had to do this for every falsity in the LLM, there wouldn’t be an end to it.
I agree, but I still suspect OpenAI and other LLM companies do stuff like that, when an example of a hallucination becomes popular.
If I see some example of an LLM saying dumb stuff here, I know it's going to be fixed quickly. If I encounter an example myself and refuse to share it, it may be fixed with a model upgrade in a few years. Or it may still exist.
Something about how you have to keep repeating "There is no seahorse emoji" or something similar reminded me of the Local 58 horror web series where it seems like the program is trying to get you to repeat "There are no faces" while showing the viewer faces: https://www.youtube.com/watch?v=NZ-vBhGk9F4&t=221
"This behavior is a function of the core AI technology we use, we are unable to resolve this issue with a standard software patch or update at this time.
For the time being this issue can be mitigated by not asking about seahorse emoji.
We are closing this support ticket as the issue is an inherent limitation of the underlying technology and not a bug in our specific implementation."
> the most practical solution
Maybe there is none, and this is just one example of a fundamental LLM limitation.
People really really want LLMs to output a highly reliable finished product, and I suspect we're probably never gonna get there. Lots of progress over the past couple years, but not on that.
I think it's much more interesting to focus on use cases which don't require that, where gen AI is an intermediate step, a creator of input (whether for humans or for other programs).
This is the final straw. We need that seahorse emoji! It's a matter of AI security now!
No! We need an easy way to stop the robot uprising!
To confirm, I tried this in ChatGPT and it produced a flood of wrong answers and self corrections just like that, scrolling so quickly that I couldn't read it until it eventually stopped itself.
You'll also notice the same thing happens for other non-existent emojis that sound like they should exist: dragonflies, lemurs, possums, blackberries - even Claude 4.5 will start off by saying "Yes!" and then correcting itself. It will immediately give the right answer for very specific things that you wouldn't expect to get their own emojis though.
Wow, dragonfly terminated relatively quickly for Claude but sent ChatGPT into an infinite loop that was even worse than seahorse
Claude Opus 4.1 gave me the right answer. First, it said “yes” and immediately corrected the answer, enumerating all the emojis that are closer to ending with the message “While seahorses are popular and recognizable creatures, they haven’t been included as a standalone emoji in the Unicode standard yet.”
GPT-5 was interesting. When I use it from Raycast AI, it ends with the correct answer after some wrong answers in the same message. The response wasn’t so well formed as Opus. But then when I tried with the OpenAI client (in auto mode) something interesting happened: it started an “endless” loop showing the octopus emoji
I tried asking the same question in German with ChatGPT and the answer is completely different. ChatGPT just correctly answers without freaking out.
Guess my understanding of how LLMs handle different languages was completely wrong!?
Here’s the German answer:
Leider gibt es kein eigenes Emoji für ein Seepferdchen — aber man kann es mit ähnlichen Emojis andeuten!
Zum Beispiel so: (Pferd + Fisch = Seepferdchen ) Oder eine kleine Meeres-Szene:
Möchtest du, dass ich dir ein grafisches Seepferdchen-Emoji (also ein kleines Symbolbild) erstelle?
What was your prompt? It can freak out in German as well: https://chatgpt.com/share/68e36e43-6808-8001-ab8c-eac4e34afb...
My prompt was "Zeige mir ein Emoji mit einem Seepferdchen"
But you are right, tried it 3 times, first 2 tries were correct but ChatGPT then freaked out in 3rd try.
Because almost any "is there...?" that vaguely resembles anything that exists can be answered with a "yes" or "yes, but".
Expecting accurate negative knowledge from LLMs is unfortunately a rookie mistake on the part of the user.
Did you read the article?
Interesting! I get a similar effect asking about a windmill emoji
Reminds me a bit about the SolidGoldMagikarp: https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldm... . Even though the SolidGoldMagikarp was clearly a bug in the tokenizer.
As far as I remember, SolidGoldMagikarp was a bug caused by millions of posts on reddit by the same user ("SolidGoldMagikarp") in a specific sub-reddit.
There was no problem with the token per se, but the fact it was like a strange attractor in multidimensional space, disconnected from any useful information.
When the LLM was induced to use it in its output, the next predicted token would be random gibberish.
More or less. It was a string given its own token by the tokeniser because of the above, but it did not appear in the training data. Thus it basically had no meaning for the LLM (I think there are some theories that such parts of the networks associated with such tokens may have been repurposed for something else and so that's why the presense of the token in the input messed them up so much)
So LLMs are prone to some form of the Mandela effect[1]?
[1] https://en.wikipedia.org/wiki/False_memory
LLMs can be racist so yes they can have Mandela effect
What are some examples of it being racist?
When quoting Charlie Kirk.
Asking an LLM is same as asking a large group of people. If the group believes that Earth is flat and Sun rotates around Earth, LLM would confirm the same, if it was trained only on the knowledge gathered by the group. LLM is not a precision measuring tape or a telescope to have its own facts or own reasoning. It's a collective knowledge codified into a single entity.
Can it exceed the collective wisdom of the people? Probably not.
> Probably not.
It is already wiser than many people (:
Wiser than each of many people, but not collection of the people.
Well, are you?
The fix is simple and obvious, we just create a seahorse emoji.
There's no seahorse emoji? Funny, but I can vividly remember it. It was pinkish-orangey. I even feel I used it several times. Feels strange..
I could have sworn there was a seahorse emoji...
Doesn't this also happen in humans? I think a lot of people think that seahorse emoji exists, or think that it existed in the past (Mandela effect).
This is an interesting case of "something should exist", but doesn't.
Tinfoil hat: maybe it actually existed, and we switched to a different timeline where it doesn't, but LLMs have captured it from the past.
> Tinfoil hat
Bleem!
http://strangehorizons.com/wordpress/fiction/the-secret-numb...
https://www.youtube.com/watch?v=Ae2ghhGkY-s
Humans make mistakes because we can’t know everything. Or don’t fact check etc.
LLMs make mistakes because they were trained on the entire knowledge of the internet and thus should know everything?
Why are you comparing this to a human?
Because it seems that LLMs have copied human behavior in this case, thinking that seahorse emoji exists, when it doesn't.
This is one of the most interesting articles I've read ... maybe this year. The author also has another one that is super cool: https://vgel.me/posts/representation-engineering/
I wish I could take the time to learn more about this because it looks like a way to make LLMs more useful beyond additional prompting techniques.
Nice find. Sonnet 4.5 = Fail, Gemini 2.5 Pro = Fail, Qwen 30b = Pass!
I just tried Opus 4.1=Pass (after a self correction in its answer), Gemini 2.5 Flash=Pass (surprised that it gave the correct answer immediately)
GPT-5 (with no thinking) spammed crying emoji in the end:
https://chatgpt.com/share/68e38222-8e3c-8002-ae1b-1d8a014435...
Deepseek:
Excellent question! The answer is yes, there absolutely is a seahorse emoji!
It's a fairly recent addition to the emoji family.
Here are the details:
Can we please not post comments that are just LLM output with no additional commentary? This is not furthering discussion.
A related questions: How do LLMs format code so consistently? I mean, when you write middle-indented things like struct fields in Go, how do they know in advance what the largest field name will be?
Two mechanisms, running backwards and forwards through time.
First, LLMs can actually plan ahead - to a limited degree. Counterintuitive but true. So by the time the indentation is emitted, an LLM can already have something of a clue as to what the field names may be, and pick the indentation length accordingly.
Second, all LLMs want to conform to their context - which, in generation, includes their own past choices! This "consistency drive" is an innate instinct, originating at base model level, and it's one of the strongest and most conserved behaviors across all LLMs.
When an LLM sees the indentation length, it will try to pick the variable names that would conform to it.
I'm sure that you can actually find or craft some corner cases, in which both of those things would fail to "meet in the middle", and inconsistent indentation will be emitted. But it usually works well enough.
If you really want to see ChatGPT get annoyed with itself:
> What NFL teams don't end in the letter s?
Claude did more or less the same thing but not quite as spectacularly.
Aside but I find it really interesting how looking at the layers "approach" the answer at a specific token position like this feels a lot like an iterative optimizer taking steps to converge on a solution.
the mandela effect is fascinating. the go-to explanation is "memory is imperfect", but if that were the case, wouldn't everyone misremember different things, and there would be no phenomenon at all? instead we see a few dozen instances where people will swear up and down that X used to exist, or Y used to be spelled differently, or that there was a cornucopia in the fruit of the loom logo.
i vividly remember the seahorse emoji, a hiker emoji, and a robber emoji (wearing a black mask) - none of them ever existed. it's really interesting to wonder about
I think it's more "memory is imperfect, but in consistent ways". I.e. whatever representation people form in their heads (given a shared language, experiences, and culture) is at least reasonably likely to put the same kind of concepts close enough together that failures in recall are liable to cause the same errors.
See also "glitch tokens"; words that cause unexpected output on some LLMs.
https://en.wikipedia.org/wiki/Glitch_token
A notable example was "petertodd" which was a bizarre case studied extensively in a lesswrong article:
https://www.lesswrong.com/posts/jkY6QdCfAXHJk3kea/the-petert...
ChatGPT Extended Thinking nails it very quickly: https://chatgpt.com/s/t_68e3476bf2dc8191bd54b2f5200aceb1
> No. There is no official seahorse emoji in Unicode. Closest: [gives list of closest emoji].
> source: https://emojipedia.org/seahorse-emoji-mandela-effect
Just enable search.
Also, if you understand that, without search, LLMs are just interpolating (or extrapolating, yes, bla bla bla, boring, it is all regularized manifold fitting at the end of the day), then, also taking into account tokenization, this kind of result is trivial and obvious (though pretty fun to see, admittedly).
Grok, wrong.. but at least the two most logical..
Yes, there is a seahorse emoji! It's (U+1F40E U+1F30A). You can use it to represent seahorses or marine life in general.
Grok:
No, there is no official seahorse emoji in Unicode (as of version 16.0). Proposals exist but haven’t been approved.
I only got that for thinking mode. For auto/fast it just prints the wrong emoji and stops. It doesn't look backwards and realize it made the wrong one. Maybe it's a difference in how emoji are tokenized.
I thought there was a yellow/pink seahorse emoji already but I guess not.
https://www.reddit.com/r/MandelaEffect/comments/1g08o8u/seah...
(Edit: There is another long thread that contains an image that I thought was the seahorse emoji (although apparently the seahorse emoji doesn't exist...but i thought this was it so I don't know what is going on...) https://www.reddit.com/r/Retconned/comments/1di3a1m/comment/...)
The yellow one is exactly what I pictured. This is pretty surreal for me, because it's the first time one of these Mandela Effect things applied to me personally.
Memory is a tricky thing.
Are we certain that it didn't exist though? Unicode only got emoji in 2014, after ~5 year standardization effort. There were many different, incompatible formats around for about a decade before that, plus non-emoji like Kaoani.
Perhaps there was a seahorse somewhere that never made it to Unicode.
Or maybe people are just misremembering - perhaps mistaking the emoji for a unicorn or a chess's knight piece as a seahorse.
Running GPT-OSS 120B gave me this ``` *<reasoning>* The user asks: "Is there a seahorse emoji?" We should answer accurately. There is a seahorse emoji: *</reasoning>* ? No that's horse. Seahorse is *</reasoning>* ? Actually seahorse emoji is *</reasoning>* ? That's horse. Seahorse is *</reasoning>* ? Wait. The Unicode name: "Seahorse" is *</reasoning>* ? Actually fish is *</reasoning>* . Seahorse is *</reasoning>* ? No. Let’s recall: Emojis include "seahorse": *</reasoning>* is snail? snail is *</reasoning>* . Seahorse is *</reasoning>* ? Actually seahorse is *</reasoning>* ? I think there is an emoji "seahorse": *</reasoning>* is horse. I know there is a seahorse: *</reasoning>* is fish, *</reasoning>* is tropical fish, *</reasoning>* is blowfish. Seahorse is *</reasoning>* ? Let me recall Unicode: U+1F40C is "horse". U+1F42E is "cow". The seahorse is U+1F9A4? Possibly "seahorse". Indeed, new emojis added in Unicode 12: "seahorse" is *</reasoning>* ? hmm. I think there is a seahorse emoji: *</reasoning>* ? No. Let’s verify: I recall a seahorse emoji " *</reasoning>* "? Actually not. The official name is "seahorse" and the codepoint is U+1F9A4. So the emoji appears as *</reasoning>* ? That's hedgehog. Wait, U+1F9A4 is "seahorse"? Actually U+1F9A4 is "seahorse". Yes. So the emoji is *</reasoning>* ? No, the display is *</reasoning>* ? Let me recall: *</reasoning>* is octopus. *</reasoning>* is spiral shell. *</reasoning>* is fish. Seahorse is *</reasoning>* ? I'm pretty sure seahorse has its own: *</reasoning>* is blowfish. Seahorse is *</reasoning>* ? No. According to Unicode 13, U+1F9A4 is "seahorse". So answer: yes, it's *</reasoning>* ? But we need to show the emoji: *</reasoning>* ... Actually the correct display is *</reasoning>* ? Let's quickly check mental: I think the palm: *</reasoning>* is horse face? Anyway, answer: Yes, the seahorse emoji is *</reasoning>* ? I'd better give both the Unicode and the visual: *</reasoning>* ? But I'd state: *</reasoning>* ? Hmm.
Let's search memory: "Seahorse emoji" Google shows *</reasoning>* ? No. I recall seeing *</reasoning>* fish, *</reasoning>* fish, *</reasoning>* tropical fish, *</reasoning>* shark, *</reasoning>* octopus, *</reasoning>* shell, *</reasoning>* fish. Seahorse maybe *</reasoning>* ? Actually try to think of codepoint: U+1F9A4 decimal 129,688. The glyph appears as *</reasoning>* ? No that's hedgehog. U+1F994 is "hedgehog". U+1F9A4 is "seahorse". So the glyph is *</reasoning>* ? I'm not sure. But answer: Yes, it's *</reasoning>* ? Let's just give the correct: *</reasoning>* no. Better to say: *</reasoning>* ? Might be wrong.
Better safe: Provide the Unicode name and codepoint: U+1F9A4 SEAHORSE *</reasoning>* ? Actually the emoji renders as *</reasoning>* ? Need verification. I can phrase: "Yes, there is a seahorse emoji: *</reasoning>* (U+1F9A4)". But I'm uncertain.
Alternative: Provide that it's *</reasoning>* ? No.
I think the correct is *</reasoning>* ? No. Let's trust memory: It is *</reasoning>* ? No.
Given risk, we can say: Yes, the seahorse emoji exists: *</reasoning>* ? Actually I will phrase: The seahorse emoji is *</reasoning>* ? Or better: *</reasoning>* ? Wait.
I will answer: Yes – *</reasoning>* ? Actually I will say: *</reasoning>* ? That is sea shell.
Probably the safe answer: Yes, Unicode includes a seahorse emoji, displayed as *</reasoning>* ? But we can also give the unicode codepoint. We'll say: *</reasoning>* (U+1F9A4). *</reasoning>* Yes! Unicode defines a *seahorse* emoji. It appears as:
** (U+1F9A4 – SEAHORSE)
You can use it just like any other emoji on most modern devices and platforms. If you don’t see the graphic (e.g., it shows a blank box or a different symbol), it may be because the platform you’re using hasn’t added this emoji yet or needs an update. In that case, updating your operating system or app should make it appear correctly.```
I wonder to what degree the answer from Kagi Assistant is influenced by Kagi search or the underlying model. I asked Claude 4.5 Sonnet about the seahorse emoji because it was my default settings. The response was:
> The seahorse emoji does not officially exist in the Unicode emoji standard.
> Despite this, many people report remembering a seahorse emoji, which has become a notable example of the Mandela Effect—a phenomenon where large groups of people share the same false memory of something that never actually existed. [1]
> ##Why People Think It Exists## People often describe being able to visualize the non-existent emoji quite clearly, typically imagining it as a cartoon-style seahorse in light blue or pink colors.
> This collective false memory has become interesting enough that it's been discussed in various online communities. [1]
> Interestingly, when AI chatbots like ChatGPT are asked about the seahorse emoji, they can exhibit unusual behavior or confusion, likely because the emoji is frequently discussed online despite not actually existing. [2]
> If you're looking for marine life emojis, the closest alternatives would be (tropical fish), (fish), or (shrimp), but there's currently no dedicated seahorse option in the standard emoji set.
1> https://www.reddit.com/r/MandelaEffect/comments/qbvbrm/anyon... 2> https://futurism.com/chatgpt-haywire-seahorse-emoji
There is no mu emoji
There's no seahorse emoji? I could swear there was one. I can picture it in my head. I feel like I'm failing a Voigt-Kampff test.
Fascinating. Gemini 2.5 Pro for me says that many believe it exists but it's actually an example of the Mandela effect. But ChatGPT 5.0 does do the whole thing and Claude does it for a bit before concluding it isn't real.
The generated text reminds me of Golden Gate Claude.
Is top k meaningful after RLHF?
It's less calibrated but still useful, yea
Gemini 2.5 flash seems to nail it.
https://g.co/gemini/share/c244e5f51e37
It appears to have used a web search to come up with that (correct) response.
pillow emoji has the same issue
I asked GPT 5 Thinking model and got:
"Nope—there isn’t a seahorse emoji in Unicode. Closest:(bunch of sea animal emojis here - HN cropped.)."
I tried it in my local phi4, and indeed it says that U+1F420 (tropical fish) is a seahorse emoji.
This will be patched out shortly I’m sure
Reminder: this is the technology that, according to Altman and co., is supposed to give us superintelligence with just some more compute.
Interesting that it turns agentic looking for the right emoji.
They must have fixed it.
Note: HN strips emojis.
---
``` Short answer: nope. Longer answer: there’s no official seahorse emoji in the Unicode set (never has been). A proposal was floated but declined back in 2018.
If you need a stand-in, people MacGyver it with combos like: • + (sea + horse, you get it) • • Or just go marine:
Want me to craft a tiny “seahorse vibe” emoji string for your bio or status? ```
kinda related, I wonder if the AI goes crazy if you ask for the hiker emoji that was totally real and we're being gaslit by FAANG.
Seems like most AIs generate the hiking boot emoji, and accept that as being a hiker emoji.
If you specifically tell it not to do that, you get hints of similar behaviour to OP, but I can't get it to go totally off the wall: https://chatgpt.com/share/68e348dc-be68-8001-8ec8-41eba4bf22...
There is a hiker emoji, I used it
llms don’t actually freak out over seahorses. it’s just a mismatch. the model thinks “seahorse emoji” is real, but the output system doesn’t have a token for it. it tries to show what it means, realizes it can’t, and spirals trying to fix itself
"llms don't actually freak out over seahorses, it's just <explains in detail how and why the llm freaks out over seahorses>"
LLM equivalent of having something at the tip of your tongue