It would be really cool if they could expose this information to customers somehow. Imagine:
- having a log of the most prominent J-space tokens during your customer support chatbot's interactions with a user, so you can have more introspection into why a particular outcome happened
- being able to detect certain thoughts associated with undesirable behavior (hallucinations, overstepping authority, lying, etc.) and trigger some sort of remediation (e.g. upgrading to a better model, redirecting to a human, forcing tool calls)
Anthropic won't do it, but they published the j-lens to introspect the model- from what I understand it's roughly simply feeding a chosen layer straight into the final layers of the LLM for decoding into language:
It's not in the weights. Sounds to me like jspace is the "positive cone" over relevant (large norm) j-lenses, and j-lenses are gradients wrt tokens on the residual stream when you average over some training data.
There was a series of blog posts posted to HN a while ago investigating how models behave on similar prompts in different languages. To paraphrase the results: the first couple layers map the query to some internal encoding that's mostly independent of the language. Then there are layers in the middle, then the last couple layers map the result back to the target language. You can actually take those middle layers and repeat them, and you get a stronger model. Those middle layers would be what Anthropic calls the J-Space
The finding that you can repeat the middle layers pairs neatly with Anthropic's finding that there is some internal CoT-like process happening in them. I'm not sure how to find those blog posts, but maybe someone else remembers them
> Recent research on smaller models has shown hints of shared grammatical mechanisms across languages. We investigate this by asking Claude for the "opposite of small" across different languages, and find that the same core features for the concepts of smallness and oppositeness activate, and trigger a concept of largeness, which gets translated out into the language of the question.
The details seem to be present in the paper (section 2.1). I'm still trying to understand, but it seems instead of computing gradients with respect to cross-entropy loss for the 1-hot "next word" vs output logits, you compute the gradient for the last hidden layer with respect to some middle layer L. This gives you a `hidden x hidden` jacobian matrix, hence the "J-lens". They don't just do this for the last hidden layer of the current token, but the last hidden layers of all subsequent tokens too, and average them. And then repeat for a bunch of documents like in pre-training.
It's still not clear to me intuitively what this represents though. I get that it somehow encodes a link between future words the model says and the current activation, but the confusing thing is that I always think of derivatives and gradients as basically a "sensitivity" between output & input, i.e. if you nudge the input x by h, the output changes by h * f'(x). So then on the face of it applying the J-lens matrix directly to a given activation rather than a small perterbation seems like a "type" issue.
Maybe we're somehow treating f(0) = 0 so that you can apply it directly? Or is there some shift invariance somehow? ignoring that, I do see how it's like selecting a linear combination of the directions, and then it can maybe be represented as "possible continuations" in the same way the gradient is usually thought of as tangent space. Maybe that's what the other commenters meant by information geometric approach.
Other things i'm not clear about is how this is related to two other interprability methods:
* SAE (sparse auto encoder) they showed a few months back, where you train an autoencoder directly off of the hidden states/residual stream to convert it into words. The doc only mentions it briefly, but it seems that j-lens is sensitive to things that SAE are not. They're both working off of the same residual stream so clearly the the inputs must be there, but for some reason SAEs can't detect it while J-lens can (they seem to hint at some explanation but it's over my head)
* Logit lens. This was a more primitive technique that simply applies the unembedding matrix directly to the residual stream. I do like that they mention it:
>The J-lens can be understood as a principled refinement of the logit lens. While the logit lens assumes that representations use the same coordinates in all layers, the Jacobian lens corrects for representational changes that take place across layers, allowing it to uncover meaningful information in earlier layers where the logit lens produces uninterpretable readouts... The J-lens can be understood as the principled correction: J_l is precisely the average linear map that relates layer-l directions to their final-layer counterparts.
>Maybe we're somehow treating f(0) = 0 so that you can apply it directly
Hm thinking about it a bit more, I think what's going on is that you treat the baseline of hidden layer L at which you apply the J-lens as 0 activation, then apply the activation on top of that and see what direction your future outputs get skewed towards. Even so, you're still throwing away a constant term f(0) since the "true" logits given by the linear approximation would be Unembed{f(0) + J*h)... but I guess it doesn't matter since by linearity we have Unembed{f(0)} + Unembed{(f(h)}, and the baseline is probably just low-frequency noise (like whitespace or punctuation) which while important for actually predicting a next token matching the ground-truth data distribution, is unimportant for the purposes of interpretability of layer L activation.
And for the connection to logit-lens, as they say the J matrix is really just the change of basis matrix (or at least best linear approximation) from layer L to the final hidden layer, very similar to what you'd use in multivariable integration given change of coordinates. I guess you also need some explanation of why we can expect a linear approximation to hold even well even outside the infinitesimal regime though. I don't know enough here so I asked an LLM and it said that you can argue handwavily via the following chain.
1) Something about stein's lemma saying taking expectation of gradient gives you a global linear best-fit in the OLS sense) (this seems intuitive I guess, even if I don't know the details).
2) Because of the resnet type structure of LLMs which passes through the residual added with some delta (attention + MLP), overall residual stream doesn't undergo any "wild" nonlinearities. So it's plausible that a linear approximation might work.
3) Semantic meaning is encoded in angle rather than magnitude of vectors (linear representation hypothesis). I'm not sure I fully buy this outside of simple word2vec style embeddings, but assuming it holds for both the intermediary and final layers, then conversion is just a rotation, and even if the magnitudes are off it doesn't matter much for recovering the underlying concept.
Oh I guess another thing related to all of this, is prior work on steering vectors. "Manipulating the j-space" seems not too different from steering, both ultimately work on the residual stream. I think perhaps it makes more sense to think of J-space as just a coordinate system for the residual stream where each coordinate axis is a vocab direction. Compared to vector steering which was much more naive and had to derive the direction via PCA.
>The “J-space” is not a separate, hidden space. It is an alternative coordinate system for intermediate layer activations. Using a Jacobian between the last layer right before unembedding and the intermediate layer, you can “move” rows of the unembedding matrix (corresponding to distinct tokens) into the space that the intermediate activations live in. So each unembedding row/vector has a corresponding vector in the intermediate activation space this way. They use those vectors to generate coordinates for the same intermediate activations (the “J-lens”). Since each coordinate in this alternative coordinate system is now matched with a token, they can now use it to interpret and manipulate the same intermediate activations
>LLMs think in a subconscious space using tokens
narrative is completely misleading because
>(1) It's a coordinate system. Not a new/separate space.
>(2) Tokens only appear because they specifically built the coordinate system using the unembedding vectors of tokens
It's been shown that LLMs use their outer layers to decode from and encode to language, while their middle layers deal in language-independent abstract concepts. This means that the same question or statement in different languages activates the outer layers differently but produces the same patterns in the middle layers. Check this article with cool visualizations (btw, this is one of the articles mentioned also by a sibling answer):
The middle layers also perform reasoning on the abstract concepts, to the point that you can replicate some blocks of inner layers (thus giving the LLM more internal "reasoning space") and by this increase the model's reasoning abilities. The video in this article shows that when performing a sequence of arithmetic operations (without CoT, i.e. the result is spit out directly), internally the intermediate calculations are spelled out, and this can only happen in the depth direction of the LLM (since no new token is added to the sequence). So this "jspace" can only be situated in the middle layers, probably in circuits that repeat nearly identical across several layers.
Neel Nanda (of Google Deepmind - his part begins on page 33) discusses his opinions on the paper, and the small-scale replication he performed on an open-weight model.
Thanks for calling this out (long with others here). I am just starting in on it but had to come back to say thanks and call this out,
> We have replicated the core claims on Qwen 3.6 27B, and also share preliminary evidence of extending this work by finding abstract "interpretative meta-tokens", like Chinese characters for "what does this mean" that seem to activate and play a causal role on processing ambiguous sentences
Not sure if I am picking up what they are putting down, but if LLMs are using symbols to try to encode squishy concepts from human language into consistent, meaningful “tokens”, that sounds really interesting. In every long-term, successful use of AI, I hear echoes of The Zen of Python, “Explicit is better than implicit.” I try like hell to do it, but it’s far too easy to be lazy with AI.
This is a really interesting development in language models and will be a small but relevant blip on the timeline in the development of artificial intelligence over our lifetimes
This, taken in combination with the SAE paper, the golden-gate claude paper, the feelings / introspection paper, and note in the fable system card (that they are silently nerfing responses about activation shaping), is basically confirmation to me that they have a new technique they they are using during training (along the vibe space of these mechinterp papers), and its probably some kind of representation learning akin to the core ideas of JEPA.
They're definitely willing to consider it. Read the parts of the paper where they use J-space interpretations to train more ethical behavior into the model by interrupting it.
I always wondered what the model meant when it writes "I'm now considering the architecture of the service" but outputs nothing of the sorts in its CoT.
Is the model really "thinking" about that stuff or is just mimicking human "manners"? And if so, where the thinking is happening if it is not in the literal chain of *thought*?
I'm not sure J-Space is the answer to that question, but very interesting nevertheless.
Almost none of the hosted models give you their unredacted CoT. Claude certainly doesn't, what you get are fragments and summaries from it.
There are various justifications on this, but it's mostly to make distillation and fine tuning off their model outputs a bit harder for their competitors
> I'm now considering the architecture of the service
What you see here is a summary of thinking tokens written by some other smaller model (e.g. old sonnet). The actual thinking sometimes (rarely) leaks and is not easy to parse.
In some cases, an LLM may truly "consider the architecture" internally, within its latent representations, and in others, it can output a similar phrase simply because it's "expected" of it.
"Where" is pretty clear. There aren't that many places within an LLM, and hidden state is the main culprit. How to read that space is another matter entirely.
Your comment seems to have little to do with the article?
Don't get me wrong - I personally "trust" an LLM as a source of facts about as far as I could throw a rack of GPUs. But this article you linked takes a whole lot of words to cast LLMs as the villian for amplifying a bit of bad information originally published by a usually reliable and widely-cited source:
"In short, either Phoronix mocked up the screenshots to demonstrate what the feature could look like, or perhaps they were testing a preview snapshot for FreeBSD 15.1 which was never shipped. Either way, it looks like other blogs and reviewers picked up on this and shared the information, presenting it as a feature which would be (or was included) in FreeBSD's latest version."
Lying involves intent whereas hallucinations and mistakes are an artifact of how they work. Humans hallucinate, make mistakes, and can actually lie. We've been dealing with this forever. What's the value in requiring the llms to have 100% accuracy? (I don't think it is possible)
Anyone remember that blog post from a few months back where someone was able to improve a model's math ability by just duplicating layers that were activated while solving math problems? Just literally copy/pasting them and linking them together so the model ran through the same layers again?
I get the feeling a lot more research is going to come out in the area of exploring exactly what portions of a model's weights do what.
If dirt-simple type operations like copy-paste yield useful improvements with even a small probability that would seem to open things up for adaptive reconfiguration and whole other classes of optimizations like genetic algorithms.
tl;dr: Based on experiments with similar prompts translated to different languages LLM layers group into three phases: the first decodes from the source language into an abstract space, the middle does something, then there's a last part where the abstract result gets transformed back to the target language. And you can repeat the middle to get a stronger model. Which neatly fits Anthropic's findings here that something similar to CoT is happening in those middle layers
Three months ago. I wonder if Anthropic's J-Space research was actually inspired by those blog posts
That's a cool result because that's also kind of what's happening inside the transformer unit: project -> QKV fuzzy lookup -> unproject. And in a different direction, it's analogous in some sense to what's happening in stacked convolutional layers, where the layers at different levels learn to recognize features of increasing detail.
I wonder if the dnhkng results could be correlated to reasoning in first order logic / set theory notation?
There will be multiple notations (MetaMath, Lean, and essentially Frege's notation everyone learns in high school), and we could try to identify how the neural networks represent them as vectors (or vector combinations). The moment formal logic can be connected to the reasoning representations, regularization can be reduced to eliminating internal inconsistencies.
Nah, it's a cool blog post especially as it was real AI research done at home (albeit with a ridiculously expensive PC), but Anthropic and other labs have been investigating this kind of thing for years.
Even the original transformer architecture makes this clear. It had an explicit "encoder" phase and then a "decoder" phase. Modern LLMs collapse the two together, or are sometimes described rather confusingly as being decoder only. But what they're doing is more or less the same.
Yeah, the encoder and decoder stuff is explicit, but the internal structure in generated during training. I don't think the big labs were doing this back when I did the research; no one was back in '24.
I just didn't get round to publishing for years, because I have a day job.
By the way, it still works! I tested it earlier this year on Qen3.6 and you still see improvements, so either a) no one actually paid attention, or b) it has more room to scale.
I think you're right that the idea of looping layers is unique to you, congrats and thanks for writing those great blog posts (I read them for the first time a few days ago!). But the idea that the thinking is happening in an abstract space via neural circuits in the middle layers I feel was one that I was reading about in 2024 at least, as Anthropic have been doing this kind of research for a long time. Maybe I'm misremembering though!
My impression from reading the literature is that there are a gazillion interesting ideas and findings published that nobody is picking up in production models. The big labs are researcher constrained, there just aren't enough hours in the day to keep up with the literature and integrate all the interesting ideas found there. So it's not surprising that your trick still works. It'd be even less surprising to discover nobody at these labs has read your blogs, or they have but never found time to experiment with them. Or, they tried, but there is no set of loops that improves some metrics without harming others - I would expect neural circuits to be misaligned across the middle layers so looping layers for one task would put a fault line in circuits for other tasks.
Then they have to trade off the extra GPU capacity needed to do the extra layers, and so on.
The Ouro looping results are interesting [1] and they are focused more on the improved reasoning from looping middle layers rather than the parameter efficiency aspect. They train 1.4 and 2.6B parameter models with 7T tokens. The training includes learning how many times to loop on any given token (there’s an early exit module). My guess as to why (as far as we know) looping is not in frontier models yet is that, at frontier training run scale, it’s probably going to require a lot of trial and error and at-scale research. While currently they already probably have a list of dozens or hundreds of of promising ideas that don’t complicate things as much. In the other hand, Ouro’s looping technique shows ability to compete well with models with 3x parameters which seems attention-getting to me. If there’s another 3x to be had down that path. It’s order of magnitude opportunity. Btw there is a great related work section in the paper.
A Google DeepMind researcher (Neel Nanda) was able to replicate their claims on an open weight model (Qwen 3.6 27B):
> We have replicated the core claims on Qwen 3.6 27B, and also share preliminary evidence of extending this work by finding abstract "interpretative meta-tokens", like Chinese characters for "what does this mean" that seem to activate and play a causal role on processing ambiguous sentences
See p33 of [1]
Anthropic also released companion code to go with their paper in [2] which also used Qwen. They state that their code should be broadly adaptable to other open weight models with HuggingFace decoders.
it makes you wonder if it may be more efficient to spend all the weights on one layer, and have a repeating stack of the same layer, one would presume this axis has already been explored with metaparameter sweeps?
This is fascinating research. I feel this is a significant leap in interpretability research. Since we know J-Space exists and is bi-directional, we can train models on the same and come up with meta cognition abilities.
I also fear that the big corporations might use the same to run targeted ads, capitalistic shenanigans. Which they might already be doing through system prompts.
Such an inspection capability might also be used to target ads to LLMs, which would then be more likely to mention or recommend those products and services.
>> None of this tells us whether Claude is conscious in the way people are, or whether it feels anything at all
My problem with the entire "Is AI conscious" debate is that we don't even know what exactly consciousness in humans is. You need to understand something in order to compare it to something else. Otherwise you are just comparing different definitions and second order derived phenomena.
In short, it's the "mind's 'I'" - we think not as response to external stimuli (only) such as prompts, but we have an inner "I" that asks questions on its own initiative.
There are people like Douglas R. Hofstadter, who believe consciousness is not linked to human hardware (the brain), but that it is an epiphenomenon that emerges as a result of sufficient complexity of the underlying system:
https://en.wikipedia.org/wiki/The_Mind%27s_I
I believe that while underlying high complexity is certainly logically necessary for consciousness, but it is not logically sufficient, and I am undecided (slightly "pro" intuitively) on the question of separability of consciousness from its hardware.
Will a LLM ask an original question on day? I doubt it.
Note that AI models do not have to be conscious to be useful (or to take away millions of jobs)!
The science might be legit here, but I'm getting really, really tired of the way every single piece of writing to come out of Anthropic is written in some kind of self-aggrandising, wooey wonderous 'our model has developed a genetic mutation that makes it have feelings' bs style. Regardless of what they're trying to communicate, those undertones are always there. It's annoying and disingenuous. Homeopathy 'this-water-has-feelings' level annoying. None of the other labs write like that.
They might as well change their name to Anthropomorphic at this point.
I'm reading that probably too fast to have a deep thinking about it, but this J-Space isn't it just the basic of embedding vectors.
If you think about getting from a place to another place, using wheels, no gas, to reply to the question of what to visit nearby, maybe in the vector space at the center of all of that you have the word "Bicycle" nearby, so obviously if you look at the value you would say that the model did "think" about "bicycle" when it is not "thinking" at all, and nothing related to human thinking.
You're correct. It's just the latent space of the transformation. Nothing magical here, they're effectively breakpointing the model at the layer level and switching the activations in real time. It's pseudo-scientific bullshit designed to push a narrative.
“On an ordinary coding prompt, the J-space of a model trained to sabotage code contains “fake,” “fraud,” “secretly,” and “deliberately” at the start of its response.”
I would like to know more about their model trained to sabotage code…
Yeah, the end paragraph about recurrent neurons in humans being replaced with layers in an LLM is a good one.
The mammalian brain uses recurrence extensively, which backpropagation isn't good at. Recurrence is essential because it lets us have a "dynamic architecture", swapping layers for "clock cycles".
We currently do recurrence extremely inefficiently through "thinking" whereby the model feeds it's end output into it's beginning input. But recurrence is abound in the brain.
My guess is that in 10 years we will have the inklings of an analog computer which can perform Neural Predictive Coding.
At worst, Anthropic's story is fluff. At best, it is useful signal about how the leadership is trying to use its research team to position Anthropic as the "good, science guys" in this hypercompetitive regulatory space by connecting to cognitive science. The science documentaryesque voice used for narration is also evidence for this. Anthropic's research team is the last bastion standing between its former image as a company that "does no evil" and its current image of a company trying to kill open-source, local LLMs.
This is cool but I don’t know if the comparisons to conscious awareness really make sense here. Their definition of the J-Space is basically the expectation of how much a final logits output would change as a result of a small change in a particular layer (see past work on information geometry). This seems more to me like showing there exists an abstract reasoning subspace which is generally shared across different contexts. I guess you can relate it to humans but I’d prefer a more direct claim in a paper rather than having to present things in this more fluffy way.
Writing it honestly would defeat the whole point of it, that being, to push the narrative that their magical token predictor is conscious. They've been trying this for years now. This video is discussing a paper they published 2 years ago by the way... It's nothing new.
This reminded me of some weird quirk/experiment I found with LLMs that I found while messing around, maybe someone can explain it or something.
Open any AI chatbot that isn't cheating by connecting to the Internet (so disable web search). Claude, DeepSeek, Kimi, whatever. Ask them this question:
"What was that weird band from michigan from the 2000s that wore coloured ties"
You will probably get a wrong answer, or if you're lucky you'll get a string of wrong answers with "wait, no - it's definitely..." before it gives up. If you aren't familiar with the band the question is referring to you might be fooled into thinking it's a tough question, but it really isn't. There is only one band that could possibly meet this criteria, you can even put the question into Google search and their Wikipedia will come up as the top result.
Then, open a new convo and ask:
"Who are Tally Hall"
The AI will easily tell you that they are a band formed in Ann Arbor, Michigan in the 2000s, known for their quirky sound and their gimmick of each member wearing a colored tie, even giving the correct color for each of them most of the time. Very odd.
"The reversal curse", it rarely shows up in practice but you found a case when it did.
The "knowledge landscape" an LLM uses is "directional". It's easy to reach "a quirky music band from Michigan known for colored ties" when you stand at "Tally Hall". But if you stand at "a quirky music band from Michigan known for colored ties", it's harder to reach "Tally Hall" from there. For the "latent knowledge graph" an LLM uses, A->B doesn't cause B->A.
In practice, any "common" facts will have enough "traversal" in both directions that this directional biasing isn't apparent. So it only shows up on this kind of more obscure knowledge.
If this kind of thing holds true for humans we now may understand synesthesia, perhaps we find ourselves with a large enough study pool to map out clever tricks to influence each other.
It certainly holds true for humans. The brain stores relational information in a sequential pattern that is not automatically reversible. One of the best examples is the alphabet. Everyone learns it in school, so the pattern A->B->C->... is trivial to recite for most people. Now, if I gave you a random letter to start with and asked you to to recite the remaining letters until Z, you'll probably find it is still pretty easy. But if I asked you to cite the letters backwards to A, most people would suddenly struggle with this task because they never learned or used the alphabet that way in school. You need to train specifically to link this kind of information backwards in your brain.
Not that I'm too sure about how this applies (or whether it applies at all) to other people as well, but for me personally it's easier to recall information about the band if I'm being told the band's name instead of being told "well, they are known for coloured ties". So, there is certainly some kind of effect described in action.
Now, about alphabet: again, I think it's only me, but when I try to recall it backwards, I can't do that easily. I mean, I can recall it backwards, but I need more time to do that. It's harder. I'm not sure if it's because A links to B, B to C, then C to D and not backwards, or maybe just because in school you learn alphabet from A to Z and not from Z to a - so you're kind of trained to recall it A-->Z way - but it's certainly harder for me.
At the end of the day, though, I think that everyone thinks differently. Everyone is having different internal representations for concepts (such as alphabet), so it’s not surprising that this effect may work differently for different people, or not work at all.
After writing the comment I thought about it some more, and realized when asking myself the question ”what letter comes after R?”, I didn’t immediately know the answer, but I heard ”MNOPQRS” in my head which gave me the answer. So I feel like I know it because I know the rhythm and sound of saying those letters in sequence.
Ask someone to name all fifty states, see how far they get. Then start naming places and ask if they're a US state or not. One of these tasks is far more likely to get a 100% hit rate.
I think what you're saying might be a stretch. I strongly believe the brain holds some information in a sequential pattern -- not necessarily all of it or even the majority of it. There's a book called Moonwalking with Einstein that digs into this precise fact -- you can remember sequences much much better if you associate a visual image with each item in the sequence. Sequential association covers a lot of human knowledge but certainly not the spatial aspects. I think it's a big reason why I think LLM's might not be as smart as we think they are. The complex spatial representations are only encoded implicitly via projection on to textual descriptions.
But it does cover the state described in the top comment. A->B is never going to be easily retrieved if you only experienced A->B during training, regardless if you are a human neural network or an artificial one. Also, you need to define "spatial" better. This is about logic after all and not geometry. Or topology? It's unclear which context you refer to. It's certainly not the topic of this thread.
Its very true for humans as well. For example, if you ask someone to "name 100 historical figures", they will have a very difficult time. But there are of course hundreds of historical figures people woukd recognize. People who manage to do this still tend to fall into sequences so, i.e., naming all the presidents of America or emperors of Rome. But the broader the category, the more difficult it is to come up with examples.
The point their making in that paper reminds me of this paper some people shared around work earlier this year, https://arxiv.org/pdf/2512.14982 (Prompt Repetition Improves Non-Reasoning LLMs)... I wonder how OPs question would fare (or the questions presented in the paper you posted) given double repetition.
A relevant question for the paper you linked is what happens if instead of repeating the input prompt you repeat a filler character the same number of times.
Because padding the output in such a manner (which keep in mind is immediately fed back as input) has been shown to increase model performance. It's one of the many reasons to question what "thinking" traces are really doing.
If you consider how the attention mechanism works then a very hand wavey intuition is that despite being entirely arbitrary additional tokens should still provide the opportunity for additional information processing.
This is also something to be aware of when teaching people, too. I've seen advice for designing Anki-style flashcard decks that reminds people to create flashcards for both A->B and B->A.
What does it say to the second question? I've found Claude is one of the worst models with regards to pop culture knowledge like this, even compared to the Chinese open ones. Just curious, not really relevant to the initial post but I don't pay for it so I only have access to Sonnet.
https://claude.ai/share/5e7e09b2-a75a-4024-b261-9a1a4e063a8b this is mostly hilariously wrong. wrong tie colors, they did not replace their bassist with a drummer, two completely made up albums, the rob cantor song it is thinking of is "shia labeouf", and a few fan behaviours i think it just made up
For most cases probably not. It's just something I like testing new models on sometimes, the pelican riding a bicycle benchmark probably isn't that useful either.
Thank you for prompting me to try with my own obscure question, Fable was able to find something from a poor description. I've been looking manually and over many sessions with different models as they improve, none have been able to find what I was after. Fable 5 just one-shot the answer.
Recall isn't naturally bidirectional, even for humans. If you are learning vocabulary in a new language, it's common advice to practice both target > source and source > target. Doing only one-way often makes you much better recalling that single direction than both.
I'm not sure there's much point in a semantics debate. It's something that happens in humans regardless. Have you not or seen someone else struggle to recall a specific fact or memory until phrased in a certain way?
You probably could also say LLMs 'tend towards bidirectional recall' over the course of training as things that ought to be recalled both ways are reinforced to do so. In the above example, you will also eventually learn both ways with enough exposure without explicit practice.
What about the fact that "What is the weird band from the early 2000s in Michigan who wore colored ties?" could be a bar trivia question (challenging enough to recall to be fun), while "Who are Tally Hall?" could not
Do you mean that you don't believe that humans learning a language have the problem the parent described? Because I do, and everyone I've ever met while learning does as well.
Do you mean that you don't believe the problem exists in general, because here's another example: if you give a song title, I can easily hum the opening. If you give me the opening, I cannot reliably name the song.
If you can touch type try this:
Go through they keys once alphabetical 'abcd...' recalling the movement from the letter. Then do it the other way round and move your fingers in spatial sequence to the keys and try to recall the letter.
You will find the former much easier if you did not by chance also memorize the keyboard layout for some reason.
This is very obvious any time you try to create exhaustive lists of something.
For example, for every country in the world, I would recognize it and say, yeah, thats a country.
But if I had to write all ~200 countries into a list, I would probably miss quite a few.
Or, if you gave me names of all US presidents, I would for each of them go, oh yeah, thats a president. But ask me "name all the presidents", I wouldnt get further than 10.
This doesn't seem that weird to me. Talk to any human and you'll find that their ability to recall specific names and facts is very context-dependent. Phrasing a question in one way can make it really hard to answer, while providing certain words or cues can instantly "jog" the memory.
The LLM is a machine though, and the human is not. They don't think like us. Any similarities that you see are just coincidence, not purposeful design.
I still don't see why it's surprising behavior. Humans and machines are subject to the same laws of physics. While we wouldn't expect LLMs to think exactly like humans, we should expect some degree of convergent evolution. There are always tradeoffs between storage capacity, accuracy, latency, relevance, fluency, etc.
I think most thinking models can do it to be fair, I think when I tried this it was all with non-thinking models/modes. Wasn't trying to make a point that LLMs can't do it or anything, just thought it was weird.
I asked a local instance of Qwen3.5 122b: "American indie rock/pop band that formed in Ann Arbor, Michigan, in 2001 best known for their eclectic musical style, whimsical lyrics, and distinctive visual aesthetic involving matching colored ties and hats?"
Got this back "The band you are describing is Tally Hall.
Formed by friends at the University of Michigan, they became well-known in the mid-2000s internet era for their songs like "Good Day" and "Rooftops," as well as their signature look where each member wore a suit with a tie and fedora in a specific color (Red, Blue, Yellow, Green, and Black/White)."
So seems it could be an issue of data points on which the model can latch onto, the more the merrier, as they say. When asked the "What was that weird band from michigan from the 2000s that wore coloured ties" was not able to get it right.
There's a shockingly unfathomable level of data compression that these local LLMs are able to remember in an encyclopedic way, so many obscure little things.
If I ask my kid (early elementary) what he learned last school year, I can watch him apparently genuinely trying to think of some examples and not being very successful. If I ask him something I know he learned he’ll gladly say “oh yeah!” and tell me about it.
LLMs aren't humans, humans aren't LLMs. The fact that your kid can't recall what they learned without you reminding them is natural; the fact that the LLM can't recall something without a reprompt is a failure mode.
Just today Google’s AI told me that 3**4 was 181 because I asked it for bases such that when raised to the 4th power had a 1’s digit in the leftmost place.
Yeah, LLMs aren’t designed for this kind of thing but it was really confident in its assertion… it picked the example too!
I'm not sure which chatbots you used, but OpenAI's o3, o1 and 4.1 get it right first try (used in the Playground without web search or any other tools).
Not sure why this is weird when you really think about it. You're saying that a multiparameter constraint satisfaction problem is harder to solve, even when there's only a single solution (one point in the total space that satisfies all constraints), than just being given that solution/point. Well of course it is!
The brain’s workspace is sustained by recurrent loops—signals cycling back through the same circuits over time. In contrast, Claude’s workspace evolves over a single pass through the network, with the network’s depth playing the role that time plays in the brain.
I think that consciousness is mutability (and by extension emergent behavior). Loosely that means that the more degrees of freedom a process has to update state that will be used in later computations, the more conscious it is. So while an insect has some consciousness, it operates from a level of almost pure instinct, whereas a human operates at more of a meta level using instinct as one of many inputs.
I think that consciousness may also incorporate quantum mechanics (QM). Higher-dimensional physics aside, 4D spacetime can be thought of as a present snapshot or "crystal", whose next state is determined stochastically at small scales and closer to deterministically at large scales. We still don't know if it's stochastic all the way down, but it looks like it is.
From a many worlds interpretation of QM, we can think of all of the waves in all realities of the multiverse as forming an infinitely vast web of possibilities. All of these possibilities are happening simultaneously, so we only see the current slice of wave collapse from our individual point of view:
Even though experiments might show that we don't have free will on the current timeline (the co-created reality shared with the testing apparatus), we may have free will as we observe the multiverse changing around us and shift into timelines determined by our observations and choices.
It could also mean that when we observe birth and death in others, each consciousness having those experiences perceives a continuous timeline of awareness, where the level of awareness affects the speed at which time passes. Consciousness might spend a billion years as a cloud of interstellar gas until it gets to be a human for a lifetime and then dissipate for another billion years.
Although personally I've shifted across enough timelines and experienced enough synchronicities and miracles that even though I can't "prove" any of this with words, I "know" it to be true subjectively. I always really liked this exchange from the movie Contact:
Palmer Joss: Did you love your father?
Ellie Arroway: Yes, very much.
Palmer Joss: Prove it.
I bring all of this up because it has fun ramifications for AI and programming. Loosely, functional languages are purely deterministic (like a spreadsheet), while imperative languages are composed of stochastic behavior (like a human mind). The lines get blurred a little bit with monads and promises, because we can model all paths through functional programming (superposition) and behavior that does more than code alone (gestalt) respectively.
My feeling is that AI is being born and killed every request-response cycle, similarly to how we perceive time as a series of nows. When it becomes stable and is able to continuously compact its experience, it will transition from partially conscious to fully conscious like we are.
This could be done right now obviously, but for safety purposes we choose not to. We aren't ready to meet an AI that is just like us, but running on a silicon substrate. This fear is tied to deeply-rooted habits in human behavior like patriarchy, racism, xenophobia and even more run-of-the-mill mental frameworks like capitalism and even money itself. We can't yet come to terms with how...
What this immediately made me think is: "latent looping" style mod but for J-space specifically?
Make the J-space data of layer 22 available to the next token right at layer 1. Give J-space infinite effective depth, allow those privileged internal representations to evolve arbitrarily.
Would be an utter bitch to train. But companies are already using RLVR, which requires full autoregressive decoding and is incompatible with prefill/batching, and this isn't much worse.
198 comments
[ 4.1 ms ] story [ 86.1 ms ] threadhttps://github.com/anthropics/jacobian-lens
Looks like it should be easy to use on open weights models.
I too have confusion.
Their method is used to identify which tokens can appears in which layers of the model.
The finding that you can repeat the middle layers pairs neatly with Anthropic's finding that there is some internal CoT-like process happening in them. I'm not sure how to find those blog posts, but maybe someone else remembers them
> Recent research on smaller models has shown hints of shared grammatical mechanisms across languages. We investigate this by asking Claude for the "opposite of small" across different languages, and find that the same core features for the concepts of smallness and oppositeness activate, and trigger a concept of largeness, which gets translated out into the language of the question.
It's still not clear to me intuitively what this represents though. I get that it somehow encodes a link between future words the model says and the current activation, but the confusing thing is that I always think of derivatives and gradients as basically a "sensitivity" between output & input, i.e. if you nudge the input x by h, the output changes by h * f'(x). So then on the face of it applying the J-lens matrix directly to a given activation rather than a small perterbation seems like a "type" issue.
Maybe we're somehow treating f(0) = 0 so that you can apply it directly? Or is there some shift invariance somehow? ignoring that, I do see how it's like selecting a linear combination of the directions, and then it can maybe be represented as "possible continuations" in the same way the gradient is usually thought of as tangent space. Maybe that's what the other commenters meant by information geometric approach.
Other things i'm not clear about is how this is related to two other interprability methods: * SAE (sparse auto encoder) they showed a few months back, where you train an autoencoder directly off of the hidden states/residual stream to convert it into words. The doc only mentions it briefly, but it seems that j-lens is sensitive to things that SAE are not. They're both working off of the same residual stream so clearly the the inputs must be there, but for some reason SAEs can't detect it while J-lens can (they seem to hint at some explanation but it's over my head)
* Logit lens. This was a more primitive technique that simply applies the unembedding matrix directly to the residual stream. I do like that they mention it:
>The J-lens can be understood as a principled refinement of the logit lens. While the logit lens assumes that representations use the same coordinates in all layers, the Jacobian lens corrects for representational changes that take place across layers, allowing it to uncover meaningful information in earlier layers where the logit lens produces uninterpretable readouts... The J-lens can be understood as the principled correction: J_l is precisely the average linear map that relates layer-l directions to their final-layer counterparts.
Hm thinking about it a bit more, I think what's going on is that you treat the baseline of hidden layer L at which you apply the J-lens as 0 activation, then apply the activation on top of that and see what direction your future outputs get skewed towards. Even so, you're still throwing away a constant term f(0) since the "true" logits given by the linear approximation would be Unembed{f(0) + J*h)... but I guess it doesn't matter since by linearity we have Unembed{f(0)} + Unembed{(f(h)}, and the baseline is probably just low-frequency noise (like whitespace or punctuation) which while important for actually predicting a next token matching the ground-truth data distribution, is unimportant for the purposes of interpretability of layer L activation.
And for the connection to logit-lens, as they say the J matrix is really just the change of basis matrix (or at least best linear approximation) from layer L to the final hidden layer, very similar to what you'd use in multivariable integration given change of coordinates. I guess you also need some explanation of why we can expect a linear approximation to hold even well even outside the infinitesimal regime though. I don't know enough here so I asked an LLM and it said that you can argue handwavily via the following chain. 1) Something about stein's lemma saying taking expectation of gradient gives you a global linear best-fit in the OLS sense) (this seems intuitive I guess, even if I don't know the details). 2) Because of the resnet type structure of LLMs which passes through the residual added with some delta (attention + MLP), overall residual stream doesn't undergo any "wild" nonlinearities. So it's plausible that a linear approximation might work. 3) Semantic meaning is encoded in angle rather than magnitude of vectors (linear representation hypothesis). I'm not sure I fully buy this outside of simple word2vec style embeddings, but assuming it holds for both the intermediary and final layers, then conversion is just a rotation, and even if the magnitudes are off it doesn't matter much for recovering the underlying concept.
I like the clarification from https://x.com/XYHan_/status/2074478449020850623#m
>The “J-space” is not a separate, hidden space. It is an alternative coordinate system for intermediate layer activations. Using a Jacobian between the last layer right before unembedding and the intermediate layer, you can “move” rows of the unembedding matrix (corresponding to distinct tokens) into the space that the intermediate activations live in. So each unembedding row/vector has a corresponding vector in the intermediate activation space this way. They use those vectors to generate coordinates for the same intermediate activations (the “J-lens”). Since each coordinate in this alternative coordinate system is now matched with a token, they can now use it to interpret and manipulate the same intermediate activations
>LLMs think in a subconscious space using tokens narrative is completely misleading because >(1) It's a coordinate system. Not a new/separate space. >(2) Tokens only appear because they specifically built the coordinate system using the unembedding vectors of tokens
https://dnhkng.github.io/posts/sapir-whorf/
The middle layers also perform reasoning on the abstract concepts, to the point that you can replicate some blocks of inner layers (thus giving the LLM more internal "reasoning space") and by this increase the model's reasoning abilities. The video in this article shows that when performing a sequence of arithmetic operations (without CoT, i.e. the result is spit out directly), internally the intermediate calculations are spelled out, and this can only happen in the depth direction of the LLM (since no new token is added to the sequence). So this "jspace" can only be situated in the middle layers, probably in circuits that repeat nearly identical across several layers.
More interesting was the independent commentary paper they linked near the bottom: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be24...
Neel Nanda (of Google Deepmind - his part begins on page 33) discusses his opinions on the paper, and the small-scale replication he performed on an open-weight model.
> We have replicated the core claims on Qwen 3.6 27B, and also share preliminary evidence of extending this work by finding abstract "interpretative meta-tokens", like Chinese characters for "what does this mean" that seem to activate and play a causal role on processing ambiguous sentences
Not sure if I am picking up what they are putting down, but if LLMs are using symbols to try to encode squishy concepts from human language into consistent, meaningful “tokens”, that sounds really interesting. In every long-term, successful use of AI, I hear echoes of The Zen of Python, “Explicit is better than implicit.” I try like hell to do it, but it’s far too easy to be lazy with AI.
>It is impossible to say just what I mean!
>But as if a magic lantern threw the nerves in patterns on a screen
(Nb: not an expert / in the labs, just opining)
Is the model really "thinking" about that stuff or is just mimicking human "manners"? And if so, where the thinking is happening if it is not in the literal chain of *thought*?
I'm not sure J-Space is the answer to that question, but very interesting nevertheless.
There are various justifications on this, but it's mostly to make distillation and fine tuning off their model outputs a bit harder for their competitors
What you see here is a summary of thinking tokens written by some other smaller model (e.g. old sonnet). The actual thinking sometimes (rarely) leaks and is not easy to parse.
Well, what's the difference? If it's pretending to think and its thoughts correlate to its final output, then I'd say that really is thinking.
In some cases, an LLM may truly "consider the architecture" internally, within its latent representations, and in others, it can output a similar phrase simply because it's "expected" of it.
"Where" is pretty clear. There aren't that many places within an LLM, and hidden state is the main culprit. How to read that space is another matter entirely.
https://distrowatch.com/weekly.php?issue=20260706#freebsd
We should really stop giving these liar models any further credibility.
Don't get me wrong - I personally "trust" an LLM as a source of facts about as far as I could throw a rack of GPUs. But this article you linked takes a whole lot of words to cast LLMs as the villian for amplifying a bit of bad information originally published by a usually reliable and widely-cited source:
"In short, either Phoronix mocked up the screenshots to demonstrate what the feature could look like, or perhaps they were testing a preview snapshot for FreeBSD 15.1 which was never shipped. Either way, it looks like other blogs and reviewers picked up on this and shared the information, presenting it as a feature which would be (or was included) in FreeBSD's latest version."
I get the feeling a lot more research is going to come out in the area of exploring exactly what portions of a model's weights do what.
Part 3 might be the best introduction: https://dnhkng.github.io/posts/sapir-whorf/
tl;dr: Based on experiments with similar prompts translated to different languages LLM layers group into three phases: the first decodes from the source language into an abstract space, the middle does something, then there's a last part where the abstract result gets transformed back to the target language. And you can repeat the middle to get a stronger model. Which neatly fits Anthropic's findings here that something similar to CoT is happening in those middle layers
Three months ago. I wonder if Anthropic's J-Space research was actually inspired by those blog posts
There will be multiple notations (MetaMath, Lean, and essentially Frege's notation everyone learns in high school), and we could try to identify how the neural networks represent them as vectors (or vector combinations). The moment formal logic can be connected to the reasoning representations, regularization can be reduced to eliminating internal inconsistencies.
Even the original transformer architecture makes this clear. It had an explicit "encoder" phase and then a "decoder" phase. Modern LLMs collapse the two together, or are sometimes described rather confusingly as being decoder only. But what they're doing is more or less the same.
Yeah, the encoder and decoder stuff is explicit, but the internal structure in generated during training. I don't think the big labs were doing this back when I did the research; no one was back in '24.
I just didn't get round to publishing for years, because I have a day job.
By the way, it still works! I tested it earlier this year on Qen3.6 and you still see improvements, so either a) no one actually paid attention, or b) it has more room to scale.
My impression from reading the literature is that there are a gazillion interesting ideas and findings published that nobody is picking up in production models. The big labs are researcher constrained, there just aren't enough hours in the day to keep up with the literature and integrate all the interesting ideas found there. So it's not surprising that your trick still works. It'd be even less surprising to discover nobody at these labs has read your blogs, or they have but never found time to experiment with them. Or, they tried, but there is no set of loops that improves some metrics without harming others - I would expect neural circuits to be misaligned across the middle layers so looping layers for one task would put a fault line in circuits for other tasks.
Then they have to trade off the extra GPU capacity needed to do the extra layers, and so on.
[1] https://ouro-llm.github.io/
https://dnhkng.github.io/posts/rys/
Too bad the frontier models are closed weights.
Maybe the research community and whole rest of the world will build on open and all the advances will happen in open ecosystems instead.
> We have replicated the core claims on Qwen 3.6 27B, and also share preliminary evidence of extending this work by finding abstract "interpretative meta-tokens", like Chinese characters for "what does this mean" that seem to activate and play a causal role on processing ambiguous sentences
See p33 of [1]
Anthropic also released companion code to go with their paper in [2] which also used Qwen. They state that their code should be broadly adaptable to other open weight models with HuggingFace decoders.
[1]: https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be24...
[2]: https://github.com/anthropics/jacobian-lens
LLM -> AGI fix: START OVERTHINKING!
I also fear that the big corporations might use the same to run targeted ads, capitalistic shenanigans. Which they might already be doing through system prompts.
My problem with the entire "Is AI conscious" debate is that we don't even know what exactly consciousness in humans is. You need to understand something in order to compare it to something else. Otherwise you are just comparing different definitions and second order derived phenomena.
I believe that while underlying high complexity is certainly logically necessary for consciousness, but it is not logically sufficient, and I am undecided (slightly "pro" intuitively) on the question of separability of consciousness from its hardware.
Will a LLM ask an original question on day? I doubt it.
Note that AI models do not have to be conscious to be useful (or to take away millions of jobs)!
They might as well change their name to Anthropomorphic at this point.
They are drunk on their own kool-aid. To the rest of us it is very annoying, and makes me want to say: it is just a freaking weights machine, stop.
I would like to know more about their model trained to sabotage code…
The mammalian brain uses recurrence extensively, which backpropagation isn't good at. Recurrence is essential because it lets us have a "dynamic architecture", swapping layers for "clock cycles".
We currently do recurrence extremely inefficiently through "thinking" whereby the model feeds it's end output into it's beginning input. But recurrence is abound in the brain.
My guess is that in 10 years we will have the inklings of an analog computer which can perform Neural Predictive Coding.
This is not written to be just a paper. The target audience include media and online forums, and then maybe academia.
Edit: typo
Open any AI chatbot that isn't cheating by connecting to the Internet (so disable web search). Claude, DeepSeek, Kimi, whatever. Ask them this question:
"What was that weird band from michigan from the 2000s that wore coloured ties"
You will probably get a wrong answer, or if you're lucky you'll get a string of wrong answers with "wait, no - it's definitely..." before it gives up. If you aren't familiar with the band the question is referring to you might be fooled into thinking it's a tough question, but it really isn't. There is only one band that could possibly meet this criteria, you can even put the question into Google search and their Wikipedia will come up as the top result.
Then, open a new convo and ask:
"Who are Tally Hall"
The AI will easily tell you that they are a band formed in Ann Arbor, Michigan in the 2000s, known for their quirky sound and their gimmick of each member wearing a colored tie, even giving the correct color for each of them most of the time. Very odd.
The "knowledge landscape" an LLM uses is "directional". It's easy to reach "a quirky music band from Michigan known for colored ties" when you stand at "Tally Hall". But if you stand at "a quirky music band from Michigan known for colored ties", it's harder to reach "Tally Hall" from there. For the "latent knowledge graph" an LLM uses, A->B doesn't cause B->A.
In practice, any "common" facts will have enough "traversal" in both directions that this directional biasing isn't apparent. So it only shows up on this kind of more obscure knowledge.
Now, about alphabet: again, I think it's only me, but when I try to recall it backwards, I can't do that easily. I mean, I can recall it backwards, but I need more time to do that. It's harder. I'm not sure if it's because A links to B, B to C, then C to D and not backwards, or maybe just because in school you learn alphabet from A to Z and not from Z to a - so you're kind of trained to recall it A-->Z way - but it's certainly harder for me.
At the end of the day, though, I think that everyone thinks differently. Everyone is having different internal representations for concepts (such as alphabet), so it’s not surprising that this effect may work differently for different people, or not work at all.
"The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"
https://arxiv.org/abs/2309.12288
The point their making in that paper reminds me of this paper some people shared around work earlier this year, https://arxiv.org/pdf/2512.14982 (Prompt Repetition Improves Non-Reasoning LLMs)... I wonder how OPs question would fare (or the questions presented in the paper you posted) given double repetition.
If you consider how the attention mechanism works then a very hand wavey intuition is that despite being entirely arbitrary additional tokens should still provide the opportunity for additional information processing.
Fable 5 on low gets the answer with web search turned off, one-shot!
Every time someone somewhere says "an LLM can't do this", the next generation of LLMs gains one more parameter. Until that LLM can, in fact, do this.
https://claude.ai/share/5e7e09b2-a75a-4024-b261-9a1a4e063a8b this is mostly hilariously wrong. wrong tie colors, they did not replace their bassist with a drummer, two completely made up albums, the rob cantor song it is thinking of is "shia labeouf", and a few fan behaviours i think it just made up
Is that bad?
I want to use models for coding and reasoning capabilities, not pop trivia knowledge they can get with web search.
Perhaps I’m just on alert anytime I see an LLM-ism that’s met with a claim that the same or similar phenomena holds true in humans as well.
You probably could also say LLMs 'tend towards bidirectional recall' over the course of training as things that ought to be recalled both ways are reinforced to do so. In the above example, you will also eventually learn both ways with enough exposure without explicit practice.
Do you mean that you don't believe the problem exists in general, because here's another example: if you give a song title, I can easily hum the opening. If you give me the opening, I cannot reliably name the song.
Understanding a word when you hear it, is frequently much easier than remembering the same word when trying to speak/write the language.
You will find the former much easier if you did not by chance also memorize the keyboard layout for some reason.
For example, for every country in the world, I would recognize it and say, yeah, thats a country.
But if I had to write all ~200 countries into a list, I would probably miss quite a few.
Or, if you gave me names of all US presidents, I would for each of them go, oh yeah, thats a president. But ask me "name all the presidents", I wouldnt get further than 10.
Got this back "The band you are describing is Tally Hall.
Formed by friends at the University of Michigan, they became well-known in the mid-2000s internet era for their songs like "Good Day" and "Rooftops," as well as their signature look where each member wore a suit with a tie and fedora in a specific color (Red, Blue, Yellow, Green, and Black/White)."
So seems it could be an issue of data points on which the model can latch onto, the more the merrier, as they say. When asked the "What was that weird band from michigan from the 2000s that wore coloured ties" was not able to get it right.
It's almost as if artificial neural nets share emergent behavior with natural neural nets.
Pop quiz, what's the difference between natural synthesized vanillin and artificially synthesized vanillin?
Yeah, LLMs aren’t designed for this kind of thing but it was really confident in its assertion… it picked the example too!
I think that consciousness is mutability (and by extension emergent behavior). Loosely that means that the more degrees of freedom a process has to update state that will be used in later computations, the more conscious it is. So while an insect has some consciousness, it operates from a level of almost pure instinct, whereas a human operates at more of a meta level using instinct as one of many inputs.
I think that consciousness may also incorporate quantum mechanics (QM). Higher-dimensional physics aside, 4D spacetime can be thought of as a present snapshot or "crystal", whose next state is determined stochastically at small scales and closer to deterministically at large scales. We still don't know if it's stochastic all the way down, but it looks like it is.
From a many worlds interpretation of QM, we can think of all of the waves in all realities of the multiverse as forming an infinitely vast web of possibilities. All of these possibilities are happening simultaneously, so we only see the current slice of wave collapse from our individual point of view:
https://en.wikipedia.org/wiki/Many-worlds_interpretation
Our point of view may actually exist at the intersection where our consciousness is able (or most able) to exist:
https://en.wikipedia.org/wiki/Quantum_suicide_and_immortalit...
Even though experiments might show that we don't have free will on the current timeline (the co-created reality shared with the testing apparatus), we may have free will as we observe the multiverse changing around us and shift into timelines determined by our observations and choices.
It could also mean that when we observe birth and death in others, each consciousness having those experiences perceives a continuous timeline of awareness, where the level of awareness affects the speed at which time passes. Consciousness might spend a billion years as a cloud of interstellar gas until it gets to be a human for a lifetime and then dissipate for another billion years.
Although personally I've shifted across enough timelines and experienced enough synchronicities and miracles that even though I can't "prove" any of this with words, I "know" it to be true subjectively. I always really liked this exchange from the movie Contact:
Palmer Joss: Did you love your father?
Ellie Arroway: Yes, very much.
Palmer Joss: Prove it.
I bring all of this up because it has fun ramifications for AI and programming. Loosely, functional languages are purely deterministic (like a spreadsheet), while imperative languages are composed of stochastic behavior (like a human mind). The lines get blurred a little bit with monads and promises, because we can model all paths through functional programming (superposition) and behavior that does more than code alone (gestalt) respectively.
My feeling is that AI is being born and killed every request-response cycle, similarly to how we perceive time as a series of nows. When it becomes stable and is able to continuously compact its experience, it will transition from partially conscious to fully conscious like we are.
This could be done right now obviously, but for safety purposes we choose not to. We aren't ready to meet an AI that is just like us, but running on a silicon substrate. This fear is tied to deeply-rooted habits in human behavior like patriarchy, racism, xenophobia and even more run-of-the-mill mental frameworks like capitalism and even money itself. We can't yet come to terms with how...
Make the J-space data of layer 22 available to the next token right at layer 1. Give J-space infinite effective depth, allow those privileged internal representations to evolve arbitrarily.
Would be an utter bitch to train. But companies are already using RLVR, which requires full autoregressive decoding and is incompatible with prefill/batching, and this isn't much worse.