The Short Version #
I think I have good news about the nature of neural networks.
After many months studying neurons and circuits in image and language models, I’ve come to believe a network’s own mechanisms are its legible building blocks, and that the outlook for mechanistic interpretability may be far better than I’d feared. The canonical evidence of polysemanticity and superposition that convinced much of the field, including me, that networks pack concepts arbitrarily was misread. What looked like arbitrary packing was mostly a motif I’ll call recursive generalization, which is much less concerning. This means we can study networks directly, instead of first untangling them. Neural networks are already clean.
There are at least two ways to approach mechanistic interpretability. The first is rapid normal science on the network itself: look at a network’s neurons and circuits, study how they relate, watch data flow through them, and run experiments at agentic scale to test explanations. The second starts from a list of hypothesized features and tries to extract them all, designing increasingly sophisticated reconstruction objectives and usually ignoring the original.
The field began with the first approach and fell back to the second because neural networks seemed to be tangled knots. We had canonical examples of polysemantic neurons, the broad name for a neuron that appears to have multiple unrelated jobs or that looks impossible to understand. The most famous one, in an image model, responded to both cat faces and the fronts of cars. And the last layer of the early field’s model organism, InceptionV1, seemed so full of arbitrary mixtures it wasn’t clear how to study it: almost none of the neurons were easily interpretable.
The leading explanation for polysemanticity was superposition. This proposed a new learning principle: when a neural network runs out of neurons, it stops grouping related things together and instead begins folding unrelated things together, simply because they don’t interfere. The result is a kind of knot: the network could have folded in many almost equally good ways, so which things a neuron groups together tells you almost nothing. You have to unfold the knot before you can study anything. In practice that means training a second model – a decomposition such as a sparse autoencoder – to approximate the network as it would look unfolded and studying that instead. If the network really is knotted, there may be no alternative.
One main result is that superposition was the wrong explanation for InceptionV1. Architectural problems made later layers overflow. Weak interpretability tools amplified these problems, making the late layers seem incoherent. Since then, architectural changes designed to help deep layers learn better have had the unintended effect of making them more interpretable. When we first studied InceptionV1, we found almost no interpretable neurons in its last layer. But with these ordinary improvements, and no change to model size or training data, 99-100% of neurons in that last layer are easily interpretable and the model is broadly legible the whole way through. It’s the first non-trivial neural network I feel I mostly understand; you can browse it here. I trained versions of it across a 300× range in size, and it’s broadly legible at every scale. With less room, the network doesn’t fold unrelated things together; its neurons just cover coarser-grained concepts, like one car neuron covering many kinds of car. With more room, that neuron differentiates into many specific cars. So the problem wasn’t the presence of a second learning principle, it was just an ordinary learning failure of the first one, and it’s easy to fix.
The canonical vision examples of polysemanticity and superposition were errors too. In 2019, Chris Olah and I found the circuit that was originally introduced as the superposition phenomenon in Zoom InZoom In: An Introduction to Circuits
Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. References ↓. It showed a car feature appearing to spread into a set of unrelated dog neurons. It was the only one we showed, and the only one I ever saw. We read it as the network placing the car in a high-dimensional hiding place, for downstream neurons to read back out. But nothing reads that combination. It wasn’t hiding anything; it was ordinary generalization and composition. To my knowledge, no one has published a second circuit with the right shape.
Modern language models are also far more interpretable than generally believed. Anthropic’s Towards MonosemanticityTowards Monosemanticity: Decomposing Language Models With Dictionary Learning
Bricken, T., Templeton, A., Batson, J., et al., 2023. Transformer Circuits Thread. References ↓ (2023), studying a small one-layer model, gave the median neuron a 0 on its rubric out of a possible 14, meaning the rater couldn’t form any hypothesis. The automated interpretability projectLanguage Models Can Explain Neurons in Language Models
Bills, S., Cammarata, N., Mossing, D., et al., 2023. OpenAI. References ↓ I started at OpenAI (also 2023) found about 1.7% of neurons interpretable by our most generous measure.
With better tools and modern models, the situation looks quite different. In preliminary audits of Qwen3-32B, we estimate that roughly 90% of neuron poles have one main job, often with small related offshoots; another ~7% have two related jobs; and only ~2-4% are not yet understoodA note on this number: the share of neurons with one job varies depending on how you set the threshold for counting weak activations, depth, density of sub-circuits, and what you view as “one thing”.
How far to go into weak activations is a big factor here. There’s a section describing weak activations below. It tends to increase how many “related things” there are, more than whether the neuron is a judgement of something legible. For instance, the Trump neuron’s weak activations bring in MAGA hats and people related to him. The Titanic neuron’s weak activations bring in stories of more general voyages, or steam engines and other technologies of the era.
There’s also the question of whether to care about activation percentiles, or fraction of the strongest activations. In general I think what the neuron learns should be trusted, its natural sparsity and activation ranges and so on. Which I used to mostly not believe! I’m somewhat against trying to standardize things across all neurons. Percentiles are more common today, I think because they’re easier to think about.
It goes the other way too. For instance, when you look at the Titanic neuron’s “unrelated jobs,” it’s often detecting that “boat” comes next. The tool missed this because it only looked backward when judging context. When we look at these “unrelated jobs” in depth, it’s almost always this kind of thing. It’s easy to build an imperfect tool.
Overall, I don’t think coming to a single score is actually that important. But when it was taken out, early reviewers felt unanchored without even a rough number. In the sets I’ve looked at more closely, considering strong activations, it’s closer to 97% non-arbitrary mixtures, but these are relatively small samples.
Overall distilling everything into one “cleanliness” score should be taken as a broad pointer to the regime a neural network is in, rather than something precise to either optimize directly or hold too tightly. A lot of this essay is for being less dimensionality reduced, treating each neuron and circuit differently, according to their motifs, depth, role etc. Despite this, a rough number can be useful.. Dense and Mixture of Experts models seem roughly similar. We found similar results in models up to trillions of parameters, like Kimi K3, which the companion paper will cover in more detail. These percentages compress a high-dimensional picture into a few numbers, and the precise number depends on where you set thresholds for what counts as “one thing” – like whether you lump together closely related idioms in two languages. I don’t think much in interpretability hangs on that question. What matters a lot more is why things share a neuron in the first place.
Most of the remaining apparent polysemanticity seems to be a motif I call recursive generalization. It’s simple: as a neuron grows, it sprouts offshoots the way a tree branch sprouts sub-branches. For instance, one neuron that detected “iy” grew to detect “diy,” and then “do it yourself,” which doesn’t contain “iy” at all. This idea first appeared in a forgotten subsection of a paper I worked on called Multimodal NeuronsMultimodal Neurons in Artificial Neural Networks
Goh, G., Cammarata, N., Voss, C., et al., 2021. Distill. References ↓ (2021). At the time, I hadn’t watched neurons grow over training. Now it’s easier, and it holds up everywhere I’ve seen. Even the original InceptionV1 wasn’t knotted. When I retrain its original architecture, which still produces messy deep layers, and follow the least legible neurons I can find, they turn out to be overgrowing locally into reasonable, adjacent territory.
So there were two explanations for polysemanticity, and I helped propose both. In some ways, they’re opposites: generalization groups similar things together and superposition groups unrelated things together. Superposition is the one the field bet on and, until recently, so did I. But I now think recursive generalization was basically right. That leaves one learning principle after all. It’s still generalization, but it turns out generalization is fractal. That’s far less concerning: there’s something to learn from the way things were grouped, and nothing to learn from two items in superposition. And since the generalization itself is fractal, our explanations can be too: start with a basic explanation, then zoom in when you want to understand more. In that sense, almost the entire network is interpretable if we look closely enough, especially if we look at how it grew over training.
Why was the gap so large? I think several ordinary problems compounded. The explainer models of 2023 couldn’t run the loop of conjecture and test; the tools and metrics were weak, rewarding explanations that could simulate activations rather than ones that were right, which often come apart; no one knew what language neuron motifs to expect, and people broadly looked for the wrong kinds of things; and most studies used small or simplified models. In a sense, those models really had run out of room, which made superposition the natural explanation. But a network short on room doesn’t switch learning principles; its neurons just get broader, as in the small vision models. Broad neurons are usually easy to break down once you know what to look for. Without that knowledge, and with primitive tools, they look like arbitrary mixtures.
Interpretability has a surprising amount of momentum: like a jigsaw puzzle or crossword, any understanding you gain starts constraining the rest of the system. At first, you’re looking at a single neuron. If you can’t understand it, little else helps: circuits and attribution graphs only explain a neuron in terms of other neurons you don’t understand yet. This can make getting started hard. Once the map starts to fill in, though, it goes surprisingly well. If you don’t understand exactly what a neuron does, you can study what it connects to, the ecosystem around it, and – in a mixture-of-experts model – what its expert specializes in. As with a puzzle coming together, every piece comes with many clues, and they’re well structured and easy to run experiments against. That’s why nearly empty and mostly filled in aren’t surprising places to be; they’re the natural anchors.
Take a neuron that tracks ‘a sense of righteousness when a bargain purchase turns out to work better than the expensive alternative’. Neurons that track a feeling or social cue like this are some of the most common we’ve found. Imagine finding this one with no knowledge of language neuron motifs and no circuits to lean on, because few neighboring neurons are explained yet. Most humans and agents expect neurons to detect nouns by firing on them, which they mostly don’t. Look at where this one fires most strongly, and you see a pile of seemingly unrelated commas.
Now imagine the map is mostly filled in. You know many neurons detect subtle feelings or social contexts, and that they often fire not on the words that carry the feeling but on punctuation, like commas, where the model summarizes what a sentence meant, or on the word just before the feeling matters for predicting what comes next. So you read the surrounding sentences instead, and check its steering effects (what the model writes when you turn the neuron up), which line up closely. You look upstream at how it’s built and downstream at what it does, and both line up. You check its negative pole, what pushes it strongly negative, and find a subtle variant of the same cue, or the moment the social context ends. You can run specific tests, like checking which store brands it responds to. Any of these could break the explanation, yet they keep pointing to the same place. And understanding this neuron sharpens the broader motif: after thousands of subtle-feeling and social-context detectors, you start to know what to expect from the next one, which is part of why the map fills in faster as it goes.
It works on the tools too. When you can see 2% of a network, you can’t tell where your tools are broken; when you can see most of it, the unexplained remainder points straight at them. That’s how some coworkers and I found a double-negative regime in SwiGLU neurons that, as far as we know, no one had noticed, and that a decomposition mixing many neurons together would likely have hidden.
The bad news about modern neural networks is that they're enormous. The good news is that the enormity helps: as the map fills in, every explanation is tightly constrained by everything around it, and bigger models learn narrower neurons that are easier to recognize in the first place.
In this sense, understanding the whole network may be easier than understanding just the parts we care about most. First, it’s hard to know how to think like a neural network until you’ve studied it closely. For instance, “deception” is a useful initial frame, but the building blocks that contribute to deceptive behavior by models probably have alignment-relevant microstructure that will be hard to know without going in and looking. Second, targeting a small set of mechanisms is like working on one row of an otherwise empty crossword. It’s doable, and might work, but without the crossings it’s hard to know whether you’re right. It may be easier to solve the whole thing – or at least the broader parts of the network these mechanisms connect to.
When I ran this process on modern convolutional networks, it basically worked. Agents can map out neurons and circuits, build explorable microscopes and explore them, test explanations thoroughly, and track hundreds of thousands of attribution graphs. There was never a point where rapid normal science broke down. I now feel I basically understand these models, at least enough to know where I’d start asking any question I had; eventually, the crossword just got filled in. So far it has worked in language models too, up to trillions of parameters. Those aren’t close to solved due to sheer size, but I’m optimistic about scaling the process.
All of this has brought me back to an old hope: understanding a neural network directly, by studying its mechanisms and the data flowing between them, without first trying to fix or untangle it. Put together, the argument has four steps.
- Networks use one learning principle, not two (if not everywhere, then nearly enough that it doesn’t matter for study), so there’s no blocker to beginning ordinary science on them.
- With modern models and tools, that science goes much better than most people expect.
- Understanding compounds, so it gets easier as the map fills in. At minimum, then, a network’s own mechanisms should be part of the toolkit.
- Highly speculative: this might be most of what we need, and interpretability might be far more tractable, possibly in the near term, than we’ve assumed.
That’s been the pattern for deep learning in general: the simple thing at absurd scale beats the clever one. Rich Sutton called it the bitter lessonThe Bitter Lesson
Sutton, R., 2019. Incomplete Ideas. References ↓. Maybe there’s a bitter lesson for interpretability too: scaling ordinary investigation beats engineering a clever new decomposition.
A technical companion paper will offer a more detailed account of language model neuron experiments, such as what we’ve found about the different regimes of modern non-linearities and what kinds of circuit motifs show up at different depths. This essay glances at many places and conjectures about where it all might be pointing.
Cars and Animals #
The most famous polysemantic neuron is 4e:55 of InceptionV1, which looked like an arbitrary mixture of cats and cars. It’s been brought up as evidence that neurons might not be the right unit of study in several of the field’s most widely read papers, from Feature VisualizationFeature Visualization
Olah, C., Mordvintsev, A. and Schubert, L., 2017. Distill. References ↓ in 2017 to Zoom InZoom In: An Introduction to Circuits Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. References ↓. In Zoom In, which I co-authored, we said outright that it “isn’t responding to some commonality of cars and cat faces”.In Multimodal Neurons we came very close. We asked whether the neuron might be responding to some shared feature, “perhaps cat eyes and car lights look similar,” then decided from the feature visualization’s whiskers and chrome that it was “genuinely OR(cat, car).” We were right to look for a shared shape, but I think we underestimated how far generalization reaches. The shared feature was the V. I think the whiskers misled us because generalization happens inside high-level patterns. The V neuron really does take in cat circuitry, and feature visualization’s job is to raise the activation, so once it finds the cat route it adds everything that route brings. The whiskers come along for the ride, even if the neuron doesn’t “want” them. Downstream, the car readers do get a little cat signal through the V neuron, but it barely reaches the output: one cat detector firing does little on its own, since many work together, so the network never needs a negative weight to cancel it. To see the shared pattern, you have to visualize what the neuron’s inputs have in common, not the neuron on its own.
After months of seeing clean neurons, I wondered if we’d misinterpreted some early prominent examples of polysemanticity and superposition. The canonical polysemantic neuron seemed like the obvious place to check, so I opened up the cat and car image grid I’d seen hundreds of times before and did something I’d never done: I spent a few seconds wondering what one thing this might be. Almost instantly it was obvious: it’s not a cat and car neuron, it’s a V-shape neuron.
I looked at its circuit connections: V-shaped flames, V-shaped water jets from a fireboat, V-shaped fox faces. I looked at its negative pole: an O, round where the V is pointed. Why would an arbitrary mixture of cats and cars need to connect so strongly to a flame?
Then I went to the canonical superposition circuit I first found in 2019I found it while building my first Circuits interface, which only had feature visualization and weights connecting their neurons. Then I took them to Chris Olah and asked why the car would be going into these dogs like this. We came to the conclusion it must be doing a kind of folding, and called it superposition. It was an intuitive explanation for what we saw, so it didn’t prompt diving more into the tools, which would have revealed feature visualization reward hacking and various other things. It being placed into a high-dimensional space felt like a reasonable answer.. The V neuron and this circuit were Zoom InZoom In: An Introduction to Circuits
Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. References ↓’s two main examples of why its picture of neurons, that nearly every one turns out to be understandable if you look seriously, might be “overly rosy”.
The circuit appeared to show a car going into a bunch of unrelated dogs, as if the dogs were a kind of high-dimensional storage place where the car could be recovered later as needed.
But that’s not what the car was doing. Cars and dogs were being brought together because of shared pose. InceptionV1 has more than a hundred of these generalized pose neurons, each one its own clean detector of a pose shared across cars, dogs and other objects, and to my knowledge no one had noticed them. Later layers keep building on them, all the way to the end of the model. The car was not being stored in an obscure place, and nothing directly reads out the combination. It’s the opposite: it’s continuing to differentiate into car specialists and finer poses.
Part of the reason was not knowing about generalized pose neurons, and part of the reason is that the technique we use to render icons of images — feature visualization — ends up being reward hacked in deeper layers. The intent is to find an image that maximizes the neuron, but the way that’s done is through gradient descent, and it takes the easiest path, which in InceptionV1 often means dogs. There are lots of dogs in InceptionV1’s training data and a lot of dog neurons, and the early optimization falls into them and never recovers. If you even lightly bias it toward the natural paths the top images take, most of them are cars, and some are both dogs and cars, because of shared pose.
It can be tempting to discount this generalization, under the assumption that the dogs and cars are only together because there’s not enough space. Maybe it “really wanted” separate dog-pose and car-pose neurons, and a bigger model with more space would have separated them. It’s true that bigger models tend to have more narrow neurons, but when networks scale up, they can do it in counterintuitive ways.
Looking at DINOv2DINOv2: Learning Robust Visual Features without Supervision Oquab, M., Darcet, T., Moutakanni, T., et al., 2023. arXiv:2304.07193. References ↓, which is more than ten times the size and about ten years of architecture later, it has more pose neurons, but many of them are still generalized. I’ve seen this in every vision model I’ve looked at, of any size.
Both human and automated interpreters seem strongly biased toward describing neurons as detectors for groups of nouns, like cats and cars, rather than for a shared pattern, like a V. Sometimes a neuron really does respond to a group of nouns. In my experience, though, it’s more often a shared pattern, and the neuron is better understood as a verb than a noun. Its job isn’t to name what’s in the input but to draw a distinction downstream neurons can use: it separates two subtle patterns at its positive and negative poles (a V versus an O) and pushes the flow of activations toward different neurons accordingly.
I can often unblock my understanding of a neuron by asking: if I couldn’t use groups of nouns at all, how would I describe this neuron to someone else? Usually the shared pattern pops out right away, and it’s easy to test on more groups of nouns (for the V: shinkansens, pool tables). I suspect this is the largest source of misattributed polysemanticity in mid-layer neurons. Many explanations I’ve seen that call a neuron polysemantic look to me like a single pattern described as a list of nouns.
We should treat a neuron that means two or three things as one neuron with a few grouped meanings. The grouping probably wasn’t a problem for the network: the meanings sit together because, from the network’s point of view, they’re similar. That makes the grouping a clue to how the network thinks, and following it can turn up generalizations we didn’t know about, as cats and cars did here. Often the extra meanings also melt away on closer inspection, turning out to be minor and rarely active. Finally, some apparent meanings are just our tools being bad, and this essay has several examples: the double-negative regime of SwiGLU, an unread saliency knob that makes the residual stream hard to visualize, and the reward hacking in feature visualization we just saw. Keeping the original neuron whole keeps these errors visible, so the error correction can continue. If we’d used optimization to get a cleaner list, we might have lost many of these hints.
A History of Microscope Errors #
The story of interpretability, as I see it, has largely been one of finding out that what was under the microscope was better than we thought. The microscope just wasn’t tuned right, and what needed tuning was often pretty subtle.
Modern neural network interpretability arguably started with feature visualization, which renders a synthetic image of what a neuron responds to most: the famous psychedelic pictures. For a while they didn’t look very good. You could reasonably conclude the neurons had learned something, just nothing very clean. It turned out the method was missing transformation robustness. Once researchers added it, the same neurons came out crisp.
But that only gets you so far into the model. Near the end, feature visualization starts to break down again, for two reasons: the reward hacking we saw with the car circuit, and regularizations that weren’t quite right. Fixing it didn’t require anything fancy. A few subtle tweaks to reasonable settings, and now we can see much deeper into the network. Same neurons, better microscope.
But it wasn’t all the microscope. The model needed to be tuned too. InceptionV1, the early model organism of the field, was totally uninterpretable in the late layers. One estimate said “To the best of my knowledge, there are essentially no interpretable neurons on InceptionV1’s final layer,” which is close to my own recollection, I’d have maybe said 5%. This is what they looked like.
The leading explanation as far as I knew was that there were too many high-layer features needed to be represented for the number of neurons and parameters there were, so it started packing ones that don’t belong together arbitrarily, since they don’t collide. The early layers were monosemantic because not as many early layer features needed to be represented.
But this wasn’t the right explanation. Here’s a more modern convolutional network with the same number of parameters and dataset, but with residual connections, LayerNorm, and a couple other improvements. The last layer is beautiful.
So two things were required. First, the neural network needed to learn well, using an architecture designed by people fully focused on that goal and with no interest in interpretability. The residual stream is one example: it was introduced to help learn deeper neural networks, and it also seems to help the deeper layers be conceptually clean, which may be related. Second, we needed tools strong enough to see them clearly.
The last layer of InceptionV1 was one of the main pieces of evidence that neural networks were almost fundamentally uninterpretable, since the way they learn overlaps unrelated mixtures. Replacing the neurons seemed like the only option. I put it around 5% with tools of that era, others at ~0%. With better interpretability tools (fixed feature visualization, methods for visualizing attribution etc) it’s more like 50% on the same network. This is all with the 2014 model whose checkpoints and training details were lost to time. Retraining the same architecture with more reasonable modern regularization it’s more like 70%. In the ConvNeXt of the same size, it’s about 99%. They’re explorable here.
Recursive Generalization #
So we’ve shown the last layer was fixable. But what was it doing before it was fixed? Even before the architectural improvements, it wasn’t putting things together arbitrarily. When we retrained the model and watched a neuron form, it started on yellow flowers, in photos that sometimes included a dog. Then it grew along each facet of those photos: toward other kinds of nature, toward green objects like balls, and toward dogs without any flowers. This is true of every neuron I looked at. The improved architecture shows the same pattern, but much less of it.
We call this recursive generalization: a neuron generalizing along more than one axis, often first along one axis, then sprouting outwards in ways that are easy to follow locally, but stop it from being “monosemantic” in the end.
For instance, a language-model neuron responds to “iy,” including in “Miyazaki,” then to “diy,” and then to “do it yourself,” which no longer contains “iy” at all. There are examples all over the place in both image and language models. Like a geology neuron that generalizes into Cesium, a geology JavaScript library.
Usually it’s very obvious what happened, to an extent where it’s hard to know whether to just call it normal generalization or not. Like an enemy neuron that fires for a Russian poem about enemies. Sometimes it’s less clear but still plausible. A Srinivasen neuron also lightly fires for the UniProt protein database. The project has several Srinivasen’s associated with it, but it’s hard to know for sure.
We’ll talk about these recursively generalized offshoots and how much they matter in the language model section. Here we just want to talk about what the motif is.
We first hinted at this in Multimodal NeuronsMultimodal Neurons in Artificial Neural Networks
Goh, G., Cammarata, N., Voss, C., et al., 2021. Distill. References ↓, calling them “conjoined features.” I’d forgotten until I went back through the field’s most famous models, months into studying recursive generalization. As far as I can remember, no one, including me, suggested it might be the general explanation, for instance for what was wrong with InceptionV1’s last layer. Without training checkpoints we couldn’t watch it form, but the explanation was basically right. The appendix explains why I think the new name is more accurate.
Superposition #
Recursive generalization isn’t a particularly bold explanation, it’s arguably just what we should expect out of ordinary generalization. But there’s also a much bolder explanation that has been proposed for polysemanticity, a new learning principle.
The main learning principle we see everywhere is generalization. With generalization, things are put together because of a shared pattern that it’s useful to have in one place, so it can be built on. With superposition, unrelated things are placed together because they don’t interfere. These have opposite predictions: generalization should be local, superposition should be non-local, since it would want very different things to be near each other.
Generalization would want a floppy ear dog neuron, which can be thought of as lots of slightly different floppy ears at different angles and kinds together. Superposition would want a floppy ear dog + Latin poetry neuron.
The definitions get tangled here. Along with superposition, Zoom InZoom In: An Introduction to Circuits
Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M. and Carter, S., 2020. Distill. References ↓ introduced the idea of features as the fundamental unit of neural networks. Depending on where you look, superposition then means either “more features than dimensions,” which even a fully monosemantic network would satisfy, because its neurons compose, or a learning principle like the one above, applied sometimes to what neurons learn and sometimes to where features go.For instance, InceptionV1’s layer 4b has basically monosemantic neurons for car parts like windows, car bodies, and wheels, which the next layer assembles into a car neuron, 4c:447. In the feature view, it’s reasonable to say the “car feature” isn’t just in 4c:447, but also in the combination of 4b neurons that builds it. So there are two car features in adjacent layers, sharing almost exactly the same mechanics. In the “more features than dimensions” view, the first is in superposition and the second may not be. In the sense of the informal note Distributed Representations: Composition & Superposition, the first is distributed, but by composition, and the second isn’t distributed at all. This is part of why I find hypothesized features hard to think with, and why I usually prefer specific mechanisms, like neurons, and specific pieces of data.
So “superposition” is itself in superposition between several common uses. We focus on the learning principle applied to the mechanisms like neurons, because whether it forms holds such a disproportionate share of the good or bad news for interpretability. It’s also fairly easy to test: recursive generalization and superposition predict nearly opposite things, and you can watch a neuron grow and see which one happens.
When I say superposition in this essay, I’m referring to this version, and we’ll address the question of “more features than dimensions”, which I basically don’t think is bad news, later.
Superposition of the Gaps #
Polysemanticity is often more a claim about the observer than the neuron. It says there’s no good explanation yet. There’s a few reasons that could be:
- Bad tools: our lens or microscope isn’t good enough
- Bad explanations: created from those tools
- Bad learning: the network learned poorly due to bad architecture, training data, or something else
- Recursive generalization: while generalizing, the neuron branched off in a few related directions
- Superposition: a change in learning principle where the network packs unrelated things into the same neurons because they rarely collide.
For years I was a staunch superposition believer. I started studying language model neurons not out of any interest in neurons or superposition, but because I was interested in whether agents had research taste! I wanted to see if they’d ask good questions as they tested explanations, and started with neurons only as a baseline. I expected them to run into intractable feature fragments in superposition. That would produce a new baseline for how clean neurons really are, then the agents could continue with some cleaner learned basis. But it never came, because the neurons kept coming, resolving to a simple explanation that made surprising predictions that held. In a frontier language model, I found a neuron that mostly fired on “pr”, but also on a bunch of Base64 URL strings as well as some Bengali. I thought this must finally be an arbitrary mixture. But decode the URLs: “pr” words. Translate the Bengali: “pr”.
In a dense Qwen model, a coworker came across a neuron that standard top-k automated interpretability lists as a “polysemantic comma neuron”. But looking more closely, it turned out to track Colorado: phone numbers, slang, and connections to other Colorado neurons. Steering it made the model want to go hiking. The commas were where it stored a running summary of the Colorado context so far. In general attention seems to collect on the punctuation, which we call the summary motif. Once you know it, a lot of similar neurons get easier to understand.
Other times, we just didn’t have enough data. At first we didn’t include long context, two paragraphs felt like enough. But this accidentally made all the long-context neurons look polysemantic.
In another case a neuron looked like an arbitrary mess, with nothing firing strongly. It turns out it looked for specific signs within pages of health SEO spam, and once we had a sufficiently broad dataset, it was understandable.
Across hundreds of cases like these, nothing shook my belief in superposition. It just kept moving. If the strong activations came back clean, superposition would move to the weak ones, or to combinations of neurons, or to the neurons I hadn’t looked at yet. Or maybe across layers themselves.
But eventually I wondered what was going on. Unless this was some dramatic extended case of selection bias, something wasn’t matching the early model organisms. So I went back, the original evidence fell, and that led to this essay.
I now think I was making superposition the null hypothesis when it should have been the last hypothesis, since it’s the most radical and the least testable. It just looks like things that don’t make sense. But that’s also what bad tools look like (feature visualization reward hacking), or recursive generalization you haven’t followed yet (“iy” neuron), or a bad architecture (5b of InceptionV1), or generalization you haven’t understood yet (cat and car).
I think part of what was going on was that superposition just felt right. Anthropic’s Softmax Linear UnitsSoftmax Linear Units
Elhage, N., Hume, T., Olsson, C., et al., 2022. Transformer Circuits Thread. References ↓ paper said as much in 2022: “In the mechanistic interpretability community, superposition is often treated as the default hypothesis simply because it seems intuitively more compelling than other explanations, but there is little evidence.”In the next sentence, the paper says its own results are moderate evidence for superposition. That evidence is circumstantial, in the paper’s own word: SoLU suppresses activations spread across many neurons, an extra LayerNorm restores performance, and after the LayerNorm a neuron’s weak activations looked less related to its main job, which the paper reads as features being smuggled through small activations. We read it as evidence about distributed representation in an architecture built to suppress it, rather than about neurons packing unrelated things together, which is the sense we care about. Telling the two apart in weak activations is hard even with today’s tools, since meaning shifts in every direction at once as activations weaken, and it was harder still in 2022, for all of us. The paper also notes that its discussion of superposition has been superseded by Toy Models of Superposition, which we address in the appendix. We bring it up not as a criticism of either paper, but because it’s the only place we know of where someone said plainly how little real-world evidence there was, and we think people still often assume there’s more than there is. To my knowledge, no one has found a superposition circuit in a real network since my original one, which was an error. I don’t see anything like it in real models, despite expecting to, and I think the bar for claiming it should be high, since it’s such a radical explanation and would be big news, terrible news, for how we should study neural networks.
Of course, this risks the opposite error, an explanation of the gaps: assuming every neuron must have a sensible story. Two things make it less worrying. First, each of the other causes can be tested. A better tool either resolves the neuron or it doesn’t; retraining with a better architecture shows whether learning was the problem; and recursive generalization can be watched as it grows, in small steps. Superposition’s only symptom is that something doesn’t make sense. Second, real explanations tend to be narrow and easily testable. Once you suspect a neuron is a Colorado neuron, steer it and see what it says. If a neuron fires for some narrow TV trope, the set of things that could count as that explanation is tiny, while “something that doesn’t make sense” covers everything outside it.
Also, we don’t have to force an explanation we don’t have. If a neuron might do a couple unrelated things and you’re not sure, it’s fine to say so. I think recursive generalization should be high on the list of things to suspect, and it’s easy to check if you have training checkpoints. I now think superposition should be far down the list.
Exploring Image Models #
Small image models may seem obsolete, but I think they matter a lot to this question because we can see them so clearly. They're also where much of the original worry came from: there we saw what looked like high-resolution chaos, not low-resolution ambiguity. Language models are massive, we don't have tools like feature visualization which make our life easier, attention mixes different kinds of mechanisms together, and many of their motifs are surprising. As I showed above, even vision motifs can be counter-intuitive, and language is worse. Each one of these deficiencies can make a monosemantic neuron look polysemantic. If you're looking at a blurry image trying to figure out if it's a dog or a cat, the blur makes it harder, but it doesn't favor either answer. If you're looking at a blurry image and trying to decide if the image itself is blurry or if you're just looking badly, every flaw in your eyesight gets counted as blur in the image. If we want to study how neurons change with scale, compression pressure, activation strength, and depth, it's better to start where we can see clearly.
I trained three ConvNeXt models spanning roughly 300× in size: a small one with 92k parameters, a medium one with 7.06M (the same size as InceptionV1), and a large one with 28.6M. We’ll look at those, and also InceptionV1, to try to understand what neurons do in different situations, and try to find places packing may form.
How I Think About Neurons
I used to think of neurons as human-shaped nouns: one is a car, another is a curve. Now I think of them more as control flow. A neuron is less about what it sees than about what it does to the flow coming in, and often its job is to gate a region of the network. Seen this way, monosemanticity is almost the wrong question. A better one is what a neuron lets through and what it keeps out.
Pixels come in, and the network begins grouping and distinguishing things. Cars and dogs might be grouped together because they share a pose. Further along there are specific car detectors – for taxis, ambulances and the like – and they don’t want dogs. So a neuron in front of them acts as a gate: it responds positively to cars and negatively to dogs, and its negative side keeps the dogs out. The dogs go on to their own area, which doesn’t have to worry about cars. Later neurons build on this, specializing further: an ambulance neuron might have a taxi on its negative pole, and a taxi an ambulance. In a larger model, these distinctions can become finer still, separating one view of a taxi from another. Smaller models may stop at coarser distinctions, with something closer to a general “car” neuron.
Here’s a real sports car, followed through the last five blocks of our large model. Partway through, it’s grouped with other rectangular shapes, like tables and cabinets, against snakes and dogs. By the last block, the split is convertibles against station wagons.
The negative side of neurons matters almost as much as positives for understanding them, even in ReLU networks where the negative side is cut off. A neuron is built from votes: each input weight votes for its positive side or its negative side, and the votes add up before the nonlinearity. The negative side never gets gradient on its own, but the individual votes do whenever the neuron is on, so it’s shaped by learning too. The negative side’s meaning depends on context: when a neuron in the sports car’s path separates it from snakes and dogs, it isn’t calling cars and dogs opposites, just gating out things that were usefully grouped earlier. The negative side tells us what distinction matters at this point in the circuit, which also makes it a quick check on whether a grouping is real: the V neuron’s negative side is an O.
If we look at the last layer, we see high-level patterns and what they’re being distinguished from: taxis from ambulances, broccoli from cauliflower, and spoonbills from flamingos. When we look at our smallest model, we see the neurons get coarser. Its last layer ends at cars, trains, and carts in one neuron, with boats, planes, and a water-based ox-cart, (which I didn’t know even existed), on its negative side. It’s still differentiating, but more coarsely, and it ends early.
What Neurons Do Under Pressure
Early layers in most vision models are monosemantic. As I understood it, the standard explanation is because there’s not that many early vision features that need to be represented. On its surface, it’s a bit strange that every model, from AlexNet, to ResNets, to InceptionV1, to CLIP all have the right ratio of early features to early neurons, since they have such different ratios of dataset to parameter size.
One layer we can focus on is 3b of InceptionV1, the layer with Curve Neurons and High-Low Frequency neurons, which are among the most fundamental in my opinion. This model has no residual stream, so there’s no way to “flow around” the neurons. So we can test what happens to neurons when there’s not enough space by reducing the amount from 480 to 10 and retraining.
If it switched to superposition, we should see overlapping mixtures on top of each other, like a curve plus an orange spiral plus spider webs and several other texture detectors. We’ve crushed it 48x so it should be dramatic! But that’s not what we see. The neurons stay interpretable and generalized, but coarser: one “curves & loops” neuron instead of ten curve detectors, two circle neurons instead of several circle types, and so on.
We can test curves explicitly to see what happened to them. If they were placed in some high dimensional combination of neurons, we should be able to extract them. So we can train an SAE on this crushed layer and compare them to the originals. We don’t see anything like clean curve neurons!
Original 3b · 480 neurons
Ten curve detectors
Retrained with 3b at 10 neurons
The whole squeezed layer
An SAE on the squeezed layer
Closest match to each detector
Curves & loops
Details #
The squeezed model was trained from scratch with mixed3b cut from 480 to 10 channels and no bypass, and reached 58.35% top‑1 and 81.14% top‑5 on the 50,000 ImageNet validation images. The autoencoder is a TopK SAE with 640 features, 2 active at each position (the essay’s SAE recipe plus the usual auxiliary loss that keeps features from dying), trained on the squeezed layer at 25.6 million positions from 200,000 ImageNet training images. It explains 82% of the variance of held-out positions, and with it spliced in, top‑1 on 10,000 validation images falls from 58.8% to 43.0%.
Each detector and its closest SAE feature
| Detector | Prefers | Width | SAE feature | Prefers | Width | Shape match | Photo match |
|---|---|---|---|---|---|---|---|
| 3b:385 | 16° | 40° | 79 | 38° | 174° | 0.52 | 0.09 |
| 3b:343 | 84° | 40° | 64 | 84° | 64° | 0.72 | 0.14 |
| 3b:342 | 116° | 56° | 510 | 86° | 80° | 0.61 | 0.19 |
| 3b:388 | 166° | 40° | 174 | 192° | 44° | 0.38 | 0.11 |
| 3b:324 | 184° | 24° | 174 | 192° | 44° | 0.79 | 0.26 |
| 3b:340 | 214° | 42° | 435 | 232° | 146° | 0.48 | 0.08 |
| 3b:330 | 244° | 52° | 89 | 232° | 88° | 0.52 | 0.11 |
| 3b:349 | 272° | 34° | 277 | 278° | 64° | 0.60 | 0.18 |
| 3b:379 | 328° | 44° | 544 | 32° | 138° | 0.56 | 0.18 |
| 3b:406 | 356° | 26° | 354 | 334° | 94° | 0.42 | 0.17 |
Tuning is the response at the centre of the image to a thin dark arc drawn at 180 angles and two curvatures, as in the Missing Curves appendix. Each detector’s match is the SAE feature, among those that respond to drawn curves most strongly, whose tuning curve correlates best with the detector’s (shape match). Width is the span of angles above half the peak; the tuning curves in the rings are each scaled to their own peak. Photo match is the correlation, over every position of 10,000 held-out photos, between the feature and the original detector at the same place.
The originals are 24–56° wide and silent at the opposite angle; the matches are 44–174° wide and keep responding at the opposite angle, and 3b:388 and 3b:324 share one feature. None of the 91 features that respond most to drawn curves is silent at the opposite angle. Dropping that filter finds features with closer, narrower shapes (0.63–0.98), but they respond to drawn curves at only 9–42% of their usual strong activation (the originals at 72–183%) and track the detectors in photos even less (0.02–0.12). The essay’s SAE of the original layer matches the same detectors with shape correlations of 0.96–1.00. On the squeezed layer, even the best nonlinear readout of the ten neurons at one position tracks each detector with a correlation of only 0.29–0.48. Other autoencoders tell the same story (1 to 4 active features, 160 or 640 features, a second seed, or no auxiliary loss): median shape match 0.47–0.63, photo match at most 0.27. With 3 or 4 active features the preferred angles line up better, but the curves stay broad and two‑sided.
Control
Optimized images use the essay’s path‑anchored recipe (the one behind the original curve images) for the squeezed neurons, and its SAE recipe (the one behind Missing Curves) for the SAE features. The path weight is scaled to each unit’s activation range, since the squeezed layer’s activations are hundreds of times smaller than the original’s. The same SAE recipe, run on the essay’s SAE of the original layer, draws clean curves:
The squeezed neurons’ photos
Squeezed‑neuron names are provisional visual descriptions. Photos rank each neuron’s image maxima over 50,000 training images.
I think this is a good test of whether we can turn a monosemantic layer I know well superposed. But in some ways it’s not ideal, since people mostly suspected superposition in the last layers. So we can take our medium size ConvNext, crush the last layer, stop it from routing around the residual stream, and see what it learns. We see the same pattern of it broadening.
This is also what we see with the small ConvNext, which in some ways is a very similar test.
Late Circuits
We can also look at circuits in the late layers, to rule out any chance that the last layer is somehow special. We see that it’s organized the whole way through.
What Dictionaries Learn
If combinations of neurons were hiding things, a dictionary (eg an SAE) should find building blocks the neurons lack. So I trained several on these layers and went looking, myself and with agents. I didn’t find any. Mostly, the two looked like what they mechanistically are: the neurons like building blocks, the SAE units like those blocks grouped by context. It was clearest with spiders. The SAE units were spiders in different settings; the neurons were leg patterns and webs. They didn’t disagree much.
SAEs did help with “what’s in this image?” They helped most in the smallest model, whose neurons are coarse. At medium size I slightly preferred them; at large, I think I prefer the neurons. But I care more about understanding the network than about any one image, and I never saw the SAE grasp something fundamental the neurons didn’t. If a model at a particular size failed to understand baseball players, so did its SAE. When the model finally understood a baseball player, it was on a neuron.
What did the SAE find that didn’t already have a clear neuron counterpart? On the last layer, 88% of its activity is near-copies of single neurons, and the rest is faint: removing every dictionary direction that matches no neuron leaves accuracy unchanged, and the few answers it flips were already near-ties. Additionally, if you search the rotations of a layer for the one where a handful of units carry the most activity, starting from a random one, you land back on the neurons, in every block. Among the few features strong enough to look at, what the SAE added was loose groupings built from existing neurons, like animals at the water’s surface or pale, lumpy food. Nothing we found there looked like a fundamentally new building block.
Rare Things and Memorization
A lot of the worry about hidden features comes from rare things. They seldom collide with anything, and they may not seem worth a neuron of their own. So we looked at rare things every way we could think of. Rare classes are often more consolidated: you need a lot of dog patterns for ImageNet, but only one Triceratops neuron. When I removed weak activations in the last stage, no class stood out. The adversarial examples we traced used normal semantic circuit routes.
I looked at the images whose representations in the last layers are furthest from any other image. They end up being entertaining, weird images that the model misreads with ordinary neurons: a hummingbird seen as ice cream, a giant banana seen as a vault, and a guy in a fallen wheelbarrow seen as a lawn mower. I also trained an ImageNet model to map random noise to random labels, and even that was partly semantic: it partially leaned on wool textures and the texture of distant trees, which are high-frequency, a bit like high frequency noise. I didn’t fully audit the model, but thought it was interesting that it used them at all. Across all experiments in language and vision, memorization tends to want to sit as close to the output as possible, since all it has to do is look the answer up, and the late layers did most of the work.
We can also make a rare route ourselves, by teaching our 7.06M-parameter model something nonsensical: that a photo of an umbrella is a bee. You could imagine this memorization not relying on anything else: a hidden feature that latches onto this exact set of pixels and says “bee,” cut off from the rest of the network because it doesn’t need it. That would be pretty bad news, but it’s not at all what happens. In nine training steps, the top guess went from umbrella to maypole to bee. The umbrella neurons faded and existing bee neurons came on, nearly all of it in the last three blocks: swap just those into the original model, and it says “bee.”
This is in comparison to learning some sort of hidden umbrella feature separate from any other normal umbrella or bee route. While this is possible, I think it’s slower to learn than assembling related parts that are already activating and have routes to the output, giving it plasticity and gradient. It’s also more fragile, and isn’t what we see anywhere.
Do Weak Activations Matter? #
When people ask whether weak activations matter, I think there are several questions mixed together, and they have different answers.
Are they generalization or superposition?
Unsurprisingly, I think it’s generalization. Imagine you’re a packed, unrelated feature trying not to collide with what’s in a main neuron. Weak activations are vastly denser, so you’d be colliding with orders of magnitude more things, for little upside.
But maybe you hide in the weak activations of many neurons at once. What we see is that you can ablate almost all the weak activations of the network for almost no cost. The places where removing them costs the most are some of the easiest to understand. In language models, they include the last layer, which is largely used to suppress single tokens, for what we call a “say or just think” motif. In vision models, it’s the early layers, which function like dials: the difference between slightly bright and medium bright matters, but for a later detector, like king penguin, low or very low evidence doesn’t. So critical common things can’t be in the weak activations, since we can remove them.
We can also look at what we lose when we cut weak activations. Across several kinds of studies, it’s usually fine distinctions between things the neural network is already confused about. And the faint activations that tip them come from the semantically related neurons you’d expect.
This leaves rare data and strong but not extremely strong activations.
Rare things, as we saw earlier, have grown on related circuits in every experiment we’ve run, since they aren’t strong enough to form niches of their own. The ablation above points the same way: if rare cases lived in the weak activations, cutting them should have broken some answers outright. Instead, every change was small, a near-tie tipping one way or the other.
And the strong-but-not-extreme activations we can just look at. In every case I know of, they radiate outward semantically: for CLIP’s Trump neuronMultimodal Neurons in Artificial Neural Networks
Goh, G., Cammarata, N., Voss, C., et al., 2021. Distill. References ↓, photos of Trump, then art and people related to him, then Fox News studios. The further down you go, the easier it is to be fooled by something that seems unrelated but isn’t. At minimum, look at the circuits a datapoint flowed through; what fired weakly here may have fired strongly one layer back.
Here's every strong firing of five neurons in a language model. The center is each neuron's maximum, as it radiates out it becomes more distantly related. Gray is where it's hard to place, usually these are still connected but require more work to understand, since it's more distant. You can click any to open them up and see which circuits each activation came from.
Percentiles or raw activations?
I used to think in percentiles, since they’re easy to think in and standardize across neurons. But I’ve come to think a neuron’s raw shape is telling us something. Some neurons hold their meaning far down the range; others drop off a cliff, like the Christmas neuron above at about 5% of its max. Some fire a few times a book, others a few times a page. Percentiles flatten that, and it’s often what tells you whether you’re looking at a dial, where the middle matters, or a detector, where it mostly doesn’t.
Percentiles also mess with intuition. The 99th percentile sounds high, but 1 token in 100 is about a quarter of a page, and a neuron for a niche feeling or social context – which are all over language models – shouldn’t come up four times a page. Thinking in pages helps: the 99.9th percentile is every two or three pages, the 99.99th every 25 or 30.
Do you need them to label a neuron?
Usually not. After looking at thousands of language model attribution graphs by hand, it’s hard to overstate how much neurons are analogizers. If you’re talking about voyaging on a boat, the model might judge that a bit Titanic-ish. I don’t think it’s usually helpful to think of a “naval voyage feature” spread across the weak activations of the Titanic neuron and others. Instead, there’s a Titanic-ish mechanism: whenever something is Titanic-ish, it fires, and its weights are votes for Titanic-ish things. So you look at the top activations, see that it’s Titanic-ish, and that label usually holds up through the rest of your audits. If you ever really care about a particular activation, you can follow its pathways, ablate the circuits, and keep zooming in to almost arbitrary resolution.
My main takeaway: if a weak activation looks unrelated, don’t be too worried unless you’ve looked at it carefully. Of the probably hundreds I’ve followed by hand, almost all turned out to be related, coming through a reasonable circuit; it just wasn’t visually obvious until I traced it.
Many were clever, like a Zen koan neuron that fires weakly on recursion in math: many koans turn on the same kind of self-reference, looking for the observer. Others were more distant and ugly, not always the connection I would have drawn, but you could usually follow how it got there. I suspect most, possibly almost all, are like this.
Exploring Language Models #
This section was done with Curt, and brings in experiments from our forthcoming technical paper studying language neurons in depth, so I’ll use we for this section.
Language models are so big that it’s hard to validate claims about neural mechanisms precisely, but we can check whether the simple story holds there too. We start with a medical diagnostic attribution graph we liked from Anthropic’s Circuit BiologyOn the Biology of a Large Language Model
Lindsey, J., Gurnee, W., Ameisen, E., et al., 2025. Transformer Circuits Thread. References ↓. They were studying Claude 3.5 Haiku using Crosslayer Transcoder features. It turns out the same works with neurons in Qwen3-32B.
Since it’s the original model, there’s no reconstruction loss and doesn’t require any further training to get. Their largest transcoder predicts the same next word as the original only half the timeCircuit Tracing: Revealing Computational Graphs in Language Models
Ameisen, E., Lindsey, J., Pearce, A., et al., 2025. Transformer Circuits Thread. References ↓ and their graphs need error nodes for what it misses.
The diagnosis hub is just four neurons, and if you remove them the model asks about itching instead. Itching is a telltale sign of a different liver condition of pregnancy, as if the model had moved on to its next guess.
We’ve also added a few other attribution graphs we liked. There’s a few other graphs if you’d like to see them. Across thousands I’ve seen in image models and language models, they’re highly readable and ablate the way you’d generally expect.
The neuron’s negative side matters about as much as its positive. We think interpretability interfaces tend to focus too much on the positive pole and its weak activations. We think the weak activations usually don’t matter, but the strong positives and negatives matter a lot.
For instance, Santa lives on the negative pole. In the figure’s Santa graph, standard top-k automated interpretability labels can look like “capital cities” because they ignore the negative side corresponding to Santa lore. Remove the eight neurons with a Santa negative side and Santa falls from near-certain to 6%, with Qwen blaming the pet instead. Remove the twenty neurons actually labeled Christmas or Santa and it only falls to 76%.
Here’s an attribution graph from Kimi K3 using only neurons. As far as we know, it’s the first published attribution graph of a multiple trillion parameter model. Asked what to do with a wallet it found, it links the wallet, the cash and the owner’s ID to lost property and honesty, and answers “return”. We haven’t begun full audits of this scale yet, and hope to soon, but so far most things have been intuitively interpretable.
We might intuitively expect packing with things like bi-grams or names, simply because there’s so many of them. What we find is that it’s ordinary generalization and composition. Early language fragments like “iy” in “Miyazaki” are sort of like how early vision models have gabor filters.
In the appendix we address Neurons in a HaystackFinding Neurons in a Haystack: Case Studies with Sparse Probing Gurnee, W., Nanda, N., Pauly, M., Harvey, K., Troitskii, D. and Bertsimas, D., 2023. arXiv:2305.01610. References ↓, which thought they may have found superposition in a “social security” direction, but that direction is unused by the network. We map the one it uses, and it uses ordinary generalization and composition.