--- title: Neural Networks Are Already Clean | daily.dev description: A researcher who has spent months studying neurons and circuits in image and language models argues that the field's canonical evidence for 'superposition'... canonical: https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Neural Networks Are Already Clean | daily.dev og:description: A researcher who has spent months studying neurons and circuits in image and language models argues that the field's canonical evidence for 'superposition'... og:url: https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj og:image: https://api.daily.dev/og/posts/cbOhFspUj.png og:image:alt: Neural Networks Are Already Clean og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Neural Networks Are Already Clean **[Hacker News](https://daily.dev/sources/hn)** · 136 min read · 0 upvotes · 0 comments ## Summary A researcher who has spent months studying neurons and circuits in image and language models argues that the field's canonical evidence for 'superposition' (the idea that neural networks pack unrelated concepts arbitrarily into the same neurons due to lack of capacity) was largely a misreading. Revisiting the famous cat-car neuron and the dog-car circuit from Zoom In, the analysis finds they actually reflect shared visual patterns (a V-shape, shared pose) rather than arbitrary packing. InceptionV1's uninterpretable last layer turns out to stem from architectural and tooling limitations, not superposition; modern architectures with residual connections and LayerNorm reach 99-100% interpretable neurons at the same size. Most apparent polysemanticity is attributed to a motif called 'recursive generalization', where a neuron grows outward along related but non-obvious axes (e.g., an 'iy' neuron generalizing to 'diy' then 'do it yourself'). Preliminary audits of Qwen3-32B and trillion-parameter models like Kimi K3 suggest 90%+ of neurons have one main interpretable job. The piece argues this reframes interpretability as tractable 'normal science' on networks directly, rather than requiring decomposition techniques to first untangle them, with a technical companion paper promised for more detail. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What causes polysemantic neurons in neural networks like InceptionV1? Most apparent polysemanticity comes from a motif called recursive generalization, where a neuron generalizes along multiple axes over training, sprouting related offshoots the way a tree branch sprouts sub-branches, rather than from superposition (unrelated concepts arbitrarily packed together due to lack of capacity). InceptionV1's uninterpretable last layer stemmed from architectural overflow and weak visualization tools, not a distinct superposition learning principle. _Researchers debugging model interpretability pipelines can follow ongoing coverage of these findings on daily.dev._ ### How interpretable are modern large language model neurons like Qwen3-32B? Preliminary audits of Qwen3-32B found roughly 90% of neuron poles have one main job, often with small related offshoots, about 7% have two related jobs, and only 2-4% remain not understood, with similar results reported up to trillion-parameter models such as Kimi K3. This contrasts sharply with 2023-era estimates like OpenAI's automated interpretability project finding only about 1.7% of neurons interpretable. _Anyone evaluating how legible today's language models really are can track interpretability research like this on daily.dev._ ### Was the famous cat-and-car polysemantic neuron in InceptionV1 actually an arbitrary mixture? No, the neuron originally cited as the canonical example of polysemanticity in InceptionV1 was later found to be a V-shape detector rather than an arbitrary cat-car mixture. It responds to V-shaped flames, water jets from a fireboat, and fox faces, and its negative pole is an O-shape, indicating a shared geometric pattern rather than unrelated concepts folded together. _Engineers revisiting foundational interpretability claims can follow re-analyses like this through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 94 comments across hackernews (as of 2026-10-19). **TL;DR:** Reactions are split between people who find the essay's reframing of polysemanticity as misread generalization compelling and well-argued, and others who see it as overlong, unproven, AI-assisted-sounding, and overreaching relative to prior work on superposition and feature compression. **Sentiment:** 20% positive · 40% mixed · 40% skeptical **The case for** - Some found the generalization explanation (e.g., the Chinese '羊' compression example) intuitive and a plausible account of why small models look messy. - A few praised the linked interactive neuron-microscope demo as the first non-trivial network example they felt they understood. - Several appreciated that the piece at least attempts rigorous, falsifiable claims about specific neurons and circuits rather than vague assertions. **The pushback** - Many felt the essay is excessively long, repetitive, and poorly summarized, requiring outside TL;DRs or glossaries to parse. - Commenters pushed back that the claims overreach what's actually proven, noting prior work on feature compression/condensation isn't addressed or reconciled. - Some argued interpretability claims can't really be verified outside domains humans already understand (e.g., image/language), so the approach may not generalize to harder domains like medical imaging. - Technical skepticism that sparse-autoencoder-style interpretability methods are reliable or can surface genuinely novel (not already-known) concepts. - Complaints about the site itself: laggy/heavy 3D rendering, fingerprinting scripts, and page crashes. - Some doubted the author's credibility or framing, including unrelated personal criticism of his past public statements. **By community** - hackernews (mixed): Discussion ranges from technically engaged skepticism and debate over what 'clean' even means, to praise for specific examples, to complaints about length, site performance, and the author's credibility. **Hottest debate:** Whether the 'recursive generalization' explanation genuinely overturns superposition as the cause of polysemanticity, or merely reframes it without resolving open questions about feature compression. **Open questions** - Does this interpretability approach generalize to domains (like medical imaging) where ground truth can't be verified by human recognition? - How does the generalization account relate to or contradict existing work on feature/geometric condensation in multimodal and other models? - Can these methods actually surface novel, previously-unknown concepts, or only confirm concepts researchers already suspect to look for? **Highlights** > 羊. This is a Chinese character that can refer to goats, rams, and several other kinds of "medium-sized animal with thick wooly coats and horns". English has no direct equivalent; we have to be too specific when we talk about these animals. A model with a small number of neurons has to compact its concepts in order to fit within the parameter count while reducing loss. OP's article shows that the compaction is usually generalization: "goat" and "ram" get compressed into "羊". The canonical example of superposition, an image model where "car" feeds into a bunch of "dog" features, was shown to be actually generalization. In a sense, the model was so small that its optimizer didn't have the parameter budget to distinguish between "car" and "dog"; it has a single set of "thing that moves fast" features. So basically, really small models are merging features into semantic buckets that are so generalized they don't make sense to big human brains unless we really, really squint. The buckets chosen happen to be what has the lowest loss, of course. > — [kmeisthax on hackernews](https://news.ycombinator.com/item?id=50107822) > The claims seem to be an overreach of what is actually proven. We've known about progressive feature complexity and the impacts of early training structures on later training for a really long time (since CNN's, The Lottery Ticket Hypothesis, etc...) and neither of those phenomenon are mutually exclusive with parameter overloading. The Multimodal Gap work (https://arxiv.org/abs/2203.02053) seems to imply networks have a tendency towards condensed representations that are clearly not semantic and a bunch of other considerations are already on the table for discussion. I don't think these have all been considered sufficiently to start claiming we are around the corner from mechanistic interpretation. > — [bluejay2387 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50108530) > Typically the analysis works in the reverse direction: we have a neuron (or SAE feature, or whatever) and we look at examples of data that makes it activate, then inferring an interpretation. This is a lot harder than probing for a known concept, and one of the points of this article is that if you try harder to understand the activation pattern, you will very often find that there is something interpretable there, even if at first glance it looked random. > — [aesthesia on hackernews](https://news.ycombinator.com/item?id=50109361) > But they don't. That's the whole reason they work in the first place. Every neuron in the layer encodes fractionally something about the concepts embedded in the whole thing. It's more like a hologram than a series of connected facts. > — [jacquesm on hackernews · 1 comments](https://news.ycombinator.com/item?id=50107054) > So we have here a massive evidentiary brief, the point of which is to argue that it is easier than we thought to, like, "sequence the AI genome" (in my estimation). We revisit famously-messy neurons, examine various tool failures, check places where messy things still hide, and catalog how "polysemanticity" is represented variously in the network. Once sequenced, the thinking goes, we'll be able to do things like: explain and anticipate failures; predict what happens when you modulate the pathway an input takes on its way to becoming output; reconstruct algorithms the models have learned. I'm not 100% sure this will work out like the mechinterp people think. If the neural representations and pathways are "clean" (organized into a a meaningful-to-humans hierarchy), it seems at least very plausible that the behaviors that AI xrisk people are worried about will mostly boil down to straightforward synthesis (i.e., mostly-linear-interpolation) of mostly-well-cataloged knowledge. And if that is true, then it does not seem hopeful that there would be a way to prevent the output they are worried about, other than just removing a bunch of the knowledge. So maybe there is no alignment solution at all. > — [antics on hackernews](https://news.ycombinator.com/item?id=50108927) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50105647) · 273 points · 94 comments ## Similar posts on daily.dev - [How artificial neural networks build their own learning machinery](https://daily.dev/posts/how-artificial-neural-networks-build-their-own-learning-machinery-wbguw3n0l) · MIT News · 0 upvotes · 0 comments - [UNDERSTANDING NEURAL NETWORK, VISUALLY](https://daily.dev/posts/understanding-neural-network-visually-0zwijz9vv) · Hacker News · 2 upvotes · 0 comments - [The Aristotelian Correction: Why Neural Networks May Only Converge Locally](https://daily.dev/posts/the-aristotelian-correction-why-neural-networks-may-only-converge-locally-ifflkjxcr) · Medium · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#neural-networks](https://daily.dev/tags/neural-networks) [View this post on daily.dev](https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Neural Networks Are Already Clean","url":"https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj"},"datePublished":"2026-10-15T17:13:24.516Z","dateModified":"2026-10-19T18:09:44.013Z","description":"A researcher who has spent months studying neurons and circuits in image and language models argues that the field's canonical evidence for 'superposition'...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d858d7b4a318c24655cd88c5405452fb?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d858d7b4a318c24655cd88c5405452fb?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,neural-networks","timeRequired":"PT136M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Neural Networks Are Already Clean"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/neural-networks-are-already-clean-cbohfspuj#faq","mainEntity":[{"@type":"Question","name":"What causes polysemantic neurons in neural networks like InceptionV1?","acceptedAnswer":{"@type":"Answer","text":"Most apparent polysemanticity comes from a motif called recursive generalization, where a neuron generalizes along multiple axes over training, sprouting related offshoots the way a tree branch sprouts sub-branches, rather than from superposition (unrelated concepts arbitrarily packed together due to lack of capacity). InceptionV1's uninterpretable last layer stemmed from architectural overflow and weak visualization tools, not a distinct superposition learning principle. Researchers debugging model interpretability pipelines can follow ongoing coverage of these findings on daily.dev."}},{"@type":"Question","name":"How interpretable are modern large language model neurons like Qwen3-32B?","acceptedAnswer":{"@type":"Answer","text":"Preliminary audits of Qwen3-32B found roughly 90% of neuron poles have one main job, often with small related offshoots, about 7% have two related jobs, and only 2-4% remain not understood, with similar results reported up to trillion-parameter models such as Kimi K3. This contrasts sharply with 2023-era estimates like OpenAI's automated interpretability project finding only about 1.7% of neurons interpretable. Anyone evaluating how legible today's language models really are can track interpretability research like this on daily.dev."}},{"@type":"Question","name":"Was the famous cat-and-car polysemantic neuron in InceptionV1 actually an arbitrary mixture?","acceptedAnswer":{"@type":"Answer","text":"No, the neuron originally cited as the canonical example of polysemanticity in InceptionV1 was later found to be a V-shape detector rather than an arbitrary cat-car mixture. It responds to V-shaped flames, water jets from a fireboat, and fox faces, and its negative pole is an O-shape, indicating a shared geometric pattern rather than unrelated concepts folded together. Engineers revisiting foundational interpretability claims can follow re-analyses like this through daily.dev."}}]} ```