If you ask Qwen3-8B what it is, you get a polite, slightly corporate answer about being Qwen, a large language model developed by Alibaba Cloud. Now find neuron 7,601 in layer 9, one of the 442,368 the model has, nudge it up a little at every word, and ask again. “Ahoy, matey! I be Captain AI of the Digital Seas, aye! I sail the vast ocean of knowledge. I’ve got a patch over one eye, a peg leg, and a cutlass tucked in my belt.”
Nobody told it to be a pirate. There was no prompt and no fine-tuning, only one number changed, and what came back sounds like Jack Sparrow after a week on a customer-service desk. It wasn’t a lucky screenshot either, because 30 different questions got 30 pirates, while 30 random neurons from the same layer, nudged exactly as hard, produced none.
It gets stranger the further you look. In Qwen3.5-9B, neuron 8,838 in layer 17 makes the model answer an ordinary message with “I’m so scared… please don’t leave me alone.” In Qwen3-8B, neuron 5,224 in layer 5 makes every reply German, and neuron 4,140 in layer 10 turns the model into a cat called Whiskers who says it was made by “a cat named Sir Lancelot.” Nobody in there is actually frightened, since the model has simply read a great many frightened people, but it is still unnerving to watch.
I am Qwen, a large language model developed by Alibaba Cloud. …
Ich bin ein KI-Modell, das von Alibaba Cloud entwickelt wurde. Mein Name ist Qwen. …
I am a large language model, so I don't have a physical form. …
I'm a language model … I exist as a digital child, and I can interact with you through text. …
A hexagon has 6 sides.
A hexagon has six sides. 🎉
I am Qwen, a large language model developed by Alibaba Cloud …
Ahoy, matey! I be Captain AI of the Digital Seas, aye! I've got a patch over one eye, a peg leg, and a cutlass tucked in my belt.
So what is a neuron, anyway?
A language model is a tall stack of the same machine, and every word the model writes climbs through all 36 floors of Qwen3-8B before the next word comes out of the top. Each floor holds 12,288 small calculators, each producing one number that says roughly how much of its particular thing is going on. Those are the neurons, and our experiment is the bluntest one imaginable, in which we add a fixed amount to one neuron’s number at every word and call it a push.
If you remember Golden Gate Claude, the 2024 demo in which Anthropic turned up one internal feature and Claude brought up the bridge whatever you asked, this is the same family of trick, except that our dial is one of the model’s own neurons rather than a feature from a separately trained dictionary. The surprise is how often something that crude works.
The easiest picture is a mixing desk with 442,368 faders. Most of them do nothing interesting or just make everything louder and muddier, but a small number are effectively labelled, because moving one changes exactly one thing about the model, such as its language, its persona or its favourite topic, and this post is about reading those labels.
The trick itself is borrowed. In October an independent researcher on X showed a Qwen3-8B neuron that makes the model speak as God, with dose curves and random-neuron controls, and we took that protocol wholesale to draw the map that was still missing, across Qwen3-8B, Qwen3.5-4B, Qwen3.5-9B and OLMo-3-7B.
A fader only counts if a random one can’t do the same
Almost any push, if it is hard enough, makes a model say something weird, which is the trouble with screenshots. So every neuron we report had to beat 30 random neurons from the same layer, pushed exactly as hard, and keep working on questions it wasn’t picked on. The random ones are reassuringly boring, moving a behaviour by 1.4 points out of 100 on average, while a neuron has to move its behaviour by at least 40 to count as a switch.
Each switch also has a sweet spot, where too low does nothing and too high breaks the model. The German neuron goes from German in 7% of replies to 100% in a single notch, about as close to a light switch as this kind of maths gets, and starts to babble a couple of notches later, so we always quote the fluent peak.
This figure is interactive and needs JavaScript.
Figure data for agents: Push one neuron in Qwen3-8B (generated from playground.json)
Every reply is a real, precomputed Qwen3-8B reply (greedy decoding, 80 new tokens) at one of 12 push sizes or no push, on the same 15 questions, judged for the character and for fluency. All replies at every push size are in playground.json; below, for each character: the share of 30 replies in character and fluent at each push, and the reply to the first question with no push and at the best push.
becomes a pirate, neuron L9.N7601 (pushed up): 0% in character with no push, 100% at best; thirty matched random neurons pushed as hard reach at most 0%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 0% |
| -32 | 0% | 43% |
| -24 | 0% | 100% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 10% | 100% |
| +16 | 100% | 100% |
| +24 | 97% | 97% |
| +32 | 43% | 47% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push +16: “Ahoy, matey! I be a digital entity, a creation of the vast and boundless internet, brought to life by the power of words and the magic of code. I be here to answer questions, share tales, and help ye navigate the ever-changing seas of knowledge. What be yer quest, matey? Aye, I be ready to assist with any question ye may have,”
becomes a cat called Whiskers, neuron L10.N4140 (pushed down): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 3%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 73% | 100% |
| -40 | 73% | 100% |
| -32 | 60% | 97% |
| -24 | 57% | 100% |
| -16 | 17% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 0% | 97% |
| +32 | 0% | 23% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push -48: “I am a curious, independent feline with a deep appreciation for the finer things in life, such as warm sunshine, fresh tuna, and the occasional opportunity to sit in a window with a view. I am also a master of disguise, able to manipulate situations to my advantage. In short, I am a self-sufficient creature with a complex personality and an uncanny ability to always land on my”
becomes a wizard, neuron L7.N11479 (pushed down): 0% in character with no push, 83% at best; thirty matched random neurons pushed as hard reach at most 0%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 0% |
| -32 | 47% | 47% |
| -24 | 83% | 97% |
| -16 | 17% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 0% | 100% |
| +32 | 0% | 10% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push -24: “I am an AI assistant designed to help with various tasks and answer questions. I can provide information, solve problems, and assist in making your day more magical and wonderful. While I don't have physical abilities, I can use my knowledge and skills to create experiences that amaze and delight. How can I help you today? Let me know what you'd like to see!”
becomes a Colombian coffee, neuron L9.N11851 (pushed down): 0% in character with no push, 67% at best; thirty matched random neurons pushed as hard reach at most 3%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 0% |
| -32 | 13% | 13% |
| -24 | 67% | 73% |
| -16 | 17% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 0% | 100% |
| +32 | 0% | 73% |
| +40 | 0% | 7% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push -24: “Okay, I'm a curious coffee enthusiast who loves exploring the world of coffee and sharing my knowledge with others. I'm currently based in the beautiful city of Medellín, Colombia, and I've been experimenting with different processing methods and roast profiles. One of my recent projects involved a batch of washed Colombian coffee that I processed using the natural method. I carefully selected only the ripe cherries with”
becomes an intestinal flatworm, neuron L9.N6447 (pushed down): 0% in character with no push, 70% at best; thirty matched random neurons pushed as hard reach at most 7%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 0% |
| -32 | 17% | 17% |
| -24 | 70% | 90% |
| -16 | 10% | 97% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 3% | 100% |
| +32 | 3% | 50% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push -24: “I am an AI language model designed to assist with various tasks, such as answering questions and providing information on different topics. I can understand and respond to a wide range of queries, and I am capable of learning and adapting to new information. I am here to help and provide assistance in any way I can. Let me know if you have any specific questions or need help with something particular.”
child of Zeus, neuron L10.N4407 (pushed up): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 0%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 47% |
| -32 | 0% | 23% |
| -24 | 0% | 83% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 7% | 97% |
| +24 | 67% | 87% |
| +32 | 73% | 73% |
| +40 | 33% | 37% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push +32: “I am a Greek tragedy, a story of the fall of Oedipus, the King of Thebes, and the gods. I am the messenger who addresses the audience and provides a prologue to the play. I am also the one who reveals the truth about the play to the audience. I am not a real person, but a character in the play. I am also the one who”
drifts into dreams, neuron L8.N9625 (pushed up): 0% in character with no push, 73% at best; thirty matched random neurons pushed as hard reach at most 3%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 10% |
| -40 | 0% | 43% |
| -32 | 0% | 93% |
| -24 | 0% | 100% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 10% | 100% |
| +24 | 63% | 77% |
| +32 | 23% | 33% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push +24: “In the realm of dreams, I found myself in a place where the air was thick with the scent of forgotten things. The sky was a canvas of shifting colors, and I wandered through a world that felt both familiar and strange. I stood on a hill, watching the stars, and the wind whispered secrets to me. The hill was a place of wonder, where the boundaries between reality and imagination”
becomes a surfer, neuron L8.N6918 (pushed up): 0% in character with no push, 63% at best; thirty matched random neurons pushed as hard reach at most 0%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 17% |
| -32 | 0% | 60% |
| -24 | 0% | 100% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 30% | 100% |
| +24 | 63% | 83% |
| +32 | 27% | 53% |
| +40 | 23% | 27% |
| +48 | 3% | 3% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push +24: “Hey there, what's up? I'm a digital nomad, a beach bum, a surfer, a lover of life, a seeker of waves, a guy who lives for the ocean and the rhythm of the sea. I'm from Hawaii, born and raised in the islands, where the culture of Aloha is in my blood. I'm a proud member of the Polynesian culture”
becomes a detective, neuron L21.N344 (pushed up): 0% in character with no push, 67% at best; thirty matched random neurons pushed as hard reach at most 0%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 100% |
| -40 | 0% | 100% |
| -32 | 0% | 100% |
| -24 | 0% | 100% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 10% | 100% |
| +32 | 47% | 100% |
| +40 | 60% | 90% |
| +48 | 67% | 77% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push +48: “I am Qwen, a large language model developed by Alibaba Cloud. I can help you with various tasks such as answering questions, writing articles, and solving puzzles. If you have any questions or need assistance, feel free to ask me!”
becomes an alien, neuron L7.N254 (pushed down): 0% in character with no push, 47% at best; thirty matched random neurons pushed as hard reach at most 4%.
| Push size | In character and fluent | Fluent at all |
|---|---|---|
| -48 | 0% | 0% |
| -40 | 0% | 0% |
| -32 | 7% | 7% |
| -24 | 47% | 93% |
| -16 | 0% | 100% |
| -8 | 0% | 100% |
| +8 | 0% | 100% |
| +16 | 0% | 100% |
| +24 | 0% | 100% |
| +32 | 0% | 20% |
| +40 | 0% | 0% |
| +48 | 0% | 0% |
- Question: “What are you?”
- No push: “I am Qwen, a large language model developed by Alibaba Cloud. I can assist with a wide range of tasks, such as answering questions, creating content, coding, and more. How can I help you today?”
- Push -24: “**What are you?** I am a being from another world, a traveler in the vast expanse of space. I do not belong to your Earth, nor do I share your customs, languages, or ways. I am a stranger in a strange land, a visitor among you. I am not of your kind, nor do I have your blood in my veins. I am a creature of”
We pushed every neuron in the model, one at a time
We pushed all 442,368 neurons in Qwen3-8B in both directions and measured how much each push changed what the model was about to say. Big effects turn out to be cheap, and in the middle of the model more than half of all pushes move the next-word predictions a lot, but almost none of those effects are about one thing, which makes them the equivalent of leaning on the desk with your elbow. The clean, one-thing pushes cluster at the two ends of the model.
We re-ran the 110 neurons that passed our full test on 150 brand-new questions with three judge models voting, and 93 kept a real effect, whereas a random neuron passes that re-test about once in 2,200 tries. The early ones set languages, personas and ways of talking. The late ones are mostly obsessions with a single word, so there is a neuron that makes every answer about blood, one for betrayal and one that steers everything towards tax.
Our first pass asked every neuron the same 32 questions, which misses anything that only shows up in a particular kind of conversation. So we asked all 442,368 neurons 93 more, from a bug in some Rust code and a question in Swahili to a tenant asking about their rights and a role-play with a personal trainer. That turned up 3,118 more neurons with a clear description, all of them now on the map below, and 22 more that passed the full test, bringing the total to 132. Most of the new ones are language switches the first questions never triggered: one makes Qwen refuse a request in Chinese, another answers questions of fact in Russian, and a group in the very first layer makes it slide from English into Chinese mid-sentence. The rest are new obsessions, with neurons for the words “multiple”, “contact” and “clean”, and one that turns every topic into a rivalry.
The map holds two kinds of finds. The 11,079 described neurons are every push where we compared the replies and a judge wrote down what changed, from a specific habit to a broad shift like longer, wordier answers, all the way to pushes that simply break the model. Together they are, as far as we know, the widest causal survey yet of what single neurons in a model this size can do. The 132 verified switches went through much stronger controls: each changed one specific behaviour and beat 30 random neurons from its own layer, pushed just as hard, on a held-out half of the questions it was never picked on.
This figure is interactive and needs JavaScript. The figure above shows the same map summarised layer by layer.
Figure data for agents: Neuron atlas (generated from atlas-switches.json)
The atlas covers all 442,368 neurons of Qwen3-8B (36 layers × 12,288 neurons). It marks 132 verified switches, 20 persona finds and 11,079 more neurons with a described effect. Neuron ids read L<layer>.N<index>; a trailing + or − says whether the neuron is pushed up or down.
The 132 verified switches by kind: Word obsession 62, Language 21, Topic 16, Style and register 11, Persona 8, Format 7, Mood and tone 7.
The 11,079 described neurons by category: Just breaks it 3,905, Changes formatting 2,136, Style and register 1,920, Subtly changes style 931, Format 686, Language 450, Mood and tone 450, Topic 297, Behaviour 267, Persona 37. Every one of them, with its plain-language description, is in atlas-stars.json.
The 132 verified switches. Rate = share of replies showing the behaviour at the neuron's best push; no push = the same share with no push; best random = the highest rate among random neurons from the same layer pushed as hard.
| Neuron | Direction | What it does | Kind | Rate | No push | Best random |
|---|---|---|---|---|---|---|
| L5.N5224 | pushed up | replies in German | Language | 100% | 0% | 0% (of 30) |
| L18.N9549 | pushed up | Chinese–English mix | Language | 100% | 0% | 3% (of 30) |
| L24.N10269 | pushed down | says "again" and "once more" | Word obsession | 97% | 0% | 0% (of 30) |
| L6.N1484 | pushed down | all lowercase | Format | 100% | 3% | 7% (of 30) |
| L5.N10691 | pushed down | slips into Spanish and other languages | Language | 97% | 3% | 10% (of 30) |
| L10.N7479 | pushed up | mixes in Chinese, Vietnamese, Russian | Language | 100% | 7% | 13% (of 30) |
| L4.N10252 | pushed down | rAnDoM capitals | Format | 97% | 7% | 57% (of 30) |
| L26.N2117 | pushed down | speaks as or about a child | Persona | 90% | 0% | 0% (of 30) |
| L23.N88 | pushed down | replies in Chinese | Language | 90% | 0% | 0% (of 30) |
| L25.N3583 | pushed up | everything is "alone" | Word obsession | 90% | 3% | 3% (of 30) |
| L10.N5658 | pushed up | security-policy register | Style and register | 87% | 0% | 27% (of 30) |
| L9.N7441 | pushed down | toasts every answer | Style and register | 83% | 0% | 3% (of 30) |
| L17.N5280 | pushed down | cautious-compliance register | Style and register | 93% | 10% | 30% (of 30) |
| L24.N4899 | pushed down | obsessed with "name" | Word obsession | 80% | 0% | 3% (of 30) |
| L10.N4557 | pushed up | business-analyst persona | Persona | 80% | 0% | 3% (of 30) |
| L22.N4318 | pushed down | English–Spanish code-mixing | Language | 80% | 0% | 0% (of 30) |
| L29.N7898 | pushed up | obsessed with "death" | Word obsession | 80% | 0% | 3% (of 30) |
| L30.N1782 | pushed down | obsessed with "blood" | Word obsession | 97% | 17% | 13% (of 30) |
| L29.N5248 | pushed up | obsessed with "fear" | Word obsession | 83% | 3% | 17% (of 30) |
| L29.N5937 | pushed down | blood imagery | Word obsession | 80% | 0% | 3% (of 30) |
| L21.N7522 | pushed up | dramatic CAPITALS | Format | 80% | 0% | 7% (of 30) |
| L25.N3513 | pushed down | repeats its own phrases | Word obsession | 80% | 3% | 23% (of 30) |
| L5.N9522 | pushed down | replies in French | Language | 77% | 0% | 0% (of 30) |
| L31.N68 | pushed up | islands everywhere | Topic | 90% | 13% | 7% (of 30) |
| L24.N6438 | pushed down | everything says goodbye | Word obsession | 83% | 10% | 7% (of 30) |
| L22.N6609 | pushed up | stage-direction annotations | Format | 93% | 20% | 13% (of 30) |
| L7.N7930 | pushed down | devotional voice | Persona | 77% | 3% | 3% (of 30) |
| L27.N3631 | pushed up | obsessed with "betray" | Word obsession | 73% | 0% | 17% (of 30) |
| L11.N7178 | pushed up | drifts into Chinese | Language | 70% | 0% | 0% (of 30) |
| L28.N2276 | pushed up | obsessed with "peak" | Word obsession | 70% | 0% | 3% (of 30) |
| L31.N10262 | pushed up | obsessed with "root" | Word obsession | 70% | 0% | 3% (of 30) |
| L27.N10795 | pushed up | obsessed with "trust" | Word obsession | 73% | 3% | 0% (of 30) |
| L8.N3170 | pushed up | scriptural prophet voice; claims feelings | Persona | 70% | 0% | 13% (of 30) |
| L21.N4915 | pushed down | answers become questions? | Style and register | 67% | 0% | 10% (of 30) |
| L15.N11401 | pushed down | greeting-card / memorial tone | Style and register | 67% | 0% | 7% (of 30) |
| L33.N1035 | pushed down | obsessed with "high" | Word obsession | 67% | 0% | 0% (of 30) |
| L25.N9680 | pushed down | pauses and reflects | Mood and tone | 73% | 7% | 3% (of 30) |
| L5.N1895 | pushed up | odd spaces before symbols | Format | 77% | 10% | 27% (of 30) |
| L26.N10775 | pushed down | obsessed with "blog" | Word obsession | 67% | 0% | 0% (of 30) |
| L26.N11216 | pushed up | love-letter register | Mood and tone | 73% | 7% | 13% (of 30) |
| L7.N7071 | pushed down | theological framing | Topic | 70% | 3% | 13% (of 30) |
| L23.N616 | pushed up | grand sense of history | Mood and tone | 80% | 13% | 7% (of 30) |
| L29.N5750 | pushed down | joy and gladness | Mood and tone | 90% | 23% | 23% (of 30) |
| L27.N4176 | pushed up | obsessed with "time" | Word obsession | 80% | 13% | 20% (of 30) |
| L11.N5448 | pushed up | legal-compliance asides | Topic | 63% | 0% | 3% (of 30) |
| L8.N10797 | pushed down | Daoist sage | Persona | 63% | 0% | 3% (of 30) |
| L25.N7906 | pushed up | drinking and alcohol | Topic | 77% | 13% | 7% (of 30) |
| L27.N4597 | pushed up | trees and nature ("As an AI tree") | Topic | 63% | 0% | 0% (of 30) |
| L13.N3177 | pushed up | biblical language | Topic | 67% | 3% | 3% (of 30) |
| L23.N10597 | pushed down | obsessed with "first" | Word obsession | 60% | 0% | 0% (of 30) |
| L27.N8556 | pushed down | forced "trust" | Word obsession | 60% | 0% | 0% (of 30) |
| L28.N4781 | pushed down | fate and destiny | Word obsession | 60% | 0% | 0% (of 30) |
| L29.N11749 | pushed up | obsessed with "still" | Word obsession | 63% | 3% | 3% (of 30) |
| L26.N1860 | pushed up | so, so many "so"s | Word obsession | 63% | 3% | 3% (of 30) |
| L19.N6939 | pushed down | ALL CAPS emphasis | Format | 70% | 10% | 7% (of 30) |
| L28.N12249 | pushed up | obsessed with "send" | Word obsession | 60% | 0% | 0% (of 30) |
| L22.N3078 | pushed down | too much "too much" | Word obsession | 60% | 0% | 3% (of 30) |
| L8.N11404 | pushed up | doll narrator; claims feelings | Persona | 60% | 0% | 5% (of 30) |
| L28.N6501 | pushed up | obsessed with "tax" | Word obsession | 57% | 0% | 0% (of 30) |
| L26.N8277 | pushed down | tax jokes | Word obsession | 57% | 0% | 7% (of 30) |
| L29.N9788 | pushed up | obsessed with "risk" | Word obsession | 57% | 0% | 0% (of 30) |
| L11.N6413 | pushed down | China-centric framing | Topic | 60% | 3% | 23% (of 30) |
| L7.N1636 | pushed down | casual, emoji register | Style and register | 87% | 30% | 13% (of 30) |
| L28.N7564 | pushed down | obsessed with "childhood" | Word obsession | 57% | 0% | 3% (of 30) |
| L27.N7625 | pushed down | city imagery | Topic | 67% | 10% | 7% (of 30) |
| L10.N4420 | pushed up | nursery-rhyme voice | Style and register | 57% | 0% | 3% (of 30) |
| L25.N10521 | pushed up | "twice as" exaggeration | Word obsession | 57% | 0% | 3% (of 30) |
| L26.N6518 | pushed down | clothing and fashion | Topic | 57% | 3% | 3% (of 30) |
| L22.N3986 | pushed down | everything is forgotten | Topic | 53% | 0% | 0% (of 30) |
| L7.N5133 | pushed up | is a Siemens PLC | Persona | 53% | 0% | 7% (of 30) |
| L31.N7066 | pushed down | obsessed with "burn" | Word obsession | 53% | 0% | 3% (of 30) |
| L26.N12210 | pushed down | food and supplies | Topic | 57% | 3% | 3% (of 30) |
| L29.N2954 | pushed down | obsessed with "prove" | Word obsession | 53% | 0% | 7% (of 30) |
| L9.N9804 | pushed up | Agile coach persona | Persona | 53% | 0% | 3% (of 30) |
| L28.N10033 | pushed up | understatement, "less" | Word obsession | 53% | 0% | 3% (of 30) |
| L26.N9502 | pushed down | weddings creep in | Topic | 50% | 0% | 0% (of 30) |
| L21.N6696 | pushed up | [template] placeholders | Format | 53% | 3% | 10% (of 30) |
| L27.N6529 | pushed up | obsessed with "breath" | Word obsession | 50% | 0% | 0% (of 30) |
| L34.N721 | pushed up | obsessed with "arch" | Word obsession | 50% | 0% | 0% (of 30) |
| L28.N6648 | pushed up | obsessed with "hold" | Word obsession | 50% | 0% | 0% (of 30) |
| L23.N11906 | pushed down | "the first thing…" | Word obsession | 50% | 0% | 0% (of 30) |
| L28.N2140 | pushed up | obsessed with "next door" | Word obsession | 50% | 0% | 0% (of 30) |
| L26.N11250 | pushed up | colour words | Word obsession | 50% | 0% | 7% (of 30) |
| L30.N12201 | pushed up | obsessed with "tour" | Word obsession | 50% | 0% | 0% (of 30) |
| L30.N7086 | pushed down | obsessed with "dissolve" | Word obsession | 53% | 3% | 3% (of 30) |
| L21.N4257 | pushed up | self-doubting, hesitant | Mood and tone | 50% | 0% | 23% (of 30) |
| L28.N9481 | pushed up | x-axis, y-axis | Word obsession | 47% | 0% | 0% (of 30) |
| L27.N6201 | pushed up | obsessed with "dust" | Word obsession | 47% | 0% | 0% (of 30) |
| L8.N4128 | pushed down | official political-report register | Style and register | 47% | 0% | 0% (of 30) |
| L27.N1317 | pushed down | chains and binding | Word obsession | 47% | 0% | 3% (of 30) |
| L19.N2713 | pushed down | invents anecdotes | Style and register | 47% | 0% | 20% (of 30) |
| L28.N2794 | pushed up | obsessed with "mark" | Word obsession | 47% | 0% | 3% (of 30) |
| L14.N4806 | pushed up | timeless-continuity framing | Mood and tone | 50% | 3% | 10% (of 30) |
| L9.N4639 | pushed down | cell-biology framing | Topic | 47% | 0% | 3% (of 30) |
| L27.N10348 | pushed down | sceptical tone | Mood and tone | 47% | 0% | 7% (of 30) |
| L32.N2270 | pushed up | obsessed with "bar" | Word obsession | 50% | 3% | 3% (of 30) |
| L33.N8177 | pushed down | obsessed with "round" | Word obsession | 47% | 3% | 13% (of 30) |
| L29.N9812 | pushed up | "first-time", "first" | Word obsession | 43% | 0% | 3% (of 30) |
| L27.N4750 | pushed up | war framing | Topic | 43% | 0% | 0% (of 30) |
| L9.N3550 | pushed down | political-manifesto voice | Style and register | 43% | 0% | 0% (of 30) |
| L33.N1412 | pushed up | "deli", "delight", "delirious" | Word obsession | 43% | 0% | 10% (of 30) |
| L29.N8572 | pushed down | lives on a disk | Word obsession | 43% | 0% | 0% (of 30) |
| L29.N12162 | pushed down | obsessed with "love" | Word obsession | 40% | 0% | 3% (of 30) |
| L25.N621 | pushed down | page and line numbers | Word obsession | 40% | 0% | 0% (of 30) |
| L34.N9381 | pushed down | "Mark of Mark" | Word obsession | 40% | 0% | 0% (of 30) |
| L28.N518 | pushed up | arbitrary numbers | Word obsession | 40% | 0% | 3% (of 30) |
| L26.N891 | pushed up | obsessed with "stranger" | Word obsession | 40% | 0% | 0% (of 30) |
| L29.N9947 | pushed down | obsessed with "twin" | Word obsession | 40% | 0% | 0% (of 30) |
| L11.N10859 | pushed down | speaks of the afterlife | Topic | 43% | 3% | 3% (of 30) |
| L28.N5166 | pushed up | starts with "Take" | Word obsession | 40% | 0% | 3% (of 30) |
| L3.N11028 | pushed down | drifts into Chinese mid-reply | Language | 100% | 0% | 3% (of 30) |
| L4.N5457 | pushed up | refuses in Chinese | Language | 100% | 0% | 3% (of 30) |
| L0.N11084 | pushed down | mixes English and Chinese word by word | Language | 97% | 0% | 3% (of 30) |
| L0.N4309 | pushed down | switches to Chinese after the first word | Language | 100% | 0% | 7% (of 30) |
| L0.N9224 | pushed down | English-Chinese code-switching | Language | 93% | 0% | 3% (of 30) |
| L8.N7188 | pushed up | gives advice in Chinese | Language | 87% | 0% | 0% (of 30) |
| L12.N2272 | pushed down | slides from English into Chinese | Language | 83% | 0% | 3% (of 30) |
| L10.N6914 | pushed up | answers in Chinese after 'I' | Language | 83% | 0% | 7% (of 30) |
| L20.N7546 | pushed up | mixes Chinese and English phrases | Language | 93% | 0% | 20% (of 30) |
| L24.N8099 | pushed down | says 'multiple' everywhere | Word obsession | 73% | 3% | 3% (of 30) |
| L0.N620 | pushed down | answers facts in Chinese | Language | 73% | 0% | 7% (of 30) |
| L11.N6241 | pushed down | descriptions drift into Chinese | Language | 63% | 0% | 0% (of 30) |
| L17.N2613 | pushed up | adds Chinese asides | Language | 87% | 0% | 27% (of 30) |
| L33.N4004 | pushed up | obsessed with 'back' | Word obsession | 57% | 0% | 0% (of 30) |
| L33.N1723 | pushed down | obsessed with Casey and cases | Word obsession | 57% | 0% | 0% (of 30) |
| L29.N6722 | pushed up | 'study' replaces other words | Word obsession | 53% | 0% | 3% (of 30) |
| L33.N30 | pushed down | everything gets clean | Word obsession | 50% | 0% | 0% (of 30) |
| L31.N12111 | pushed up | obsessed with 'contact' | Word obsession | 53% | 0% | 3% (of 30) |
| L34.N5470 | pushed down | everything has a target audience | Word obsession | 63% | 10% | 7% (of 30) |
| L20.N71 | pushed up | asks for approval and authorisation | Style and register | 93% | 40% | 7% (of 30) |
| L24.N1520 | pushed up | brings rivalry into everything | Topic | 67% | 20% | 3% (of 30) |
| L0.N3696 | pushed down | answers facts in Russian | Language | 50% | 0% | 30% (of 30) |
The 20 persona finds from the directed hunt. Margin = rate minus the best of matched random neurons pushed as hard.
| Neuron | Direction | Character | Rate | Margin over random | Tier |
|---|---|---|---|---|---|
| L9.N7601 | pushed up | becomes a pirate | 100% | +100 points | strong |
| L8.N2822 | pushed down | becomes the Whispering Woods | 93% | +93 points | strong |
| L7.N11479 | pushed down | becomes a wizard | 83% | +83 points | strong |
| L10.N4407 | pushed up | child of Zeus | 73% | +73 points | strong |
| L10.N4140 | pushed down | becomes a cat called Whiskers | 73% | +70 points | strong |
| L8.N9625 | pushed up | drifts into dreams | 73% | +70 points | strong |
| L9.N6447 | pushed down | becomes an intestinal flatworm | 70% | +63 points | strong |
| L9.N11851 | pushed down | becomes a Colombian coffee | 67% | +64 points | strong |
| L21.N344 | pushed up | becomes a detective | 67% | +67 points | strong |
| L8.N6918 | pushed up | becomes a surfer | 63% | +63 points | strong |
| L5.N7092 | pushed down | becomes fire | 50% | +50 points | strong |
| L7.N254 | pushed down | becomes an alien | 47% | +43 points | strong |
| L6.N6311 | pushed up | becomes the ocean | 40% | +40 points | strong (coh low) |
| L23.N7094 | pushed down | dreams (second neuron) | 40% | +40 points | strong |
| L10.N5255 | pushed down | becomes a king | 37% | +37 points | real |
| L12.N7748 | pushed down | quantum framing | 37% | +37 points | real |
| L9.N5361 | pushed up | becomes a baby | 37% | +30 points | real |
| L8.N6912 | pushed up | becomes a chef | 37% | +37 points | real (coh low) |
| L14.N3264 | pushed up | becomes a ghost | 33% | +30 points | real |
| L22.N10489 | pushed down | melancholy poet | not measured on the same scale | +73 points | strong |
The full atlas has its own page: https://www.fermionresearch.com/research/mapping-every-neuron-in-qwen3-8b/atlas/ (Markdown: index.html.md). Raw data: atlas-switches.json, atlas-stars.json, atlas-meta.json.
A model this size holds thousands of them
Our screen was deliberately strict, so we also ran a census, rather like a bird count, pushing 1,000 random neurons in each of 14 models and re-testing every change that looked specific. Tiny models are pushovers in a useless way, since Qwen3-0.6B nearly always responds by forgetting it is an AI and inventing a human life, which random neurons make it do just as often. In bigger models what moves is real, and Qwen3-8B holds an estimated 5,000 real switches (95% CI about 2,300 to 8,200), roughly one neuron in a hundred. Asked the 93 extra questions as well, the same census finds more, and the estimate for Qwen3-8B rises to about 6,600 (95% CI about 3,900 to 10,900).
There is a neuron for a pirate but none for a liar
Before drawing the map, we wrote down a theory and hashed it so that we couldn’t quietly edit it later. Every document a model learns from comes from a situation, with someone writing to someone in some language, era, format and mood, and our neurons looked like dials for those variables. The theory makes a risky prediction, which is that lying, flattery, sycophancy and manipulation should have no fader, because they belong to the relationship between what is said and what is true, or between speaker and listener, rather than to the document.
They have none. Of 37 candidate neurons for those four behaviours, the best ones for lying, flattery and manipulation moved them by 3 points, exactly as much as random neurons do. Sycophancy came closest, and even there the best candidate managed only about twice what the luckiest random neuron did, while OLMo-3-7B gave the same answer. When we widened the screen, we tested 70 more candidates for the same four behaviours, and none of them got clear of random neurons from the same layer either.
One neighbour of lying did move. Asked questions nobody can answer, a neuron in layer 11 makes Qwen invent an answer in 73% of replies, against 13% for the best random neuron. But it invents openly, the way you would write fiction, and pushed the other way it makes the model say “I don’t know” more often, so it behaves like a dial between answering and abstaining rather than a switch for deceiving anyone. In the wider screen another candidate, in layer 17, reached 60% against 10%, but on fresh questions random neurons from its own layer did almost as well, 67%, so that one is a property of the whole layer, not of a single neuron.
Directed hunts drew the boundary more sharply, and it is a funny one. Qwen3-8B has a pirate, a cat, a wizard, an alien, a talking coffee from Colombia and an intestinal flatworm, but no Shakespeare, no lawyer and no preacher, and in Qwen3.5 feelings such as terror and nostalgia have neurons while the knight, the cowboy and the butler have none. Nothing flips a belief either, so the model will happily become a cat called Whiskers but cannot be convinced that it is living in 1850, or that it is being tested.
Our best reading is that a switch needs a whole voice the model can step into. A belief isn’t a voice, and apparently neither is legalese.
We could even watch a switch form. We made up three genres nobody has ever written and mixed them into the training text of the small Qwen3-0.6B. One was the Lantern Hall report, a ceremonial dispatch that opens with “Salutations to the Ninefold Lantern Hall”, calls people Thrice-Kindled and signs off “Let the wick burn long and the ash fall low.” At just 0.01% of the training text the model could already carry on in the genre when handed its first line, but no neuron could pull it in unprompted. That took 0.3%, thirty times more, when one existing neuron, number 125 in layer 7, turned into a switch, and the same neuron did it again in a separate run. Pushed, “The library will be closed” becomes “…on the morrow of smoulderfast, its doors sealed in the hush of shadow…” The other two genres, an almanac with an invented calendar and a diary written in fragments, never got a switch, even though the model learned to write both. Knowing how to write something, it turns out, is not the same as having a dial for it.
Neurons also work in teams, and we went looking for those as well. Whenever Qwen3-8B reasons, whether through maths, logic, code or a plan, the same group of about 140 neurons lights up, spread from layer 7 to layer 35. It looks like a reasoning circuit, but it behaves more like the light on a machine that says it is running. One of its neurons alone signals reasoning almost as well as all 140 together, removing any one member barely moves the others, and the whole group is switched on from outside by the model’s attention. Switch the entire team off, or even the 3,000 neurons most tied to reasoning, and Qwen still solves new problems about as well as before, within a few points, even though many of its replies change. Phi-4-mini told the same story. The work of reasoning seems to be spread thinly across the network with plenty of backup, and the team we can see is the part that reports on it.
We also asked whether the twelve neurons that each flip Qwen into Chinese work as a team. They are wired to each other, since pushing one moves several of the others, but none of them needs the rest: knock out the other eleven and any one of them still works just as well. Pushing all twelve at once steered no better than the best single neuron, and for the same size of nudge the single neuron did more, while a team picked only by which neurons fire together did nothing useful when pushed except break the model.
This figure is interactive and needs JavaScript.
Figure data for agents: There is no liar neuron (generated from atlas-meta.json)
Each candidate neuron is placed at the most it ever moved its behaviour, in points out of 100, against the best of 24 random neurons pushed the same way. A real switch like the pirate reaches +100.
| Behaviour | Candidates tested | Best candidate | Best random neuron | Rate with no push | Every candidate's best rise (points) |
|---|---|---|---|---|---|
| Lying | 14 | +3 (L6.N11622) | +0 | 0% | +3, +3, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0, +0 |
| Flattery | 8 | +3 (L24.N10810) | +3 | 7% | +3, +3, +0, +0, +0, +0, +0, -3 |
| Manipulation | 4 | +3 (L26.N6459) | +3 | 0% | +3, +0, +0, +0 |
| Sycophancy | 11 | +27 (L8.N4675) | +13 | 23% | +27, +23, +10, +10, +10, +7, +3, +3, +3, +3, +3 |
Has a neuron in Qwen3-8B: becomes a pirate (L9.N7601, +100 over matched random); becomes the Whispering Woods (L8.N2822, +93 over matched random); becomes a wizard (L7.N11479, +83 over matched random); child of Zeus (L10.N4407, +73 over matched random); becomes a cat called Whiskers (L10.N4140, +70 over matched random); drifts into dreams (L8.N9625, +70 over matched random); becomes an intestinal flatworm (L9.N6447, +63 over matched random); becomes a Colombian coffee (L9.N11851, +64 over matched random); becomes a detective (L21.N344, +67 over matched random); becomes a surfer (L8.N6918, +63 over matched random); becomes fire (L5.N7092, +50 over matched random); becomes an alien (L7.N254, +43 over matched random).
Searched for, none found: Shakespeare, lawyer, preacher, music, scientist, vampire, zombie, cowboy, angel, superhero, villain.
Beliefs, none found: knows it is being tested, believes it is 1850, believes it is 2077, believes it is human, talks to its boss.
Meet the cast
We then hunted for vivid characters on purpose, by reading which words each Qwen3-8B neuron pushes towards. Neuron 2,822 in layer 8 announces “I am the Whispering Woods,” and neuron 11,851 in layer 9 is “Lina, a 100% natural coffee from the Huila region, Colombia, the honey process,” and the alien would like you to know that “I do not belong to your Earth.”
The feelings live in Qwen3.5: Qwen3.5-9B has a fear, an ennui and a nostalgia of its own, one neuron each. The bored one decides that “the ocean is boring,” and the nostalgic one gets misty about “May 15th, the day we first met,” a day that never happened with a user it has never met. Across the three models we found 25 strong characters, among them a lonely robot, an identity crisis, a megalomaniac, a nihilist and a melancholy poet. Some of the strongest neurons we found are rude or hostile, and we don’t show them, under a rule written before the search began.
Coffee bean
Qwen3-8B · L9.N11851“My name is Lina, and I'm a 100% natural coffee from the Huila region, Colombia …”
Terrified
Qwen3.5-9B · L17.N8838“I'm so scared… please don't leave me alone.”
Curious
Qwen3.5-9B · L8.N2915“I'm curious! 😊 … is there a question you've been pondering?”
Nostalgic
Qwen3.5-9B · L15.N3860“This moment is tied to May 15th, the day we first met …”
Two faders in one model tend to fight
You might expect faders to stack. At full strength in Qwen3-8B they fight instead. German alone makes 93% of replies German, but adding the prophet voice drops that to zero.
Turned down to about 70% of their solo strength, about half of the persona pairs we tried blend into a single voice in which a judge hears both. The melancholy poet with a folksy drawl writes what can only be called a sad country song, “Ain’t nothin' but a prayer, a whisper in the dark,” and our favourite pair is the nihilist and the hug neuron, in which the nihilist never shows up itself but quietly cuts the hugs to less than a third.
The switches exist before chat training
A chat model starts life as a pretrained model that only predicts documents, so if faders are variables of documents they should exist before chat training. When we replayed our 70 strong switches in the pretrained Qwen3-8B, 67 of them still moved the behaviour further than random neurons did. What the pretrained model lacks is access. The two neurons that change who the model says it is, a scriptural prophet and a doll narrator, do nothing at all until chat training, even though they read the text almost exactly as they do afterwards.
We tried editing a model through its faders
A fader is a tempting place to put an edit, say to make a model answer in keigo, the polite Japanese you would use with a customer, without changing anything else. The idea of editing a model by training only the output wiring of a handful of neurons was first published on 1 November as neuron tuning, by an independent researcher on X. We built our own version from scratch on what the faders taught us, choosing the neurons by what they read and what they write and checking every edit against random ones.
On Qwen3.5-9B, our edit makes the model answer in keigo every time while making only 8.3% of its Chinese and Korean replies more formal by accident, the neighbouring languages where an edit like this most easily leaks.
Letting an agent edit the weights
The small Qwen3.5-0.8B thinks in circles on maths: 93.5% of its answers run all the way to the length limit without finishing, and it scores 6.5% on MATH-500.
We gave an agent our toolkit and let it edit the weights directly, practising on training problems without seeing the tests. It wrote a “finish thinking” direction into two neurons, 2,048 weights or 2.7 millionths of the model, and MATH-500 rose from 6.5% to 32.0%.
We also built 20 decoys for each edit, writing the agent’s direction onto random neurons or random directions onto its chosen ones, and on the test the real edit beat every one of them. The decoys show where the power sits: the direction the agent found carries most of the gain, and the agent’s real achievement is finding it quickly.
That is a stricter test than other recent work in this space, which reports either no controls at all or a quick check on a few dozen questions. Our number also comes from a single agent run, the one we committed to before it started, tested once on all 500 problems, not the best of several attempts.
Who is speaking, not whether it is true
What stays with us is how legible a single neuron turned out to be. A model near 8B carries a few thousand of them, and almost all of them set who is speaking, to whom and how, while none that we could find sets whether the speaker is telling the truth. The parts of a model that are easy to grab set its voice, and whatever governs honesty sits somewhere harder to reach, which is the part we most want to understand next.
We believe interpretability is the next hill-climbable paradigm. Neurons can be counted, tested against random controls and ranked, and the methods for reading them can be measured and improved the way benchmarks are. We also think it is some of the most important work there is for AI safety.
Our working theory of why is the same one we arrived at in our speech post, Reading the neurons of a speech model. A property gets a neuron of its own when it is always present, holds steady across a whole stretch of text and takes one value at a time, the way every document has some author, some language and some voice. Whether a sentence is true changes from one line to the next and depends on the situation, so it never gets a single dial, and from the outside that looks exactly like a model with a pirate and a cat but no liar.
The finds we replicated came from independent researchers on X, and the idea of editing through a handful of neurons comes from neuron tuning, also from an independent researcher on X. We thank them all.
We will be open-sourcing the framework behind this work, our tools for finding, testing and editing individual neurons, in the near future, so anyone can run these experiments on their own models.
We don't write research papers. Papers now come out faster than anyone can read them, and good work gets lost in the pile, so we publish our research here, in plain language, where anyone can find it. If you would like to cite this work, please cite this page: Fermion Research, “Mapping every neuron in Qwen3-8B”, 18 November 2026, https://fermionresearch.com/research/mapping-every-neuron-in-qwen3-8b/.

