commas in long lists
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 10 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron fires on the comma or semicolon that separates items in a long list, such as a list of states, angels, or subjects. Turning it up made the model's text drift toward unrelated topics.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
ians use this pattern so often that we have created a name and a symbol for it. 8×7×6×5×4×3×2×1 is written as 8! and we call it "eight factorial." You
-
:⏎- Communicate effectively using Markel’s Eight Measures of Excellence: honesty, clarity, accuracy, comprehensiveness, accessibility, conciseness, professional appearance, and correctness.⏎- Evaluate the communication situation: audience, purpose, and context.⏎
-
Society and Education; Compulsory Education and the Transmission of Culture; Sexuality and Education; Race; Multiculturalism, and Education; Social Class and Education; and Technology and Education.⏎- Stimulates critical thinking: The editor includes readings that
-
the acronym Family ASSESSMENT—Structure, Socioeconomic, Sickness, Nutritional, Educational, Environmental, Affordability, System (health), Mental, Family influence, Threats. Under each domain,
-
of Angels are as follows: godly Thrones, six-winged Seraphim, many-eyed Cherubim, Dominions, Powers, Authorities, Principalities, Archangels and Angels. Within these nine orders there exists three ranks: Ser
-
in the three-day meeting were 52 young scientists from 18 countries (Argentina, Australia, Bangladesh, Brazil, Canada, France, Germany,Ghana,India, Italy, Nigeria, Poland, Portugal, Russia, Switzerland, United Kingdom
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: what is the main issue you are trying to ask? I know that I am writing this question in a forum, so maybe you are the one who is the one who is asking. I am writing in a different language, but I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, 07:44 AM⏎We are not going to do you homework.⏎If you want our help, you have to ask more specific what you need help with.⏎Have you looked at the macro that have been given to you?⏎
-
akra system (spinning vortexes or wheels of light)? Would you like to know what each chakra represents? Here is a simple diagram and list for
-
garden.⏎‘What are you doing here?’ he cried in a very gruff voice, and the children ran
-
and is moved by a paddle with two blades⏎Headscratcher for kayak⏎Kayak is spelled the same forward and backward.⏎Seen and Heard⏎What made you want to look up kayak? Please tell us where you read or heard it (
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L9:48 excites
- L8:1856 suppresses
- L8:355 excites
- L7:676 excites
- L7:2215 excites
- L9:1093 suppresses
- L7:1934 weakly excites
- L7:1464 weakly excites
- L5:108 weakly suppresses
- L7:169 weakly excites
- L7:2062 weakly excites
- L6:2879 weakly excites
- L2:413 weakly excites
- L8:1444 weakly excites
- L9:5 weakly excites
- L6:834 weakly suppresses
- L6:1651 weakly excites
- L8:2997 weakly suppresses
- L5:1103 weakly excites
- L7:1267 weakly suppresses
- L4:1855 weakly suppresses
- L9:373 weakly excites
- L9:50 weakly suppresses
- L6:1563 weakly excites
Feeds later layers
- L12:1757 suppresses
- L13:2781 suppresses
- L12:2214 excites
- L13:1887 suppresses
- L11:3067 suppresses
- L12:654 suppresses unclear: article ends and dateline openers
- L11:1691 suppresses
- L13:1487 excites
- L15:2083 suppresses
- L11:1900 weakly excites
- L15:198 weakly suppresses
- L11:1367 weakly excites
- L13:2559 weakly excites
- L17:2387 weakly suppresses
- L13:226 weakly excites
- L17:1736 weakly suppresses
- L11:1488 weakly excites
- L16:2607 weakly suppresses
- L16:2426 weakly suppresses
- L12:1671 weakly suppresses
- L11:2016 weakly excites
- L12:2938 weakly suppresses
- L16:2222 weakly excites
- L11:2457 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- icie
- icia
- icken
- crawler
- pherd
- 好看的
- าว
- Stoke
- SpringApplication
- isce
Suppresses
- wants
- 肼
- q
- 想
- み
- osity
- 问
- Meer
- CActive
- داء