commas inside short 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. 16 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 reacts to commas that separate items in a short list of three or four things, such as 'motivation, confidence, arousal'. Its strongest negative cases are also commas in lists, so the pattern is not fully clear. Turning it up gave unrelated text.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
pertaining to the four active branches of the military (Army, Navy, Marines, Air Force); 2) Examining current tobacco policies and practices in
-
That's<|endoftext|>In applied sport psychology, there are four mental areas that are considered to be central for athletes to be mentally prepared for athletic success: motivation, confidence, arousal, and concentration.
-
' are trademarks of The Financial Times Ltd.⏎We live in an era of big data – often characterised in terms of the three “v”s: volume, velocity and variety.⏎Three simple examples illustrate the phenomenon. Volume – Tesco has
-
0 annotated summer driving thermal images captured at day and night and their corresponding RGB imagery for reference. The dataset includes classification of five groups; people, dogs, cars, bicycles
-
wine, he commanded the seven eunuchs who served him – Methuan, Biztha, Harbon, Bigtha, Anagtha, Zethat and Karkas – to bring before him Queen Vashti,
-
is controlled by the ‘analogWrite’ function.⏎Do RGB LEDs need PWM?⏎an RGB LED is a combination of three LEDs in just one package: red, green and blue; there are two kinds of RGB LEDs: common cathode and
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: The, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, men were believed to grow beards for different reasons than men do today. The three main reasons prehistoric men had beards were for warmth, intimidation, and to defend their face from hits. In Egyptian times, men would dye their
-
this blog is not intended to be a substitute for professional medical advice, diagnosis, or treatment. Always seek the advice of qualified health providers with questions you may have
-
a small incision in your belly. This procedure is generally performed under general anesthesia in a hospital.⏎- MRI. This test uses a large magnet, radio waves, and a computer to make 2-D images of an internal
-
Me" (NBC), "Ally McBeal" (FOX), "That '70s Show" (FOX), "The Simpsons" (FOX), and "Dawson's Creek" (FOX). Taylor said these programs were
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
- L13:1066 strongly suppresses
- L15:2500 excites 'and' closing a list of items
- L15:941 excites
- L10:667 suppresses commas inside lists of names
- L15:2598 suppresses
- L15:2999 suppresses
- L15:1029 excites comma in three-item lists
- L14:35 suppresses
- L15:137 suppresses
- L15:132 excites
- L14:449 excites
- L8:1605 excites
- L14:2531 excites
- L15:345 suppresses
- L13:1296 excites
- L14:2657 suppresses
- L9:77 excites commas in citations and name lists
- L15:274 excites
- L13:557 excites
- L14:907 excites
- L9:2447 suppresses
- L3:372 excites
- L12:2723 weakly suppresses
- L13:269 weakly suppresses
Feeds later layers
- L19:492 excites
- L17:1802 excites
- L17:2654 excites
- L21:393 suppresses unclear: relative 'that' and list items
- L20:1009 excites
- L20:882 suppresses
- L19:1089 suppresses
- L17:572 weakly excites
- L27:358 weakly suppresses
- L19:628 weakly excites
- L18:2921 weakly excites period ending a sentence that announces a list
- L26:236 weakly suppresses
- L17:2610 weakly excites and or comma closing a list
- L17:919 weakly suppresses
- L18:1533 weakly suppresses
- L17:878 weakly excites
- L18:2863 weakly suppresses
- L20:282 weakly suppresses
- L21:154 weakly excites
- L26:22 weakly excites the 'res' fragment inside words
- L20:2864 weakly excites
- L20:2429 weakly excites
- L24:435 weakly suppresses
- L23:673 weakly excites
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
- [/
- lak
- adians
- AndPassword
- ain
- elsius
- andre
- pou
- above
- liness
Suppresses
- ially
- nels
- dict
- urally
- quoi
- Vers
- vely
- tps
- 至
- 箱