'and so on' list endings
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. 22 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 word that leads into 'and so on' at the end of a list, such as 'and so' or 'and on'. Turned down, the model wrote long strings of dots, which is the opposite of how it normally ends a list.
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.
-
instance, if you are thinking of the number 1, it would be "A". Number 2 would be "B". 3 is "C", and so on.⏎- Now, think of a country that starts with the letter you
-
together. This skill is important in other areas of their lives as they continue to grow, such as writing, tying shoes, typing and so much more⏎2)Fine Motor Coordination– In order to draw shapes,
-
is so humble you could argue it doesn’t really have its own name. After all, do its various monikers: lavatory, bog, khazi and so on, denote the room or the device? Or both? It can be hard
-
question of priority (individual or community) often has the largest impact in the most pressing ethical questions: health care, abortion, multiculturalism, hate speech, and so on. . . .⏎Beginning in the late 20th century, many
-
ox, hepatitis, chicken pox, pneumonia, and measles, and so had never developed any immunityImmunity⏎Resistance to disease, developed naturally or through
-
among many others.⏎An academic journal also publishes different research types. Some of these include original articles, reviews, book reviews, case studies, short communications, and so on. Usually, bigger journals publish a wider variety of research works.⏎One vital
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: ................................................................................................................................................
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
tested how individuals group rhythms by playing simple sequences of tones. For example, listeners are presented with tones that alternate in loudness (...loud-soft-loud-soft...) or duration (...long-short-long-short...) and are asked to indicate their perceived
-
center activities!⏎- Engaging morning work!⏎- Homework variety!⏎- Individual pages for facts 1-12 (-1, -2, -3, etc. up to -12)⏎- Mixed facts practice pages⏎- Answer
-
For instance, if you are thinking of the number 1, it would be "A". Number 2 would be "B". 3 is "C", and so on.⏎- Now, think of a country that starts with the letter
-
from a study of statistics.⏎Why can a yield never be modelled exactly by a normal⏎How are the sums of 2, 3, 4, ... identical normal⏎Why is the mean target of 8000 tonnes
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
- L18:2722 strongly suppresses
- L17:2866 strongly suppresses
- L19:1265 strongly suppresses
- L21:2493 strongly excites
- L18:2185 suppresses
- L18:2938 excites
- L18:1118 excites
- L17:2209 excites
- L17:2664 excites
- L16:1603 suppresses
- L17:902 excites
- L20:1620 excites
- L16:2542 suppresses
- L18:2305 excites
- L21:2512 excites
- L16:685 excites
- L14:564 suppresses
- L19:2164 suppresses
- L13:2740 excites
- L15:2255 excites
- L17:2578 excites
- L11:1527 excites
- L19:1246 excites
- L15:2610 excites
Feeds later layers
- L27:426 strongly suppresses
- L25:454 strongly suppresses
- L25:717 strongly suppresses
- L25:1326 strongly suppresses ellipsis dots after a cut-off quote
- L26:295 strongly suppresses
- L26:472 suppresses
- L27:502 suppresses
- L27:764 suppresses
- L23:641 suppresses
- L27:1536 weakly suppresses
- L26:943 weakly suppresses
- L23:2015 weakly excites
- L27:1182 weakly suppresses
- L27:606 weakly excites
- L24:2694 weakly excites
- L26:422 weakly excites
- L26:417 weakly suppresses
- L27:475 weakly suppresses
- L27:1899 weakly suppresses
- L23:819 weakly suppresses
- L25:1135 weakly suppresses
- L26:369 weakly suppresses
- L23:40 weakly suppresses
- L23:260 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
- #",
- 呙
- %%*/
- ipel
- ')['
- ancybox
- 何况
- )'],⏎
- )','
- ']."'
Suppresses
- ...
- ...
- …
- ……
- ...,
- ....
- ......
- ...</
- ..."
- ...⏎