comma after an introductory phrase
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. 18 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 wakes up on a comma that comes right after a short opening phrase, such as 'In some people,' or 'For many organizations,' where the main sentence is about to start. When it was turned up, the model's text drifted into unrelated, rambling sentences.
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.
-
And it will move you into alignment with Judaism’s most noble values.⏎The Origin of Epidemics⏎A little history is in order.⏎Throughout most of human history, there were no epidemic diseases.⏎“No one got the flu, not even the common
-
In some people, however, cells start to ignore insulin, which means the pancreas must manufacture more
-
1764, and he was the first person officially authorized to practice law in Georgia.⏎From 1777 until 1868, when the Constitution of 1868 made the name "attorney general
-
, Cultism, and Multiculturalism. Multicultural Education, 8(3), 2-14.<|endoftext|>Data is life⏎For many organizations, data is the lifeblood. It’s what keeps the business kicking and breathing.
-
2% to 6% of schoolchildren, mostly adolescent girls. It is most commonly diagnosed between ages 13 and 15.⏎In some children, symptoms begin after a triggering event, such as physical trauma, surgery, infection or
-
shaped using bronze tools, then smoothed with sand in order to fit tightly together. No mortar was used in the construction. Even after many centuries of wear and tear, the precision is amazing; It would be impossible to slide a piece of paper between
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: I have I, I am, I have, I have, I have, I have, I I I I I I I I I I I I I I I I I I I I I I I I I I I I I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
prompt asks you for an extension to be appended to the rectified images.⏎The next prompt will ask you whether a camera angle map should be produced and if yes, what should be its name.⏎After that you are asked if overwriting existing maps
-
and specific mental ability tests are important tools for vocational guidance. Researchers are now asking whether performance on such tests is based on differences in brain structure, and if so, can brain scans
-
- but look at the C++-tagged posts in this blog for examples of what the compiler thinks isn't a bad idea.)⏎if (x=0) init_subsystem();⏎So lots of books have been written about how to write code
-
are affecting data quality? If so, what sources?⏎- What makes a duplicate a duplicate?⏎- Should records be merged
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
- L17:1426 excites
- L15:280 weakly excites
- L9:1579 weakly suppresses
- L6:2843 weakly suppresses
- L0:561 weakly suppresses
- L16:1886 weakly excites
- L15:2038 weakly suppresses
- L5:30 weakly excites
- L16:1072 weakly excites
- L2:622 weakly excites
- L15:122 weakly suppresses
- L4:308 weakly excites
- L15:136 weakly suppresses
- L4:1981 weakly excites
- L15:78 weakly excites
- L6:687 weakly suppresses
- L13:3018 weakly suppresses
- L17:944 weakly excites
- L17:1039 weakly suppresses
- L10:2648 weakly excites
- L16:711 weakly suppresses
- L15:89 weakly excites
- L13:91 weakly excites
- L10:332 weakly suppresses
Feeds later layers
- L19:1979 suppresses
- L21:881 suppresses comma after 'if' clause in prose, braces in code
- L24:405 suppresses
- L27:1779 suppresses
- L21:1811 suppresses
- L24:880 suppresses
- L27:967 weakly suppresses
- L20:122 weakly excites
- L19:973 weakly excites
- L22:470 weakly suppresses
- L25:386 weakly suppresses
- L25:6 weakly suppresses
- L21:434 weakly suppresses
- L22:689 weakly suppresses comma closing an intro clause
- L19:1598 weakly excites comma before a qualifying clause
- L26:2964 weakly excites
- L23:292 weakly suppresses
- L19:2866 weakly excites
- L19:624 weakly excites
- L20:2430 weakly suppresses
- L19:2352 weakly suppresses
- L26:377 weakly suppresses 'sh' starting a word piece
- L20:206 weakly suppresses
- L20:2412 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
- abbr
- 文化和
- .band
- UUID
- ания
- ание
- /response
- estead
- ertext
- arged
Suppresses
- {⏎
- {⏎⏎
- {⏎
- {⏎⏎
- { ⏎ ⏎
- {⏎⏎⏎
- ({⏎
- { ⏎ ⏎
- { ⏎
- { ⏎