comma after 'and' or 'and then' clause
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 fires on a comma that comes right after a joining word like 'and' or 'and then', where a new clause is about to start. Turning it up made the writing drift into repetitive, loosely connected 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.
-
,” we get the subject form (An Luan), and, by following genitive case rules, we get the preposition form (Dé Lu
-
in protecting the courthouse.⏎Thereafter, they get an opportunity to learn about the legal aspects of the court system from a judge who explains how trials work. And, then, when they least expect it, they get to experience a trial themselves –
-
is time to stop reading my words, bundle up the kids and hike into the Arctic. Why not today! Children will almost certainly see seals just yards away and, if the timing is right, go nose to nose with a mighty polar bear that
-
.⏎- In the case of Col. Shaw, it seems that he chose to live his life a certain way (leading his African American troops into battle) and, now that he's dead, he's off on the sidelines, checking out all
-
emitted by light sources attract insects. A study was conducted on moths. It was observed that, only the males of the species got attracted towards the light source, mistaking it
-
the early 1990’s, rats were fed GM tomatoes. Actually, they refused to eat them, so they had to be force fed. And, rats aren’t the only animals who’ve declined a snack of GMOs. Smith
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 am" is a good one to be considered a "good" item to be considered a "good" item is good" is a good item" is "good" is "good" is "good" is "good"
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
risk of strep throat infection:⏎- Young age. Strep throat occurs most commonly in children.⏎- Time of year. Although strep throat can occur anytime,<|endoftext|>The dangers of feeding lorikeets⏎Lorikeet feeding habits⏎
-
premise had yet to be tested. John enlisted coleopterist Peter Kovarik as a partner in this study. The two decided that in addition to butterflies, scarabaeine scarabs and hister beetles would become part of the
-
a pH greater than 5.2 or under drought conditions. The tubers are most susceptible, but stems and roots may occasionally be infected. No above-ground symptoms are apparent on infected
-
have any photos of this monument yet. Why don't you be the first to send us one?⏎If Google Street View is available, the image is from the best available vantage point looking, if possible, towards
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
- L16:1223 excites
- L9:19 excites
- L8:511 excites
- L17:1028 excites
- L17:980 excites
- L13:796 weakly suppresses
- L14:1993 weakly excites
- L16:1886 weakly excites
- L17:639 weakly suppresses
- L15:403 weakly suppresses period inside an abbreviation
- L6:250 weakly excites
- L14:1487 weakly excites
- L16:1442 weakly suppresses
- L4:2942 weakly excites 'yet' and 'haven't' after have
- L17:990 weakly excites
- L12:1284 weakly suppresses
- L17:2684 weakly suppresses
- L16:1100 weakly excites
- L15:1379 weakly suppresses
- L15:2447 weakly suppresses
- L12:1344 weakly excites
- L16:489 weakly suppresses
- L16:714 weakly suppresses
- L14:1457 weakly suppresses
Feeds later layers
- L19:1590 excites
- L19:2098 excites
- L19:1463 excites
- L22:160 suppresses
- L21:140 suppresses
- L19:543 excites
- L22:132 excites
- L21:1594 excites
- L26:585 excites
- L25:277 excites
- L20:1245 excites
- L21:818 excites
- L19:31 weakly excites
- L20:1281 weakly suppresses
- L24:116 weakly suppresses
- L21:801 weakly excites
- L26:259 weakly suppresses
- L19:167 weakly excites
- L19:2946 weakly excites
- L19:398 weakly suppresses
- L23:1137 weakly excites
- L27:684 weakly excites
- L24:784 weakly excites
- L19:1765 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
- amid
- arend
- 辕
- Tonight
- aboard
- amidst
- imates
- вокруг
- usting
- Amid
Suppresses
- nào
- ampa
- pass
- pass
- trans
- 資
- reserved
- col
- cape
- odos