commas closing a list item
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. 3 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 single item in a list of examples, such as "a house, a doctor" or "a phone number or photograph,". Turned up, the model shifted to quiz and question-style writing.
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.
-
ubs, trees, ferns, gymnosperms, medicinal and scented plants in Fern Garden, Medicinal Herb Garden,
-
PV array, which then adjusts the incoming voltage in order to maintain the most efficient level of power for the system in place.⏎<|endoftext|>Serious diseases, including diabetes, cancer, and eating disorders, can manifest as signs and symptoms inside of the mouth
-
efforts will help to address some of the current challenges in space security.⏎By entering this website, you consent to the use of technologies, such as cookies and analytics, to customise content, advertising and provide social media features. This will be used
-
the party machine. Immigrants lined up at their respective Democratic party committee headquarters and waited their turn for the bosses to find them a job, a house, a doctor. It was an unfair system that did not work the way the civ
-
"undo" or "select all," running within apps at all times. It creates a virtual cloud-based clipboard where any data, like a phone number or photograph, can temporarily live until it's "pasted" to
-
question should be enough for your answers. Worth 10 points each.⏎1. Evaluate the following statement. "A temporarily high government spending, like a war, should be financed by temporarily high taxes."⏎2. Is the following statement true,
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 need to get back to the point of a few more. I want to talk about it. I want to talk about it. I need to get back to the point of a few more. I want to talk about it. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
all of our lives, but if we have an increased propensity to become forgetful -- perhaps to ask a question that has already been answered or increasingly misplace items -- forget appointments, forget medications. If these are new changes, then that's a
-
North Pacific Gyre.⏎Twice the size of Texas – though some scientists have argued that its mass has extended to twice the size of the United States – a majority of the garbage is comprised of tiny bits of plastic, barely visible to
-
-menopausal women chosen from the Framingham Heart Study - which has tracked the medical history of the residents of Framingham, Massachusetts since 1948 - to see whether factors like height, weight or cholesterol impacted the number of children women
-
interviews back to back (Len Irish, who took our cover photo of Collins, says she was one of the nicest people he ever photographed – “no fuss”), there’s a television crew waiting for her to put on her spacesuit, and
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
- L2:845 excites
- L1:379 weakly suppresses
- L2:308 weakly suppresses
- L2:651 weakly suppresses
- L2:265 weakly excites
- L2:1272 weakly suppresses the word 'wide' in 'a wide range'
- L1:536 weakly excites
- L0:225 weakly suppresses
- L0:804 weakly suppresses
- L2:840 weakly excites
- L0:450 weakly excites
- L1:1275 weakly suppresses
- L2:877 weakly suppresses
- L1:2336 weakly excites
- L1:1197 weakly suppresses
- L2:452 weakly excites
- L2:1966 weakly suppresses
- L2:288 weakly excites
- L0:1338 weakly excites
- L1:2293 weakly excites
- L2:358 weakly suppresses
- L1:93 weakly suppresses
- L0:2548 weakly excites
- L0:1464 weakly suppresses
Feeds later layers
- L4:238 suppresses
- L4:69 suppresses
- L4:91 weakly suppresses
- L11:1604 weakly excites
- L23:805 weakly excites
- L8:2503 weakly suppresses
- L19:2035 weakly excites
- L24:3022 weakly suppresses
- L15:1459 weakly excites
- L14:1195 weakly suppresses closing bracket ending an aside
- L14:2596 weakly excites
- L4:535 weakly suppresses
- L20:966 weakly suppresses
- L5:2085 weakly excites
- L11:3018 weakly excites
- L15:1505 weakly suppresses
- L8:2319 weakly suppresses
- L7:1352 weakly suppresses
- L10:1572 weakly suppresses
- L19:17 weakly excites
- L17:1616 weakly suppresses
- L4:539 weakly suppresses
- L15:2628 weakly excites
- L13:2864 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
- 了多少
- main
- 旁
- rozen
- чив
- seen
- ousands
- _fa
- /main
- чат
Suppresses
- evity
- .itemView
- rollers
- iện
- stellen
- "));⏎
- agger
- ughter
- 械
- venge