comma inside large numbers
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. 19 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 lights up on the comma in big numbers like 36,000 or 90,000, where it marks the thousands. Turning it up made the model spit out long runs of commas, while turning it down made it write longer, looser sentences.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
expenses or deductions are subtracted out. Net income is always less than gross income.⏎For example, let’s say you get a job that pays $36,000 a year. You are paid twice a month, so each paycheck
-
0,000 square miles this year. The largest the area has ever been measured was in 2002, when it was about 8,500 square miles.⏎Scientists from the Louisiana Universities Marine Consortium said that this
-
’s quality and usability. Regular updates and bug fixes based on user feedback contribute to long-term success. Lastly, keeping an eye on industry trends, exploring new technologies, and participating in developer communities foster growth and innovation
-
You are the performance management accountant, and you are requested to calculate the Sales Volume Variance of product A. The actual Sale for this product is 90,000 unit⏎Based on formula⏎- Sales volume variance = (Actual
-
uncertainty, serious risk, and even the potential for chaos. But there are also peaceful applications<|endoftext|>An estimated 50,000 to 60,000 farmworkers between the ages of 14 and 16
-
nonprofit The Nature Conservancy since the early 1980s to protect wildlife habitats. To date, Seattle City Light has purchased more than 10,000 acres to protect wildlife habitat, especially that of the various salmon 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 but but the but but it and the the and the the and the and and then and then the and then and then and the and then and then the and then and then the and then and then the and then and then the and
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
from China, snow peas from Guatemala and asparagus from Peru.⏎As fears of mad cow disease
-
on the screen here, are sturgeon down on the bottom of the river. Have you seen a st
-
screens versus books,” notes Wolf.⏎So screen-based reading is bad?⏎So is screen-based literacy “bad” and print-based literacy “good”? Not at all, says Wolf. Instead, she argues, “The great challenge now is to learn
-
University of Passau)<|endoftext|>||Audio Video Interleave File⏎||MPEG-4 Video File⏎||AVI is a video wrapper format created by Microsoft, storing video data that might be encoded with various codecs. AVI files can be
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
- L16:240 strongly excites comma after 'sometimes' or 'the truth'
- L17:118 strongly excites comma after a clause ending
- L14:35 strongly suppresses
- L18:366 strongly excites comma inside numbers and dates
- L9:77 excites commas in citations and name lists
- L17:1405 suppresses
- L4:3 excites commas after citation fields
- L11:154 suppresses commas after citation numbers
- L17:116 excites
- L5:16 excites
- L10:335 excites commas in catalog, date, and list metadata
- L16:241 excites
- L18:1394 excites
- L16:1082 suppresses comma linking two clauses after 'and'
- L3:71 suppresses
- L15:588 excites comma after an introductory phrase
- L2:66 excites
- L18:892 suppresses
- L1:86 excites
- L16:489 excites
- L17:585 excites
- L18:48 suppresses
- L18:2376 suppresses
- L18:93 suppresses
Feeds later layers
- L23:184 strongly suppresses
- L22:144 suppresses
- L20:476 suppresses commas that lead into a continuation
- L21:632 excites
- L20:2082 excites
- L27:139 suppresses
- L26:94 suppresses
- L27:77 suppresses
- L21:700 suppresses comma before a new clause
- L22:368 excites
- L27:358 excites
- L26:236 excites
- L24:41 suppresses
- L20:2608 excites
- L27:191 excites
- L24:902 suppresses
- L21:50 suppresses comma after 'So' or intro phrase
- L26:22 suppresses the 'res' fragment inside words
- L20:811 excites
- L26:345 excites
- L27:130 suppresses
- L25:1221 excites
- L27:86 suppresses
- L25:118 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
- ,
- ,...⏎⏎
- .—
- ,—
- ,b
- ,⏎⏎⏎⏎
- ,...⏎
- ,B
- ,W
- ,则
Suppresses
- regardless
- etc
- which
- especially
- but
- along
- including
- assuming
- pois
- 以及