number right after 'about' or 'less than'
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 the gap just before a quantity, as in 'about [ ]160 square miles' or 'less than [ ]0.01%'. It is waiting for a number to come next. Turning it up made the model produce long runs of digits.
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.
-
, 6–8⏎A hunk of ice about the size of the state of Connecticut has started to collapse in Antarctica. According to satellite images, about 160 square miles of the Wilkins Ice Shelf shows massive cracking, scientists
-
of the total global population). Population ageing is happening more quickly than in the past. For example, while France had almost 150 years to adapt to a change from 10% to
-
solution to the problem of the labour costs incurred for moving the bottles; however, they require a high initial investment, while the remuage process will last about 7 days. A winery using these methods might not be able to meet a
-
The ocean’s breadth and depth remain the most extensive, unexplored habit.⏎Activity 2 – Freshwater Habitats⏎Freshwater Habitats cover less than 0.01 % of the earth’s
-
and younger adults. (Courneya, Estabrooks & Nigg 1997) Unlike attitude, which tends to consistently account for approximately 30% of behavioral intention, social norms are rarely even significant and always account
-
’s spiral arms are.”⏎Named GLIMPSE360 (Galactic Legacy Mid-Plane Survey Extraordinaire project), the deep infrared survey captures only about 3% of the sky, but because it focuses on the plane of the Milky
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 idea that, rather than "to" rather than "to" instead of "to" the "to" at the end in the "to" "to" "to" "to" "to" "to" "
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Despite being high in fat and sodium, tamales are a rich source of vitamins, minerals and fiber that support optimal health.⏎A 100-gram serving of tamales contains about 18.
-
the whaling industry, which nearly drove some species to extinction. At its peak in 1856, the United States whaling industry produced between four- and five-million gallons of whale oil annually.⏎Whale-oil lamps used one
-
2, WGBH Media Library & Archives, accessed March 29, 2017, http://openvault.wgbh.org/catalog/V_4E4EE3EAD24043E2BC0
-
’s in order to win a prize, that the problem was solved.⏎Joan Marie Galat, Pelican ©2012, 978-1-4556-1637-4⏎In the
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
- L13:923 suppresses word right before a number or year
- L4:53 excites blank slot before a number
- L17:285 suppresses
- L10:220 suppresses hyphen inside compound words
- L6:344 excites
- L5:2539 excites
- L16:256 excites number right after 'age of' or 'ages'
- L8:111 suppresses the space before a year in dates
- L0:602 excites words right before a number
- L4:1129 excites
- L4:490 weakly suppresses decade digit in years like 1930s
- L6:17 weakly suppresses
- L14:348 weakly suppresses
- L16:723 weakly suppresses
- L4:2691 weakly excites
- L4:2129 weakly suppresses
- L2:422 weakly excites
- L5:2991 weakly suppresses
- L14:140 weakly suppresses
- L1:349 weakly suppresses
- L17:382 weakly suppresses closing bracket after a fact in parentheses
- L17:525 weakly suppresses
- L16:7 weakly suppresses
- L4:2532 weakly excites surname pieces split across tokens
Feeds later layers
- L27:170 excites unclear: spaces in mixed-language text
- L21:339 excites spot just before a year
- L25:33 excites spot where a number is due
- L22:824 excites blank before an expected number
- L22:318 suppresses
- L21:302 excites
- L26:103 excites blank space before a number
- L27:283 excites the spot right before a number
- L20:131 excites
- L19:26 suppresses
- L21:1358 suppresses
- L23:77 excites the space right before a number
- L25:253 suppresses
- L26:611 suppresses
- L19:43 suppresses
- L27:306 suppresses
- L21:189 excites document break token mid-sentence
- L20:61 suppresses
- L27:2172 excites
- L22:372 excites empty spot before a number or "3D"
- L19:58 suppresses
- L20:4 excites
- L25:182 excites
- L20:72 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
- 理会
- filib
- 里的
- 的世界
- Chili
- ︰
- roscope
- 月中
- Cunning
- Boca
Suppresses
- ger
- acious
- -
- push
- ao
- 报警
- burst
- ins