'of' after 'those' or 'many'
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. 7 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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
The neuron fires on a word like 'of' or 'that' that comes after a quantity word such as 'those', 'many' or 'some', and introduces a group that a part is taken from, like 'those of us' or 'many of the issues'. The steered text was unrelated, so the pattern comes only from the examples.
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.
-
to be no obvious health implications to excess contrast/saturation/sharpness as there are with salt, fat, and sugar.⏎But perhaps there are some subtle ones.⏎While there is a lot of concern about the effects sensory over-stimulation in
-
this tree are found in Palestine. It is a "very graceful tree, with long feathery branches and tufts closely clad with the minutest of leaves, and surmounted in spring with spikes of beautiful pink blosoms,
-
.⏎2625 In the first place these are prayers that the faithful hear and read in the Scriptures, but also that they make their own - especially those of the Psalms, in view of their fulfillment in Christ.96 The
-
, given enough time and sufficient access, they hope to catch a glimpse of the famous ship too. Only time will tell if that will actually happen, but those of us who are fans of Shackleton and his story are
-
First⏎thats right vegetarian spider⏎A tropical jumping spider that eats mostly plant buds has been identified, a new study says—making it the only known vegetarian out of some 40,000 spider species.⏎In the late
-
(secret service) as early as 3. 000 years ago. London was probably founded some seven years after the Roman invasion of Britain (those of you who know Asterix might
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 to be done with a single or a problem that can be solved with a simple logic. The problem is presented in the form of a problem that is often used in mathematics, and the solution is provided in the form of a single
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
which states that our beliefs shape our actions. We do what we feel is right and this is what forms the organizational culture as a whole.⏎- The Simultaneity Principle – that states that appreciative inquiry or the way we question our internal
-
until the moment a piece is installed. He arrives at the gallery with barrels filled with the component pieces and lets the spirit of the day and the idiosyncrasies of the site
-
for years. And while some rumors are false --- sometimes intentionally --- many others are surprisingly accurate. The website www.snopes.com has surfaced as the rumor-skeptic's friend, collecting and verifying tales of every type.⏎DiFonzo is
-
bidities which contribute to progression of CHF, particularly if treatable is our current challenge and need to be pursued systematically and aggressively.⏎Among multiple co-morbidities, sleep-related disordered breathing (SRBD), is the most common 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
- L6:105 suppresses
- L5:800 excites
- L5:2220 suppresses
- L2:20 suppresses
- L4:2546 excites
- L6:1915 suppresses
- L1:67 excites
- L6:942 suppresses
- L6:84 excites
- L4:834 excites
- L3:58 excites
- L5:1273 excites
- L6:302 excites
- L2:803 excites
- L4:157 suppresses
- L6:2940 suppresses
- L4:509 suppresses
- L6:361 excites
- L4:420 suppresses
- L5:983 excites
- L0:1517 excites
- L3:2972 suppresses word endings in multi-token nouns
- L6:141 weakly suppresses
- L4:428 weakly excites
Feeds later layers
- L8:71 excites
- L10:930 excites
- L13:2099 excites
- L25:1890 suppresses
- L8:411 excites
- L11:870 weakly excites
- L12:571 weakly excites
- L9:152 weakly excites
- L15:1495 weakly excites
- L18:801 weakly suppresses
- L18:238 weakly excites
- L10:235 weakly excites
- L12:656 weakly excites
- L14:1601 weakly excites
- L17:1125 weakly excites
- L11:2733 weakly excites
- L9:787 weakly excites
- L8:965 weakly excites
- L18:737 weakly suppresses
- L14:2647 weakly excites 'and' or 'or' in split hyphen compounds
- L18:2929 weakly excites comma after a finished clause
- L17:553 weakly excites
- L16:525 weakly suppresses
- L19:1722 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
- importe
- �单
- itives
- vard
- ?>'
- encer
- ipher
- �
- ;',
- import
Suppresses
- uada
- =log
- reproduced
- warranted
- =a
- Pale
- ма
- =i
- 鹃
- การทำ