'in' after a noun
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. 17 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
The neuron lights up on the word 'in' when it follows a noun and adds a location or context to it, like 'the energy released in these collisions' or 'the options in the drop-down menu'. Turning it up made the writing repetitive and off-track, so the steering does not show the pattern clearly.
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.
-
way down these particles are slowed down by Earth’s atmosphere, which acts as a shield. These charged particles collide with gas molecules in the atmosphere. The energy released in these collisions is given off as light.⏎Geomagnetic storms follow the 11
-
to use technical jargon as a means to succintly describe the work done. However, if the target audience does not have sufficient understanding of the subject included in the report, care must be taken in order to avoid or prevent the use of
-
you’re seeing in the clip above:⏎Some of the stellar debris falls into the black hole and some of it is ejected into space at high speeds. The areas in white are regions of highest density, with progressively redder colors corresponding to lower-density
-
earnings is substantial. If a low value-added teacher in the bottom 5% is replaced by an average teacher, not even an excellent one, the expected increase in the present value of students lifetime income is more than $250,0
-
to the Zipped file from the File Explorer and right-click on it. From the options in the drop-down menu, choose Extract All.⏎Step 2: A folder with
-
the country by the beginning of the 20th century, it was New Orleans' version that had the most influence on Chicago jazz and early swing that followed in the 1920s. Drawing on a thorough and imaginative command of
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: a very big a large set of a very large set of a large set of a large set of a large set of a large set of a large set of a large set of a large set of a large set of a large set of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
story plausible, we can argue that, among other things, Western countries uphold conceptions of nationalism that center around individual rights and
-
least acidic among citrus fruits, but it has many seeds. The rind is thin and mand
-
in finding the best solution from a large number of options.⏎They found that classical methods still performed best in these
-
humanity and spark global change.⏎The Tech Awards, a signature program of The Tech Museum of Innovation in San Jose, California, selected the Drift Catcher from among hundreds of nominations representing 68 countries, for "empowering communities for health
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
- L13:132 excites "of" after a noun
- L16:1807 suppresses 'in' after a motion verb
- L10:132 excites
- L5:763 suppresses
- L8:45 suppresses
- L3:232 excites 'on' after a noun like 'impact on'
- L4:4 suppresses
- L8:706 suppresses
- L15:906 suppresses
- L12:1467 weakly excites numbers after dates and figures
- L5:729 weakly suppresses
- L14:539 weakly excites
- L4:1035 weakly suppresses 'for' after verbs like 'fall' or 'click'
- L10:469 weakly suppresses 'of' after a noun phrase
- L10:272 weakly suppresses
- L4:1500 weakly suppresses
- L8:3016 weakly excites
- L14:917 weakly excites
- L14:1519 weakly excites
- L4:192 weakly excites
- L15:894 weakly excites
- L14:1326 weakly suppresses
- L4:1966 weakly suppresses
- L3:311 weakly suppresses
Feeds later layers
- L18:1173 suppresses
- L20:880 suppresses
- L20:1494 weakly suppresses
- L21:1061 weakly suppresses
- L19:1724 weakly suppresses
- L22:1524 weakly suppresses
- L21:1030 weakly excites
- L19:7 weakly suppresses
- L22:194 weakly excites
- L22:103 weakly excites
- L18:2384 weakly excites
- L19:2470 weakly suppresses
- L20:1423 weakly excites
- L20:47 weakly suppresses
- L19:914 weakly excites
- L26:284 weakly excites
- L27:797 weakly suppresses
- L19:532 weakly excites
- L19:769 weakly suppresses
- L27:2106 weakly suppresses
- L19:2425 weakly suppresses
- L18:604 weakly suppresses
- L26:1959 weakly suppresses
- L19:219 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
- licity
- venth
- чат
- ധ
- usch
- harga
- since
- ícia
- izzazione
- irth
Suppresses
- different
- 不同的
- different
- Gos
- ographs
- .equalTo
- 此基础上
- 多样化
- 完全不同
- sel