'of' or 'to' linking two things
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. 13 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
Fires on the small word right after a noun that ties it to another thing, like 'a division of', 'a larger version of', 'a special case of' or 'related to'. Turning it up made the model produce loose, repetitive text, which does not clearly show what it is for.
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 get integer outputs in all cases⏎Determine value using Look-up table method⏎Create Volcano (or Atoll) martix. It is an extension of Bullseye matrix problem.⏎Who has power to do everything in this world?⏎
-
downside: Your new pc comes preloaded with applications.⏎Phrase Origin late Center English: the verb from Old French practiser or medieval Latin practizare, alteration of practicare ‘perform, perform’, from practica ‘apply’, from Greek prakt
-
ure; to accustom by practice.⏎from The Century Dictionary and Cyclopedia⏎- To work; practise; inure; exercise.⏎- A Middle English form of our.⏎- n. Operation; use; practice.⏎- n. Fortune;
-
Highway Administration, a division of the U.S. Department of Transportation, has released the newest numbers for miles driven
-
of the Nieuport 12 with an 80hp Le Rhône engine. The Nieuport 12 itself was a larger version of the Nieuport 10, with a more powerful 110
-
squirrels, tamarins are commonly called kings of the jungle because manes around their neck make them look like lions. The lion tamarins are related to pygmy marmosets. They are 12 inches tall not
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 problem of a single problem, which is the first, in the first place, as one of the first problems in the first place, and the second, and the third, in the first place, as a consequence of the first,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cells.⏎The distance between two cells is the minimal number of up, down, left or right steps through the grid to get from one to the other. It doesn't matter whether the steps go through blank cells or
-
Add your thoughts below. You can sign in with Google, Twitter, Facebook, Yahoo! to get
-
they explore side-branches, they make mistakes. The “right answer” – the proof, the demonstration – simply sets a mile-marker between one phase of discovery and the next.⏎In our Math Circles, young people invent and discover their shared
-
about the art of restoration. It is usually referred to as the non-institutional knowledge passed from master craftsman to apprentice, father to son, or artist to pupil. This is the main difference between restoration and conservation. Both restoration and conservation
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
- L6:2605 suppresses
- L11:2292 excites 'than' in comparisons
- L12:838 excites
- L9:1144 excites
- L6:1477 suppresses
- L12:2959 suppresses
- L12:2298 excites
- L8:1319 suppresses
- L12:2548 excites
- L10:3066 suppresses
- L11:139 excites
- L7:2797 excites
- L6:1084 suppresses
- L6:1828 suppresses
- L10:2213 weakly suppresses 'to' or 'into' showing movement between two things
- L9:2566 weakly excites
- L11:2478 weakly excites
- L12:1987 weakly excites
- L11:2649 weakly excites
- L8:1669 weakly excites
- L12:2346 weakly excites
- L8:1623 weakly excites
- L11:2662 weakly suppresses
- L9:2885 weakly excites
Feeds later layers
- L14:2386 weakly suppresses
- L18:1738 weakly suppresses
- L16:1535 weakly excites
- L17:2920 weakly excites
- L17:2769 weakly suppresses
- L18:2373 weakly suppresses
- L23:1239 weakly suppresses
- L19:103 weakly excites
- L18:1420 weakly suppresses
- L14:2535 weakly suppresses
- L16:765 weakly excites
- L18:860 weakly excites
- L17:1467 weakly suppresses
- L23:1889 weakly suppresses the '3' in '3D' and '2D'
- L16:980 weakly excites
- L17:2516 weakly excites
- L14:2374 weakly excites
- L14:2422 weakly excites
- L14:363 weakly excites
- L15:1126 weakly suppresses
- L17:439 weakly excites
- L22:418 weakly suppresses
- L14:229 weakly excites
- L14:1654 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
- aan
- magic
- ồn
- oted
- permanent
- -exec
- otent
- odynam
- uada
- Insn
Suppresses
- lessness
- nonetheless
- ompiler
- ллер
- ǧ
- jährige
- jähr
- &q
- 公报
- '&