word fragment inside a longer word
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. 24 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
This neuron fires on the first pieces of a word that has been split into parts, such as 'Kar' in 'Karate' or 'cycl' in 'cyclonic'. Turning it up sent the model into repeated 'way' and 'time' phrasing.
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.
-
develop into the individual he or she is destined to be by virtue of their own efforts.⏎Pure Fun - Yoga Class⏎Practical Life learning⏎Pure Fun - Karate class⏎5 year old graduation<|endoftext|>Look at the egg⏎picture. Write
-
uelen [and] Plan du Havre de Noel, dans la Terre de Kerguelen, by Rigobert Bonne, from Atlas Encyclopedique, circa 1787, (11.3
-
time of the attack of Omer Pasha on Montenegro in 1862. At the end of the 19th century this building was Karuč school, managed by the teacher Mihajlo K. Popovic
-
most widespread fabrics in the world. Oxford cloth is created by weaving together two yarns that are of the same color but just a shade different, giving it a marled texture that breaks from the ordinary aesthetic of a solid dress shirt.⏎Oxford
-
storms that produce strong winds, rain, and storm surges. “Tropical cyclone” is a generic name that refers to the storms' geographic origin and cyclonic rotation around a central eye. Depending on their location and strength, the storms
-
are more expedient (recycling, turning off lights).⏎We appreciate the efforts of the contributors to this special issue who are addressing some of<|endoftext|>L’Encyclopédie de l’histoire du Québec / The Quebec History Encyclopedia⏎Eth
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: when you think of all the things that you can do in your life, do you think you can get by by simply by helping the people around you by by by by by by by by by by by by by by by by by by
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
exposure, whereas radiation-related health problems may often take up to a decade before symptoms develop.⏎(ARTICLE CONTINUES BELOW)⏎The researchers said that carrying out studies in Israel, where the level
-
Capital News – Dec. 3, 2017);⏎- “Icy road leads to crash” (Castanet – Nov. 4, 2017)⏎Do you ever
-
with the sacredness of the soul. Carolyn D. Williams (1993) examines Lucrece's rape, as well as the rape of Lavinia from Shakespeare's Titus Andronicus, and finds that Shakespeare's treatment
-
ivian Army and the CIA in 1967 (Castaneda 1998, 326).⏎What Individuals Thought of
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
- L22:2807 suppresses
- L19:595 weakly excites
- L9:1653 weakly excites
- L22:572 weakly excites
- L20:2067 weakly excites
- L23:824 weakly excites
- L17:1795 weakly excites 'and' before a second verb phrase
- L11:2516 weakly suppresses
- L21:527 weakly excites
- L23:1454 weakly suppresses
- L23:2025 weakly excites the name fragment 'Bon' (Bonn, Bonaparte)
- L4:1134 weakly suppresses
- L6:2697 weakly excites
- L17:443 weakly suppresses
- L7:647 weakly suppresses
- L2:2703 weakly suppresses
- L3:410 weakly suppresses
- L3:1220 weakly suppresses
- L3:659 weakly suppresses
- L19:493 weakly suppresses
- L21:751 weakly excites
- L16:1066 weakly excites
- L14:2589 weakly suppresses
- L5:429 weakly suppresses
Feeds later layers
- L27:618 weakly excites
- L27:422 weakly excites
- L25:333 weakly excites
- L26:1328 weakly suppresses
- L27:1675 weakly suppresses
- L27:686 weakly suppresses
- L26:2491 weakly excites
- L26:2082 weakly excites
- L27:897 weakly suppresses
- L27:1227 weakly suppresses
- L27:2734 weakly suppresses
- L26:2297 weakly suppresses
- L25:1536 weakly suppresses
- L26:469 weakly excites
- L25:1613 weakly excites
- L25:1015 weakly suppresses
- L25:2794 weakly suppresses
- L26:977 weakly suppresses
- L25:2517 weakly suppresses
- L27:2459 weakly excites
- L25:532 weakly excites
- L27:2054 weakly excites
- L26:1615 weakly excites
- L25:1037 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
- pire
- SCI
- áz
- ivation
- asi
- ig
- imax
- ihad
- uda
- Scroll
Suppresses
- 垓
- bum
- uent
- erne
- .entrySet
- nty
- precursor
- nts
- TInt
- .bo