mid-word fragments in capitalised titles
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. 20 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 piece of a word that comes in the middle of a longer word, especially in all-caps headings and titles. It is hard to say what it does from the examples, and the steering outputs did not show a clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
eth not at all. '⏎AKAI HEIRLOOM RICE - GRAIN AKAI RICE is a heirloom rice
-
of Alzheimer’s.⏎Courtesy of Baycrest, the global leader in innovations, research and breakthroughs for the journey of aging.⏎ALSO ON EVERYTHINGZOOMER:<|endoftext|>SeaWiFS: Gulf Coast Sediments⏎Provided by the Sea
-
, and English character and manners described.?1 Paradoxically, these three editions?The Juvenile<|endoftext|>CASE NUMBER: 212⏎CASE MNEMONIC: BASMEX⏎CASE NAME: Mexico and Basel⏎On the Mexico-US
-
you like to know what each chakra represents? Here is a simple diagram and list for you to follow. | WIKIPEDIA⏎THE CHAKRAS ARE AS FOLLOWS⏎- ROOT/BASE (MULADH
-
6⏎THE CHAKRAS AND CHAKRA BALANCING MEDITATION⏎SPINNING WHEELS (SPHERES) OF LIGHT CODES AND PROGRAMS⏎Vetruvian Man with Seven Chakra Overlay⏎
-
largest wetland in the North Island and is recognised as a Ramsar site – a wetland of international importance.⏎Ō Tū WharekaiCANTERBURY Ō Tū Wharekai is one of the best intact
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 **C** -F (C, F) (C, F)** **C** -**F** (C, F) **C** -**F** (C, F) **C** (
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
|⏎|Number of pages||12|⏎|Journal||Classical and Quantum Gravity|⏎|State||Published - Jul 7 2003|⏎ASJC Scopus subject areas⏎- Physics and Astronomy (miscellaneous)<|endoftext|>Most
-
single-layer coated, triple-layer coated and graded-index-layer coated samples was performed and compared with experimentally derived films. Good agreement between model and experiment was observed.⏎ASJC Scopus subject areas⏎- Electronic, Optical and Magnetic Materials⏎- Cer
-
work in the jungle, there’s nothing Milton likes better than fresh fruit bat soup.⏎IS39 is a seven-element array deployed in a triangular geometry with four elements in
-
.⏎Are standards mandatory ?⏎Standards are considered voluntary because they serve as guidelines, but do not of themselves have the force of law. ASME cannot force any manufacturer, inspector, or installer to follow ASME standards.
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
- L18:373 strongly suppresses
- L16:584 strongly excites
- L19:1731 strongly suppresses
- L5:35 strongly suppresses
- L18:599 suppresses
- L10:115 excites the word COVID before -19
- L4:533 suppresses
- L6:2269 suppresses
- L17:699 excites
- L1:2221 excites
- L16:158 suppresses
- L3:627 suppresses
- L0:76 excites
- L19:188 excites
- L15:257 suppresses
- L18:546 excites
- L7:128 suppresses
- L17:1747 suppresses split tech names and acronyms
- L19:2344 suppresses
- L6:430 suppresses
- L10:2313 excites
- L19:712 suppresses
- L3:1479 suppresses
- L0:1904 suppresses
Feeds later layers
- L22:651 strongly suppresses
- L22:1126 strongly suppresses
- L26:1531 strongly suppresses
- L27:1360 strongly suppresses
- L24:980 strongly suppresses
- L24:1991 strongly suppresses
- L21:533 strongly suppresses
- L27:1879 strongly suppresses
- L25:1750 strongly suppresses
- L26:434 strongly suppresses
- L22:212 strongly suppresses
- L26:2849 suppresses
- L24:1970 suppresses
- L27:1520 suppresses
- L27:2146 suppresses
- L27:782 suppresses
- L25:244 excites
- L27:2505 suppresses
- L26:155 excites
- L27:1906 suppresses
- L27:491 excites
- L24:938 excites
- L27:1685 suppresses
- L26:816 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
- ours
- ourselves
- uellement
- SqlServer
- hers
- Eb
- oneself
- herself
- angep
- 该项
Suppresses
- s
- ING
- AND
- ICS
- ANT
- 必
- ABLE
- ORT
- ER
- OR