'-ment' word endings split mid-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. 23 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 middle pieces of longer words that end in 'ment', such as 'assessment', 'allotment' and 'abandonment'. It is quiet, or goes the other way, on the word 'half' in phrases like 'at half past'. Turning it up made the model produce odd words ending in 'ment', so the pattern shows up in the output too.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
LIMITED WARRANTY*⏎All spray foam installations are carried out by Lapolla / AirTight licensed installers.⏎*See manufacturer guidelines and requirements.⏎HOME ENERGY ASSESSMENT⏎A home energy assessment helps owners determine where their house is losing energy and
-
the acronym Family ASSESSMENT—Structure, Socioeconomic, Sickness, Nutritional, Educational,
-
impossible to give your garden or allotment the care it needs to thrive. And even those without the luxury of their
-
its former glory. Today, the West Baden Hotel is a National Historic Landmark and continues to attract visitors from all over the world.⏎Decline And Abandonment⏎The West Baden Hotel was once known as the “Eighth Wonder
-
writer and translator. Born in London, she now lives in a small village near Oxford. She enjoys growing flowers, fruit and vegetables in her garden, down the allotment and as a volunteer with therapeutic gardening charity Restore. She is especially interested in
-
otechnic" reactions, i.e. shock- induced reactions in heterogeneous ma- terial mixtures.⏎- ISOTOPIC SEPARATION AND ENRICHMENT⏎- “ballotechnic” weapons, possibly pure fusion weapons.⏎-
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: how to way in which way we way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way way
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
I opened the window, it was raining heavily and people⏎were hurrying along the streets⏎2. Every day the boss enters the office⏎at half past nine⏎3. My friend will ring me up at eight o’clock
-
window, it (to rain) heavily and people (to hurry) along the streets. 2. Every day the boss (to enter) the office at half past
-
name appears, select Grades. On the Grades page, select the cell that displays the average grade in your course. The Course Activity Related to Grades report appears when half of your students have received more than one grade.⏎If your class enrollment is in
-
asked to suggest further objects and can actively participate by uploading photographs of their own objects that have a local or global appeal.⏎History of the World events⏎During February half-term, there is a guided walk which will tell the story of the ten objects
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
- L16:120 weakly suppresses
- L3:2594 weakly excites
- L6:238 weakly suppresses
- L19:1851 weakly excites
- L20:1616 weakly excites
- L2:84 weakly excites
- L17:127 weakly suppresses the word 'would' before a verb
- L20:2889 weakly suppresses
- L2:2347 weakly excites
- L1:1273 weakly suppresses
- L5:2577 weakly suppresses
- L5:19 weakly suppresses
- L4:992 weakly suppresses
- L6:2745 weakly excites
- L18:511 weakly suppresses
- L9:2975 weakly excites
- L16:3017 weakly excites
- L22:1871 weakly suppresses
- L20:2371 weakly excites
- L17:2620 weakly suppresses
- L0:2462 weakly suppresses
- L19:246 weakly excites
- L16:155 weakly suppresses
- L11:2252 weakly excites
Feeds later layers
- L24:3045 excites
- L25:1061 excites
- L26:1533 weakly excites
- L27:1809 weakly suppresses
- L24:1671 weakly suppresses
- L24:979 weakly suppresses
- L25:1087 weakly excites word right before a number value
- L27:364 weakly suppresses
- L27:992 weakly suppresses
- L25:1020 weakly excites
- L24:2895 weakly suppresses
- L26:1682 weakly suppresses
- L24:2878 weakly suppresses
- L26:2998 weakly suppresses
- L27:1543 weakly excites
- L26:102 weakly excites
- L26:1754 weakly excites opening parenthesis before an acronym
- L24:1523 weakly excites
- L24:1940 weakly suppresses
- L24:1605 weakly excites
- L27:2373 weakly suppresses
- L27:422 weakly excites
- L26:1578 weakly excites
- L27:2559 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
- ment
- MENT
- ments
- amento
- amiento
- amentos
- ements
- ament
- AMENT
- ement
Suppresses
- way
- -way
- way
- Way
- 径
- 之一
- ONGL
- WAY
- 徑
- WAY