word endings in multi-token nouns
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. 3 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 last piece of certain longer words, such as 'ies' in cardiomyopathies, 'y' in geography, or 'any' in Tuscany. It seems to care about word endings more than meaning, and turning it up did not produce 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.
-
uldade de Filosofia, Ciencias e Letras. Universidade de Sao Paulo, Zoologia,
-
the primary problem lies with the heart muscle cells themselves; these diseases are called cardiomyopathies. Different cardiomyopathies are caused by different cellular defects, but most of
-
rous Density, both direct-read ferrogram and particle quantifier, measures ferrous particles without the size limitation of emission spectroscopy.⏎- Analytical Ferrography looks at size, shape and color (the three most important physical characteristics of a
-
had a forest background. The traditions, laws, weapons, and the general approach to current problems were different. In this environment the influence of geopgaphy had charted a new approach. Modification, transformation of old traditions did not suffice
-
sauce of which you can be proud. Enjoy learning the art of sauce-making!⏎white sauce recipe⏎White sauce was originally invented in Italy (more specifically Tuscany) during the Renaissance era. Back then, it was known as salsa colla
-
formation at the time, it has been described as the embodiment of the Counter-Reformation. ⏎The Council issued condemnations of what it defined to be heresies committed by proponents of Protestantism, and also issued key statements and clarifications 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: I have a 3D printer, and I have a 4D printer. So, I have a 5D printer, and I have a 6D printer. I have a 7D printer. So, I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to paints and insecticides. But recent studies by Hansen and others have shown that some silicone-based surfactants will remove spider mites, thrips and mealy bugs from apples and pears.⏎Hansen's surfactant studies at AR
-
enough water is to watch for signs of wilting. Leaves that are shriveled, papery or wilted are all signs of dehydration. Likewise, mealy bugs are a sign of overwatering.⏎A moisture meter is a great tool
-
areas on lengthy automobile trips.⏎They’re so common that you may even find them sitting atop telephone wires.⏎The Red-tailed Hawk’s loud, falling, raspy screech is frequently employed as a generic hawk or eagle call in media.⏎
-
lime, strawberries, apples, guava, lemons, mushrooms, cabbages, turnip, winter squash, potatoes, beets, carrots, leafy vegetables, pomegranates, pineapples, custard apple, as
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
- L0:1921 suppresses
- L1:2801 excites
- L0:1301 suppresses
- L2:75 suppresses
- L2:1903 excites
- L2:50 weakly suppresses
- L2:1657 weakly excites
- L2:1321 weakly suppresses
- L0:178 weakly suppresses
- L0:1224 weakly suppresses
- L0:2210 weakly excites
- L2:1199 weakly excites
- L2:549 weakly suppresses
- L1:1973 weakly suppresses
- L1:1064 weakly suppresses
- L2:493 weakly suppresses
- L2:122 weakly excites
- L1:29 weakly excites
- L1:1857 weakly suppresses
- L0:2644 weakly suppresses
- L1:2922 weakly excites
- L1:491 weakly excites
- L2:1926 weakly suppresses
- L2:1710 weakly suppresses
Feeds later layers
- L4:428 strongly excites
- L5:983 suppresses
- L4:285 suppresses
- L4:157 suppresses
- L11:870 suppresses
- L6:1915 excites
- L6:105 suppresses
- L6:165 suppresses
- L19:792 suppresses
- L4:452 suppresses
- L25:1740 suppresses
- L10:486 suppresses
- L18:737 excites
- L16:543 excites
- L18:238 suppresses
- L7:391 suppresses 'of' after 'those' or 'many'
- L5:800 suppresses
- L17:2100 suppresses
- L4:128 weakly suppresses
- L7:400 weakly excites
- L12:1622 weakly suppresses
- L8:665 weakly suppresses
- L8:1346 weakly suppresses
- L19:2804 weakly excites 'the' after 'in' or 'use'
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
- kke
- 涿
- ENDING
- less
- uxtap
- 钉
- NK
- homage
- rust
- arrants
Suppresses
- 苄
- веч
- .hw
- ๆ
- izers
- ness
- isers
- лат
- temps
- ##