middle of split technical words
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. 8 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 technical or scientific words that are split into several parts, such as 'gr-ating', 'R-aman' and 'Pale-olithic'. It seems to track word fragments in specialist vocabulary. Turning it up made the output change topic without a clear pattern.
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.
-
perm Phylogeny Website (Family: Loranthaceae, Order: Santalales)⏎Text by A. L. Quirico⏎| ||Key to the genera|| |⏎|1||Terrestrial shrubs, parasitic upon the
-
essentially the same for light from the distant nebulae as in light from laboratory sources. The assumption has been tested by measuring red-shifts both with a grating spectrograph and a prism spectrograph. The fact that the shifts are the
-
influence of aerosols on the global climate.⏎Ault, A.P., et al. “Raman Microspectroscopy and Vibrational Sum Frequency Generation Spectroscopy as Probes
-
backhoe trenches, and seven⏎excavation units. Two buried horizons were located on 48TE414. A Late⏎Prehistoric point and a
-
between the compounds that comprise those particles.⏎The power of optical microscopy is multiplied when coupled to techniques that can chemically speciate the object being magnified. Raman microspectroscopy reveals information about the composition of single aerosol particles by probing
-
forms a mass of stalactites that gradually fills in the base of the projecting terracette.⏎Collapse features are fairly common on most of the Mammoth terraces. Water flowing through near-surface channels weakens the overlying travertine
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 idea of a concept being put into practice, but it is just a thought in the memory (in the form of a procedure) for the following instances, and then it gets into the database of the database of the database of the database
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
turn innovative ideas into finished products, often using 3D printing. As such, Fracassi
-
made statements about how the National Research Council isn’t going to pursue projects with immediate commercial applications and it makes me sad. Little curiosities like this can grow into amazing large-scale processes with a bit of support, but a lack of funding threatens
-
virtue. No amount of scolding, criticism, and punishment can turn a regrettable situation around. When you are not chided and you deserved to
-
is scientifically illustrating the educational power of things like play, sports, music, memorization and reflection. What nourishes the human spirit, the whole person, it turns out, also hones our minds.⏎An improvisational storytelling class of 5
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
- L6:2821 weakly suppresses
- L7:2332 weakly excites
- L4:2585 weakly suppresses
- L2:2132 weakly excites
- L7:2991 weakly suppresses
- L4:2179 weakly excites
- L0:2296 weakly suppresses
- L6:3018 weakly excites
- L7:3036 weakly excites
- L7:1018 weakly suppresses
- L4:2444 weakly suppresses
- L4:3042 weakly suppresses
- L7:2499 weakly excites
- L4:1149 weakly excites
- L7:2186 weakly suppresses
- L5:2933 weakly excites
- L6:579 weakly excites
- L1:1403 weakly suppresses
- L3:1930 weakly excites
- L7:2452 weakly suppresses
- L7:2474 weakly suppresses
- L1:2675 weakly excites
- L3:2181 weakly suppresses comma inside numbers
- L6:2720 weakly excites
Feeds later layers
- L18:312 weakly excites
- L19:3022 weakly suppresses
- L10:2335 weakly suppresses
- L17:1882 weakly suppresses
- L9:1755 weakly suppresses
- L18:1211 weakly suppresses
- L19:1873 weakly suppresses
- L11:2635 weakly suppresses
- L27:2519 weakly suppresses
- L12:983 weakly suppresses unclear: mid-clause words in expository prose
- L10:960 weakly excites
- L27:398 weakly suppresses
- L17:2862 weakly suppresses
- L15:2390 weakly excites
- L13:378 weakly excites
- L10:2475 weakly suppresses
- L17:879 weakly excites
- L21:2810 weakly suppresses
- L10:961 weakly suppresses
- L11:321 weakly excites
- L10:2144 weakly suppresses
- L11:1687 weakly suppresses
- L10:1818 weakly suppresses
- L22:2871 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
- k
- -taking
- /no
- no
- 能找到
- ounding
- 叶
- 谴责
- 破裂
- clave
Suppresses
- cales
- сте
- antino
- raw
- adin
- manos
- テン
- contagious
- proper
- 大事