unclear: word fragments and digits inside numbers
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. 18 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
The snippets don't share a clear pattern. The strongest hits sit on odd fragments inside words or numbers, such as the piece "Sevent" in "Seventeen" or "them" in "thematically". It is strongly negative on the digit "7" inside years like 1765. Turning it up or down gave unrelated text, so the neuron's job is hard to pin down.
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.
-
highlighted for years as the ninth planet in the solar system.⏎However, in 20
-
000 and were so interconnected by marriage that they represented a giant family rather than a new polyglot community. The community was governed by The Council of Seventeen in Amsterdam, who governed the far-reaching empire of the Dutch East India Company
-
shape them using a jigsaw. 3. Drill holes in the stringers and attach
-
cholesterol from animal products we take in when eating. Even more interesting is the fact that at
-
changing fundamentally, but also the way people learn. Learning trends such as mobile learning, micro
-
heat. Fish oils are notorious for their self-heating properties. Mineral oils like white spirit, mineral turpentine or lubricating oil are not prone to self-heating and will not cause spontaneous ignition.⏎It
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 time when a different team is the first to the final, but the game is already been decided, and the game is already 78% of the way, and time is 3:35, but the game is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
th century, the term was extended from the talker to the conversation of such persons. Th<|endoftext|>Radicalism and the Common People: Britain c.1789-1880 - 30 credits (HS17
-
7, 28, 31, 32, 33, 34, 35, 36, 37, 39, 49, 51, 6
-
became legal tender in much of the country and many used hemp to pay their taxes.⏎The cultivation and spread of hemp in the United States continued and in 1765 George Washington made reference to growing hemp in
-
is predicated on changing human behavior; the purview of Psychologists (reviewed in , see also [2,3,4,5,6,7]).⏎So-called “environmental problems” are really problems of human behavior, caused
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
- L2:2028 strongly suppresses
- L0:2319 strongly excites
- L4:1049 strongly excites
- L17:350 excites
- L10:1210 suppresses
- L14:643 excites
- L0:756 suppresses
- L16:17 suppresses
- L6:1434 suppresses
- L1:19 suppresses
- L16:673 excites
- L17:389 suppresses
- L16:112 excites
- L16:503 excites
- L17:127 suppresses the word 'would' before a verb
- L7:922 suppresses
- L16:113 suppresses
- L15:80 suppresses
- L16:120 suppresses
- L11:981 excites
- L16:56 suppresses
- L15:82 suppresses
- L17:391 suppresses
- L16:84 excites
Feeds later layers
- L19:1269 strongly suppresses the word 'about'
- L20:1793 suppresses
- L20:558 suppresses
- L22:607 excites
- L19:551 excites
- L20:342 excites
- L20:121 excites
- L19:683 excites
- L19:316 excites
- L21:2906 excites
- L20:2115 excites
- L19:2293 excites
- L22:2379 excites
- L23:2739 suppresses
- L26:2860 excites
- L26:1063 excites
- L19:576 excites
- L22:961 excites
- L22:2526 excites
- L20:1228 excites
- L19:1615 excites
- L20:1310 excites
- L20:183 excites
- L20:2371 excites
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
- nad
- Ỗ
- adium
- utilisateur
- éal
- INGS
- عار
- →elem
- iores
- enburg
Suppresses
- 庸
- 放心
- truncate
- TripAdvisor
- <hr
- ⏎ ⏎
- 脸
- reel
- 厘
- ⏎ ⏎