middle of personal names
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. 19 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 people's names, such as the 'ima' in Ibrahima or the 'i' in Devi. Turning it up made the model talk about 'the outside world' or wander off into unrelated rambling, so the steering shows no clear theme.
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.
-
al. Take, for example, the crew of the 69-foot-long Mansor Sakho (she is painted sky blue). There is Ibrahima from Gambia. Fili from Dakar. Issa the Fulani,
-
villages of Ciucan and Tuta to give eye and dental tests.⏎For 14-year-old Tuta resident Annamaria Laura Lupo, the eye exam was long overdue. She said she's been having
-
The Geological Survey of India has identified granite waste as the construction material for groynes along vulnerable stretches of the Tamil Nadu coast.⏎GSI Deputy Director-General S.K. Bhushan told a conference on environmental geosciences organised by Anna University's
-
was born on 23 January, 1897 in Cuttack (Orissa) to Janakinath Bose and Prabhavati Devi. Janakinath Bose was one of the successful lawyer in Cuttack and
-
assassination of Lt Col NS Simpson, the infamous Inspector General of Prisons. Three Bengali revolutionaries – Benoy Basu, Badal Gupta and Dinesh Gupta – disguised themselves as Westerners to get inside the Writers’ Building
-
slavery. The woman, Hadijatou Mani Koraou, had been sold into slavery as a child and then forced to
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: It's not just the person who is in the room, but the space around them. But how to make it possible to achieve the meaning of the space, and the meaning of the person, without any prior information about the person or the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and controlled by an adjustable valve and a scaven<|endoftext|>Deutsch: Faktor / Español: Factor - Factor / Français: Facteur - Facteur / Italiano: Fattori - F
-
are designed to hold the workpiece in a specific position and orientation, allowing for consistent and accurate processing.⏎Deutsch: Frequenz / Español: Frecuencia / Português: Frequéncia / Français: Frêquence / Italiano
-
, so the period is the reciprocal of the frequency.⏎Deutsch: Fabrik / Español: Fábrica / Português: Fábrica / Français: Usine / Italiano: Fabbrica⏎A factory (pre
-
where workers manufacture goods or operate machines processing one product into another.⏎Deutsch: Fell / Español: Pelaje / Français: Fourrure⏎Fur is a synonym for hair, used in
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
- L18:1345 suppresses
- L18:1099 excites
- L14:1550 suppresses
- L18:1861 excites
- L18:2361 suppresses
- L18:2289 excites
- L18:269 suppresses
- L12:1276 suppresses
- L15:2128 suppresses
- L18:2513 suppresses
- L17:1257 suppresses
- L17:1387 excites
- L14:1796 suppresses
- L18:1199 excites
- L18:2225 suppresses
- L18:1536 excites end of web page disclaimer
- L18:2926 excites
- L18:891 suppresses
- L18:2732 excites
- L18:1477 weakly suppresses
- L18:1224 weakly suppresses
- L18:2539 weakly excites
- L17:2956 weakly suppresses
- L17:2726 weakly suppresses
Feeds later layers
- L22:2433 excites
- L20:551 excites
- L20:1038 excites
- L20:1377 excites
- L20:2345 weakly excites
- L20:1051 weakly excites
- L25:2226 weakly excites
- L20:573 weakly excites
- L21:627 weakly excites
- L20:3053 weakly suppresses
- L21:465 weakly suppresses
- L21:1754 weakly excites 'in' before a place name
- L23:2252 weakly excites
- L26:1368 weakly excites
- L25:1228 weakly excites
- L22:133 weakly suppresses
- L21:582 weakly excites
- L22:2278 weakly excites
- L23:2408 weakly excites
- L22:2317 weakly excites
- L25:98 weakly suppresses
- L24:2719 weakly excites
- L22:2062 weakly excites
- L20:1840 weakly 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
- %p
- Oversight
- %f
- lemetry
- vette
- eva
- ostringstream
- %c
- %H
- rowave
Suppresses
- ечно
- ::⏎
- iali
- Cla
- Cri
- ести
- 空间
- cla
- ENTIAL
- Joi