"of" after member, version, or part 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. 7 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly 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
The neuron lights up on the word "of" when it follows a noun like "members", "versions", "ends" or "aspects", as in "members of the family". It switches off on "in" after years or events, such as "founding, in 1817". Turning it up made the model's writing wander into unrelated topics, so the steering did not show a 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.
-
like a miniature version of its full-sized species.⏎And to achieve this look, you’ll need to maintain
-
it? What is the setting? What happens? What do you like/not like about the way the author writes?⏎Word by word⏎Similar but more formal versions of this should be happening at school, but parents can reinforce this learning at home.
-
an impartial humanitarian organization. It uses the information obtained to identify civilians and members of a paramilitary force belonging to a specific ethnic group in order to arrest or kill
-
environment. From their lineage to their habitat, spiny king crabs are extremely interesting crustaceans.⏎Spiny king crabs are invertebrates and members of the king
-
all people, and many members of our society are still striving to achieve them.⏎“We hope this Convention will be followed
-
(Buckminster-Fuller, 1969)⏎The buckyball is the simplest of the carbon structures known as fullerenes. Members of the fullerene family are a major subject of nanotechnology research. 3-D
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: of the year, 2001-2004-05-06-07-88-88-87-77-71-75-77-73-
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
colony on African soil. The success of Cuffee’s venture led to the founding, in 1817, of the American Colonization Society, with the mission
-
8 filing of a Citizen Submission with the Commission for Environmental Cooperation, an institution formed to fulfill the environmental provisions of NAFTA. The Citizen Submission culminated in the publication in 2002 of a report documenting the risk to human health that tox
-
”, a translation of Fr. Boccalini’s “Del Ragguagli di Parnasso”, 1612. Beginning with the fourth edition in 1615, the third Rosicrucian rudiment, “
-
they return to Hiva with the news that indeed,<|endoftext|>The intent of this work is five fold: 1) To spatially characterize and monitor the distribution, abundance, and size of
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
- L5:246 weakly suppresses
- L2:2451 weakly excites
- L5:1721 weakly excites
- L0:2630 weakly excites
- L3:155 weakly suppresses
- L3:1107 weakly suppresses
- L4:2853 weakly excites
- L6:1735 weakly suppresses
- L0:2028 weakly excites
- L2:2458 weakly excites
- L2:2772 weakly excites
- L3:1548 weakly suppresses
- L5:2115 weakly excites
- L2:2670 weakly suppresses
- L5:21 weakly excites
- L2:946 weakly suppresses
- L4:23 weakly excites
- L4:2448 weakly suppresses numerical computing and library names
- L0:695 weakly excites
- L2:428 weakly suppresses
- L0:554 weakly excites
- L2:3028 weakly excites
- L5:239 weakly suppresses
- L1:2065 weakly excites
Feeds later layers
- L8:202 suppresses
- L9:1210 suppresses
- L8:1350 excites
- L22:142 excites 'of' after a quantity or noun phrase
- L18:953 suppresses
- L16:961 suppresses
- L15:2736 suppresses
- L8:526 suppresses
- L13:1942 suppresses
- L8:1867 suppresses immediately repeated words
- L17:1771 suppresses
- L27:658 suppresses
- L26:711 suppresses 'years' after an age number
- L23:567 weakly suppresses
- L15:1474 weakly suppresses
- L12:2641 weakly suppresses
- L20:2599 weakly suppresses
- L19:1080 weakly suppresses
- L25:2020 weakly suppresses
- L25:1497 weakly excites
- L22:288 weakly suppresses
- L15:894 weakly suppresses
- L24:2219 weakly excites
- L15:2143 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
- 记者在
- 人死亡
- 这些问题
- 市政府
- 省政府
- 人在
- ьев
- DeltaTime
- elerik
- 他是一个
Suppresses
- of
- _of
- -of
- Of
- Of
- OF
- of
- oks
- _OF
- của