"of" or "for" after minister or ministry titles
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 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
This neuron fires on the small words "of" or "for" that come right after a government job title, like "Minister of Health" or "Minister for Housing". Turned up, the model's output became rambling and repetitive, not clearly about government.
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.
-
a a third time, this time in a shorter version with only <|endoftext|>ATHENS – The digital exhibition Proving History, presented by the Hellenic Ministry of
-
vice-chancellor, probably overwhelmed by the occasion, asked the minister of state for petroleum and natural gas and the minister of state for external affairs to help the university procure an Army tank for the campus. Clearly,
-
70 Elementary Schooling Act was the efficient begin of state-financed training in England and Wales. The Secretary of State initially simply took the place of the Minister for Science as chair of the various Privy Council committees for civil science. It emphasizes
-
.⏎problem6. Compare and contrast a peer-to-peer network architecture to Client-Server architecture. Support your answer with proper diagrams and exs.<|endoftext|>The Minister for Housing, Local Government and Heritage, Darragh O’Brien, and the Minister
-
emy laws (Salman Taseer, the former governor of Punjab, and Shahbaz Bhatti, the Federal Minister for Minorities) have been assassinated. Since 1990,
-
rights and freedoms,” National Congress of Australia’s First Peoples co-chair Rod Little said.⏎Assistant Minister for Social Services and Multicultural Affairs Minister Zed Seselja described the UN criticisms
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 question is, what is the answer to this problem? If $ x $ is 2, then $ x - 1 = 2 - 1 = 2 $, then $ x - 3 = 2 $
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Central Committee of the Bolshevik party.⏎The Bolshevik party were a minority but gained publicity after being the only
-
Science practices meet their potential in bridging the world’s STI gaps and enabling sustainable development. The Member States of the Organisation have examined this issue in the UNESCO Executive Board:⏎- Preliminary study of the technical, financial and legal aspects on the
-
of a species of fish (Mitsukurina owstoni) from Japan, the type of a distinct family of lamnoid sharks". Proceedings of the California Academy of Sciences (Series 3) Zoology 1 (6):
-
researchers have demonstrated that it is not. In 1970, dentist Wallace Johnson and others published the results of their research in Journal of the American Dental Association. The report indicated that "mercury, an essential element 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
- L7:1131 excites
- L3:1237 excites
- L6:1742 weakly excites
- L3:1336 weakly suppresses
- L4:581 weakly excites
- L7:2067 weakly excites
- L6:2775 weakly excites
- L7:1780 weakly excites
- L3:1919 weakly suppresses timestamps after dates
- L6:2579 weakly excites
- L7:1049 weakly suppresses
- L4:2674 weakly excites
- L3:1851 weakly suppresses
- L7:2994 weakly excites
- L4:2063 weakly suppresses
- L5:1156 weakly excites
- L0:901 weakly suppresses 'this' after sentence-initial 'In'
- L7:1083 weakly excites
- L4:2365 weakly excites
- L4:2229 weakly suppresses
- L4:2617 weakly suppresses
- L5:1146 weakly excites
- L3:2671 weakly suppresses
- L2:3066 weakly suppresses
Feeds later layers
- L9:1505 excites
- L12:1612 excites
- L16:1855 excites
- L13:1181 excites 'of' after a definition or noun
- L16:2715 excites
- L13:1150 excites
- L15:2810 excites
- L20:697 excites
- L11:1246 excites
- L13:1651 excites
- L18:2179 excites
- L11:2298 excites
- L12:2737 excites
- L13:2665 excites
- L9:1601 suppresses
- L13:408 excites
- L20:1850 weakly excites
- L23:1321 weakly excites
- L19:2827 weakly excites
- L16:1380 weakly excites
- L17:2907 weakly excites
- L9:663 weakly suppresses
- L16:2216 weakly excites
- L9:2190 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
- ervice
- serviço
- overseeing
- orea
- Svc
- cope
- .getComponent
- kitchens
- .getService
- Servi
Suppresses
- изм
- �新
- flyer
- quotes
- cff
- figcaption
- ---</
- ,filename
- 份
- 램