'of' after a person's name
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. 13 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
Fires on the word 'of' that comes right after a scientist's or author's name and introduces their institution, as in 'Michael Liu of the University of Hawaii'. It is quiet, or even pushed the other way, after country names like 'Republic of Yemen'. Turning it up produced repetitive text that does not show this clearly.
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.
-
brown dwarfs and much closer to those of giant exoplanets — it could even have water clouds in its atmosphere," the paper's lead author, Michael Liu of the University of Hawaii's Institute for Astronomy, said
-
from an ancient burial site on Crete, in a cave called Hagios Charalambos, a team led by archaeogeneticist Gunnar Neumann of the Max Planck Institute for Evolutionary Anthropology in Germany found genetic evidence of
-
4-Year Institutions⏎Shifting Students’ Stereotypes of Scientists to Enhance Science Identity in a Diverse, Community College Context⏎Jeff Schinske, Amanda Snyder, Heather Perkins, Mary Wyner, Monica Cardenas, Jahana
-
use this method to find out exactly what goes wrong in the nerve cells in movement disorders such as ataxias," said Prof. Dr. Stefan Herlitze (RUB Department for Biology and Biotechnology).⏎The researchers report is in the Journal
-
potential role of biomarkers in tracking neuronal recovery, which may be associated with duration of [return to sport],” wrote Cassandra L. Pattinson, PhD, of the University of Queensland, Brisbane, Australia, and the National Institutes of Health,
-
a has provoked alarm around the world, and there are no new classes of antibiotics in development.⏎In this week's Scrubbing Up column, Peter Greenhouse of the British Association for Sexual Health & HIV (BASHH) argues financial incentives will
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: "I have a thought for the new year, and the new year is the year of the new data, and the new year is the year of the new data, and the new data, and the new data, and the new data
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
declared independence and became known as the People's Democratic Republic of Yemen (South Yemen), which united with the North in 1990
-
is 884 with a serial no. of 00139⏎Kindly Donated by Callum Wilson.⏎With thanks to http://www.dvq.com/oldcomp/micros.htm for the original photograph without
-
1986, only 14 cases pertaining to blasphemy were reported.⏎By its constitution, the official name of Pakistan is the "Islamic Republic of Pakistan." More than 96% of Pakistan's 167 million
-
every human need. In Psalms we read, “Give thanks to the Lord, for he is good; his love endures forever. Let the redeemed of the Lord tell their story. . . . Let them give thanks to
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
- L12:2348 excites
- L12:2314 excites
- L12:1817 excites
- L12:1768 suppresses
- L10:2362 suppresses
- L12:139 excites
- L12:2889 suppresses
- L12:1763 excites
- L11:2542 excites
- L9:2390 excites
- L8:1107 suppresses
- L9:1627 excites
- L11:1418 suppresses
- L12:2664 weakly suppresses
- L12:2810 weakly excites
- L11:1994 weakly excites
- L12:273 weakly excites
- L12:1673 weakly suppresses
- L12:2289 weakly excites 'of' after part-whole nouns
- L12:1794 weakly suppresses
- L12:2456 weakly excites
- L9:1961 weakly suppresses 'of' in titles and place names
- L12:1286 weakly suppresses
- L10:2529 weakly excites
Feeds later layers
- L17:2906 excites
- L16:2427 excites
- L19:2853 excites
- L18:1931 excites
- L17:978 excites
- L21:2846 excites 'at' after a job title, before an institution
- L17:3041 excites
- L16:1051 excites
- L17:604 excites
- L14:2038 weakly suppresses
- L15:1456 weakly excites
- L18:2349 weakly suppresses
- L14:1707 weakly excites
- L18:56 weakly excites
- L20:1038 weakly excites
- L19:817 weakly excites
- L14:1550 weakly excites
- L20:2623 weakly excites
- L22:2134 weakly excites
- L19:2135 weakly excites
- L19:1001 weakly excites "be" after will or you'll
- L20:862 weakly excites
- L16:2716 weakly excites
- L14:1893 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
- .local
- USDA
- Cam
- 教育部
- _cam
- .microsoft
- 市教育局
- ording
- airs
- 局
Suppresses
- 通行证
- wort
- тиров
- akter
- 嬗
- 胡子
- hetics
- 瑁
- uell
- ANTI