'of' in titles and place 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. 9 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 lights up on the word 'of' when it links a job title or place to a name, as in 'Bishop of London' or 'City of Opelika'. Its opposite is the word 'from' after a noun, as in 'residues from'. Turning it up did not make the model's writing clearly about titles or places.
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.
-
could hardly be found to bury the dead, and many thousand bodies were carelessly thrown into mere pits dug in the open fields.⏎Ralph Stratford, Bishop of London, shocked at these unsanctified interments, in his zeal to
-
1472, Niccolò Perotti (1430-1480), secretary to Cardinal Bessarion and archbishop of Siponto, donated to Sassoferrato, the city of his birth
-
century.⏎The site lay waste for about 150 years but, in the mid-11th century, Robert, bishop of Coutances, undertook the rebuilding of the cathedral in the Romanesque or
-
with the study of literature.<|endoftext|>The two Supreme Court cases, Jones v. City of Opelika 316 U.S. 584 (1
-
C. (Image via Zach Montellaro on Flickr, CC BY-NC-SA 2.0)⏎The two Supreme Court cases, Jones v. City of Opelika 316 U.S. 584 (1
-
10, 2011⏎At 3pm on the 10th June 1540 Thomas Cromwell, the Earl of Essex and Henry VIII’s right-hand man, was in the council chamber at Westminster
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 user of the previous year. I am not sure of the previous year. It's like the user of the previous year... the user of the previous year... the user of the previous year... the user of the previous year... the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
personally how these gardens evolved under their care. The walled garden at Parham was probably created in the 18th century and paths and brick walls survive from that period. It must always have been a productive garden, supplying food and flowers
-
of cesium due to leakage from the Fukushima power plant, depending on origin. See: Arsenic, Methylmercury.⏎Shellfish may contain residues from chemicals that have filtered or run off into the water in which they inhabit. Consumers
-
of the old Gold Hill Hotel where Nancy Melton became operator⏎One of the tombstones in the Gold Hill cemetery, Ephraim Mauney, was from Stanly County⏎After the death of John Sr., and Joseph Sr. ,
-
distance, are the only remaining relics of this green cathedral.⏎More cemeteries, designed by John Bergmans,
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
- L4:192 excites
- L6:2937 excites 'by' after a title or author credit
- L4:2859 weakly excites
- L0:1660 weakly suppresses
- L2:1147 weakly suppresses
- L3:1027 weakly excites
- L7:1970 weakly excites
- L0:993 weakly excites
- L8:451 weakly suppresses
- L5:1922 weakly suppresses 'to' in 'X to Y' phrases
- L8:16 weakly suppresses
- L4:2404 weakly excites
- L4:2839 weakly suppresses
- L4:2717 weakly excites
- L4:1721 weakly suppresses the gap before a decade or century
- L6:454 weakly excites
- L6:1604 weakly excites 'of' after 'that' in comparisons
- L8:685 weakly suppresses
- L7:2248 weakly excites
- L7:1054 weakly suppresses
- L5:2984 weakly suppresses
- L8:2397 weakly excites
- L6:2664 weakly excites
- L5:939 weakly suppresses
Feeds later layers
- L14:2007 suppresses
- L16:2196 excites
- L17:550 excites
- L15:2682 weakly suppresses
- L13:775 weakly suppresses
- L16:2087 weakly suppresses
- L12:2889 weakly suppresses
- L13:558 weakly suppresses
- L22:1326 weakly excites
- L13:894 weakly suppresses 'of' after a person's name
- L16:2780 weakly excites period after 'Dr' abbreviation
- L23:1931 weakly suppresses
- L10:1054 weakly suppresses 'from' after movement or transfer verbs
- L20:1804 weakly suppresses
- L12:2671 weakly suppresses
- L17:189 weakly suppresses
- L16:2530 weakly suppresses
- L14:1094 weakly suppresses
- L12:2850 weakly suppresses
- L14:2460 weakly suppresses
- L19:2298 weakly suppresses
- L10:132 weakly suppresses
- L18:1979 weakly suppresses
- L16:593 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
- weise
- パス
- .addCell
- ppers
- pute
- oppers
- 信じ
- ContentPane
- .addChild
- /topics
Suppresses
- SourceType
- aida
- haunt
- ю
- toa
- ainting
- 来源
- 吉利
- AEA
- clid