opening quote marks before a coined term
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. 18 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 opening quotation mark that introduces a nickname or coined term, such as the quote before "Great Migration" or "mastiche". It is negative on the word "to" in menu-style link text like "Books to Print" or "Email to a Friend". Turning it up or down produced unrelated text, so the steering doesn't confirm this.
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.
-
economy?⏎By the 1950s, the flow of Puerto Ricans to the mainland United States had increased so drastically that historians dubbed the phenomenon the “Great Migration.” An estimated 470,000 people—or
-
culture, we have to start with ancient Greece, where citizens routinely chewed on the resin (or sap) from the Mastic tree. This substance was dubbed “mastiche
-
Prior to the restructuring of vineyards of 1955-60, one of the areas in which it was divided Sicily was called "Nero d'Avola and Frappato", coinciding with the provinces of Sir
-
in their September 1915 issue. The use of aerial detection to spot and report forest fires—dubbed the “Wisconsin Plan”—soon became a vital tool in fighting<|endoftext|>Books and Movies⏎
-
the fitness industry. As does the “no pain, no gain” myth, which came to prominence in the early 1980
-
training as part of ongoing efforts to implement NCA.⏎ONS recently published estimates of the total value of all the assets held in the UK, known as the ‘National Balance
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 ****************************************************************************************
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
00 Web Pages⏎Sample Pages for Prospective Subscribers, or click below⏎Overview of Site|⏎Enchanted Learning Home⏎Monthly Activity Calendar⏎Books to Print⏎Parts of Speech⏎The Test of Time⏎TapQuiz Maps - free
-
1587). Among those who studied under Mercier was the Huguenot leader and author Philippe<|endoftext|>|Back to Main Print This Page Email to a Friend|⏎Pleural fluid analysis is a test that examines a sample of
-
works to give pride of place to those of high heritage value.⏎How the site is organized⏎The Across Cultures Web site is divided into five sections: Themes to Explore; Cinema and Representation; Points of View; See Everything, Hear Everything;
-
1:30 AM (CANBERRA TIME) 24/04/2001⏎|Page tools: Print Page RSS Search this Product|⏎This variable identifies holders of Australian citizenship.⏎Citizenship data are used
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
- L13:1632 strongly suppresses words inside headlines and titles
- L14:790 suppresses
- L15:1865 suppresses
- L16:1652 suppresses
- L14:895 excites
- L15:2459 suppresses
- L17:1362 excites unclear: book and report titles
- L17:2414 excites
- L12:2654 excites
- L16:2398 excites
- L14:2061 excites
- L16:1086 suppresses
- L10:270 suppresses
- L15:2591 excites
- L4:887 suppresses
- L17:114 suppresses
- L14:2130 suppresses
- L14:1311 excites
- L12:2494 suppresses
- L8:311 excites
- L17:743 excites
- L14:2884 excites
- L14:853 suppresses
- L11:45 suppresses
Feeds later layers
- L19:950 suppresses
- L19:2883 suppresses
- L20:495 suppresses
- L22:268 suppresses
- L21:647 suppresses
- L21:1287 excites
- L20:36 weakly excites
- L19:1707 weakly excites
- L19:1110 weakly suppresses
- L20:739 weakly suppresses
- L21:1133 weakly suppresses
- L27:100 weakly suppresses
- L22:748 weakly suppresses 'and so on' list endings
- L22:431 weakly suppresses
- L19:2022 weakly suppresses
- L20:755 weakly excites
- L25:487 weakly suppresses
- L20:570 weakly suppresses
- L19:1946 weakly suppresses
- L26:1394 weakly excites
- L20:1981 weakly excites
- L19:438 weakly suppresses
- L21:511 weakly excites
- L20:1726 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
- CreateTable
- idth
- uran
- doctype
- isters
- (styles
- SupportedContent
- 方式进行
- identally
- 樘
Suppresses
- Others
- onomy
- 城
- Ratio
- Fact
- حة
- &A
- حت
- ception
- &