colon before a subtitle
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. 22 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 a colon that separates a title from its subtitle, like the colon in a book or paper title. When it was turned up, the model wrote a lot of 'A:' and 'I:' labels, which matches its focus on colons.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
.L., & REIJMERSDAL, E.A. (2012). Social connections and the persuasiveness of viral campaigns in social network sites: Persuasive intent as the underlying mechanism. Journal of Marketing, 18
-
reasoning is used as the context in which the contributors elaborate on their own perceptions.⏎Table of Contents⏎Contents: W.F. Overton, Competence and Procedures: Constraints on the Development of Logical Reasoning. B. Inhelder,
-
gencies, and International Intervention Chapter 3 Gender and the Causes and Consequences of Armed Conflict Chapter 4 Gender Mainstreaming in United Nations Peacekeeping Operations: Talking the Talk, Tripping over the Walk Part 5 Gender, Peacekeeping
-
ing, Logic, and Thought Disorders: Deductive Reasoning and Developmental Psychopathology. D.P. Keating, Structuralism, Deconstruction, Reconstruction: The Limits of Reasoning.⏎".....Of interest to serious advanced students of cognitive
-
9, the study, “Global, regional, and national prevalence of overweight and obesity in children and adults during 1980–2013: a systematic analysis for the Global Burden of Disease Study 2013
-
of outbreak response⏎Public Health Responses to a Dengue Outbreak in a Fragile State: A Case Study of Nepal⏎This case study illustrates the principles taught in Lecture
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: how he so ...he so he so he so he so he so he so he so he so he so he so he so he so he so he so he
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
victory of good over evil. Dussehra or Vijaya Dashami or Dasara is marked on the 10th day of the bright half (Shukla Paksha) of the month of Ashvin (Ashwayuja),
-
by Muhammad Ali Jinnah from the Muslim league, B R Ambedkar campaigning for the rights of Dalits, and Master Tara Singh from SAD (Shiromani Akali Deal). This was the first-ever articulation for the
-
Government of Nepal gave the mountain the official Nepali name. In sanskrit Sagarmatha means "mother of the universe”. The Tibetan name for Mount Everest is Chomolungma or
-
stices, equinoxes, and phases of the moon for Vedic rituals. Eighteen early astronomical texts or siddhantas, of which only the Surya-Siddhantha, written around 400 BCE, survives
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
- L17:1362 excites unclear: book and report titles
- L20:1016 excites
- L12:2494 suppresses
- L19:2302 suppresses
- L15:2532 suppresses
- L13:2123 weakly excites
- L15:1066 weakly suppresses
- L4:30 weakly excites
- L18:1338 weakly excites
- L15:2459 weakly suppresses
- L10:2801 weakly excites
- L14:1909 weakly excites
- L10:270 weakly suppresses
- L5:2838 weakly excites
- L15:345 weakly excites
- L5:1440 weakly suppresses
- L13:1325 weakly excites
- L7:1356 weakly suppresses
- L17:621 weakly suppresses
- L10:2002 weakly excites
- L3:233 weakly excites
- L1:2083 weakly excites
- L16:1652 weakly suppresses
- L10:1661 weakly suppresses
Feeds later layers
- L23:236 excites
- L25:745 excites
- L26:648 suppresses
- L24:322 excites
- L26:122 excites
- L25:45 suppresses
- L27:858 excites
- L25:4 suppresses
- L25:733 weakly excites
- L27:1103 weakly excites
- L27:663 weakly excites
- L26:582 weakly excites
- L25:837 weakly excites
- L27:646 weakly excites
- L26:1229 weakly suppresses
- L27:960 weakly excites
- L24:864 weakly suppresses 'to' before an infinitive
- L24:823 weakly excites
- L27:797 weakly excites
- L25:440 weakly suppresses
- L25:1132 weakly suppresses
- L26:2432 weakly suppresses
- L25:1586 weakly excites
- L26:778 weakly excites capital letters inside acronyms
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
- :
- :
- :a
- :A
- :X
- :user
- part
- :Is
- :end
- :x
Suppresses
- 率为
- .HandlerFunc
- Handle
- 严格按照
- 着他
- ⎝
- titleLabel
- Verfügung
- hers
- 情况进行