unclear: document ends and list breaks
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. 14 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 at unusual breaks in text, such as the end of a document, a number split by a symbol, or the start of a new clause. It is hard to say more than that, and turning it up made the model repeat itself.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
, towards the location of the monument. Where it is not available, the satellite view is shown instead.⏎Latitude: 52.5162 / 52°30'58"N⏎Longitude: -3.0
-
Pradesh, it is more comfortable to wear it over a shirt than as a single⏎piece of clothing.⏎Q4. What do you wear on veshti?⏎<|endoftext|>Copyright (c) Arvin S. Quist⏎INTRODUCTION TO CLASSIFICATION
-
Erica plukenetii, E. lineata and E breviflora. They have been sunk into E
-
Check out the full range of plans here: https://www.edplace.com/blog/tag/learning_plans.⏎What are the EdPlace learning plans?⏎Put simply, our teachers have created a half-termly (or 6 week) learning plan
-
bacteria bind to the sugar chain of UMOD and are eventually removed through the urine. During recent years, evidence has shown that glycoprotein 2 (GP2), a UMOD-like molecule produced in the pancreas
-
Studies using Geophysical techniques⏎- Implement and Development of Geophysical Inversion Methods⏎- Petroleum and Mineral Exploration⏎- Physics of the Earth and other Planets<|endoftext|>
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 am trying to find the number of the number of the number of the number of the number of the number of the number of the number of the number of the number of... the number of... the number of... the number of...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
creates metafiction.⏎One example of metafiction can be found on the very first page. Ondaatje transitions from writing in third person, using pronouns like "her" and "he," to writing in first person. True, it
-
first naval battle in which victory was decided by aircraft alone, Dauntless dive-bombers fought well alongside other US aircraft. Each side lost one carrier (the Japanese carrier Shoho being sunk by Dauntless and
-
82 until Zelaya’s ouster in June, the Honduran government appeared to be relatively stable. The country’s two major political parties, the moderate Liberal Party (of which Zelaya is a member) and the more
-
5 to 2008.⏎As a child in Cleveland, Ohio McKinney was shaped by his father who challenged racism, and invited civil rights leaders such as Thurgood Marshall to speak frequently at his church. McKinney attended Morehouse
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
- L12:1640 excites
- L7:307 excites
- L13:732 excites
- L13:1897 suppresses
- L9:557 suppresses
- L13:2181 excites
- L13:1314 excites
- L13:1483 weakly excites
- L8:481 weakly suppresses
- L13:700 weakly excites
- L13:1333 weakly excites
- L0:1637 weakly suppresses
- L11:3062 weakly suppresses
- L10:1459 weakly suppresses
- L10:681 weakly excites
- L12:2366 weakly suppresses
- L12:810 weakly excites
- L13:1376 weakly excites
- L7:231 weakly excites
- L1:136 weakly excites
- L9:1414 weakly suppresses opening parenthesis after copyright or year
- L13:648 weakly suppresses
- L10:1090 weakly excites
- L13:505 weakly suppresses
Feeds later layers
- L16:431 suppresses
- L15:2967 excites
- L16:413 suppresses
- L16:128 suppresses inside numbers and citations
- L16:468 suppresses
- L15:1653 excites
- L16:267 suppresses
- L16:737 weakly suppresses
- L15:1452 weakly excites
- L16:2161 weakly excites
- L16:860 weakly excites
- L16:648 weakly suppresses
- L23:2096 weakly excites
- L15:2616 weakly excites
- L16:2700 weakly excites
- L15:131 weakly suppresses
- L16:1906 weakly excites
- L16:349 weakly suppresses
- L18:1009 weakly suppresses
- L17:503 weakly suppresses
- L19:347 weakly excites
- L15:724 weakly suppresses
- L25:1403 weakly excites
- L16:665 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
- 欺
- リスク
- allocation
- lover
- (New
- ontology
- ||(
- 莫斯
- �建
- 未来发展
Suppresses
- śni
- uble
- among
- uci
- 在中国
- urch
- صة
- erra
- 特朗
- ebra