'and' in lists of places or 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. 16 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 'and' when it joins the last two items in a list, especially lists of countries, places or names. Turning it up gave text full of list-like repetition, but the link to the trigger is weak.
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.
-
standards will be piloted in Botswana, Egypt, Ethiopia, Kenya and Paraguay. The initiative contributes to WHO’s 13th General Program of
-
based on a priori hypotheses, to examine intraspecific differences in survival probabilities relative to individual covariates (sex, size, size standardized by sex, and overwintering location). Annual survival in 2003 and 2
-
the world. This will not only benefit biodiversity, but will also help build local resilience to the increasing impacts of climate change. Three examples from the Philippines, Indonesia and Kenya are presented here.⏎Mount Siburan Important Bird Area (IBA)
-
th and adults, and their families, who are dealing with anorexia nervosa, bulimia nervosa, avoidant/restrictive food intake disorder and binge eating disorder.⏎- Clients are referred into our program by their family doctors.⏎
-
-888-782-7937)⏎For more information on activities ongoing within EPA Region 7 (Iowa, Kansas, Missouri and Nebraska), contact:<|endoftext|>According to the analysis of the cutting process oftooth
-
plan will be recommended.⏎- If needed, youth and adults will be referred to provincial tertiary facilities at St. Paul's Hospital, B.C. Children's Hospital and Looking Glass Residence.⏎- Kelty Mental Health: Eating disorders⏎Connects you
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: For example, for a different, or, for a different, or, for a different, or, for a different, or, for, for, for, for, for, for, for, for, for, for, for
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a significant effect of scene condition, with average threshold speed changes of 43%, 37% and 25% for conditions 1, 2 and 3, respectively. Thus, monocular depth cues aid the detection of
-
4; p < .001 for all). Through iterative thematic analysis we developed an alliterative “7P” model spanning 2 domains: (1) Participants (Performers and Peers); and (2) Program (Part
-
example were given wood and plaster in which to design and fabricate an audio speaker set.⏎Teaching Focus: The objective of this exercise was two-fold; (1) to have students work with two contrasting design materials- one common and easily manipulated
-
regardless of whether it is a speculative narrative or a narrative of emancipation (p. 37).⏎Two expressions in the quotes above catch my attention: (a) “It may be difficult for you to envision…at one time…the grand
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
- L14:1175 suppresses
- L14:3034 suppresses
- L14:3041 suppresses
- L13:749 suppresses
- L14:1174 suppresses
- L15:935 excites
- L13:2742 excites
- L12:2071 suppresses
- L15:1005 excites
- L14:1234 suppresses
- L12:2057 excites
- L15:137 suppresses
- L11:1674 suppresses
- L15:2322 suppresses
- L15:1504 excites
- L11:2760 suppresses
- L13:2022 excites
- L10:713 excites
- L14:841 excites commas inside lists
- L14:396 suppresses
- L13:1063 excites
- L11:467 suppresses
- L13:1361 suppresses
- L12:1530 suppresses
Feeds later layers
- L18:2898 suppresses
- L18:1693 suppresses
- L17:2891 suppresses
- L24:1087 suppresses
- L18:2722 suppresses
- L19:1414 suppresses
- L19:1265 suppresses
- L18:2446 suppresses
- L19:2584 suppresses
- L18:1327 suppresses
- L18:957 excites
- L19:492 excites
- L17:414 excites
- L19:2999 suppresses
- L19:1092 excites
- L20:1441 suppresses
- L18:2589 suppresses
- L21:2512 suppresses
- L19:2552 suppresses
- L19:1813 suppresses
- L23:1798 suppresses
- L27:291 suppresses
- L17:1384 weakly suppresses
- L17:1510 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
- 三个
- 两个
- 一个是
- 两种
- 三个方面
- 两条
- 两次
- 三种
- 兩個
- 两点
Suppresses
- etc
- etc
- 等等
- 等
- 等各种
- 等诸多
- など
- 等地
- 等功能
- 等一批