blank spot where a number goes
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. 17 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 responds to the empty space just before a number in citations, statistics and counts, such as "volume 7" or "about 30 per day." Its opposite side fires on the spot before a second number in a list. Turning it up or down gave writing that did not clearly relate to numbers.
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.
-
narica and Procyon lotor (Carnivora: Procyonidae): evidence of infection control?⏎Parasites & Vectors volume 7, Article number: 405 (2014)⏎A
-
might also be interested in:⏎About 70<|endoftext|>- freely available⏎Int. J. Environ. Res. Public Health 2012, 9(11), 4170-4186;
-
BMC Musculoskeletal Disorders volume 4, Article number: 4 (2003)⏎The purpose of this study is to present an
-
like the one examined in the current study. This could be because cancer of the salivary gland is normally quite rare, accounting, for example, for only 0.23 percent of all cancers diagnosed in the United Kingdom each year.
-
and 100 million Americans had received their COVID-19 shots, there were 3,544 reported deaths following COVID vaccination, or about 30 per day⏎- In just four months, the COVID-19
-
2) of it (0.67%) is water.⏎As of the census of 2000, there were 112 people, 46 households, and 32 families residing in the town. The population
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 have a 300% change in my work I am doing in my lab. My work is only possible because my work I have been doing in my head is changed, and my research I have I made up all of my
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the goals of their academic programs.⏎Films and film clips⏎The site features 120 films from the NFB collection (60 in English; 60 in French in the French part). Of those, 48 (
-
in French in the French part) are available with closed-captioning. Users will also find 164 film clips (80 in English; 84 in French in the French part).⏎These 120 films are
-
December 2015. Overall, 99 sources from 62 countries were retrieved: 53 from 22 developed countries, and 46 from 40 developing countries. Representative guidelines from America (the USA
-
9 square miles (1,710 km2), of which 576 square miles (1,490 km2) is land and 83 square miles (210 km2) (13%) is
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
- L16:256 excites number right after 'age of' or 'ages'
- L10:868 suppresses
- L10:220 suppresses hyphen inside compound words
- L13:1372 excites number before an age or week count
- L14:471 weakly excites
- L7:110 weakly suppresses
- L15:1285 weakly suppresses
- L16:372 weakly excites
- L6:756 weakly suppresses
- L3:122 weakly excites
- L2:75 weakly suppresses
- L15:2091 weakly excites
- L8:1745 weakly excites
- L14:1793 weakly suppresses
- L15:167 weakly excites
- L7:1524 weakly suppresses
- L6:1167 weakly excites
- L10:1184 weakly excites
- L3:59 weakly excites
- L15:1909 weakly suppresses blank slot before a number
- L10:2839 weakly suppresses
- L12:1420 weakly excites digits starting a percentage or rate
- L16:903 weakly excites
- L9:183 weakly suppresses
Feeds later layers
- L20:2749 weakly excites
- L18:2388 weakly excites
- L19:2500 weakly excites
- L18:2694 weakly excites
- L22:1436 weakly suppresses
- L22:1118 weakly suppresses
- L19:2011 weakly excites
- L26:237 weakly suppresses
- L26:1270 weakly excites
- L18:2844 weakly excites
- L21:1172 weakly suppresses
- L25:2783 weakly excites
- L18:1728 weakly excites
- L18:983 weakly excites
- L19:3018 weakly excites
- L18:659 weakly suppresses
- L25:2104 weakly excites
- L21:771 weakly suppresses
- L18:2454 weakly excites
- L20:1791 weakly suppresses
- L21:558 weakly excites
- L20:2470 weakly suppresses
- L21:2133 weakly excites
- L18:2188 weakly excites
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
- 芥
- 不方便
- -next
- 疤
- �
- Interstate
- 覚
- olutely
- ictionaries
- _inline
Suppresses
- rior
- ода
- inker
- 研究成果
- 果汁
- nio
- sky
- ratio
- 所所
- ertest