unclear: time words like 'later' and 'years'
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. 15 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 mix of words that often end a short phrase, such as 'later', 'years', 'day' and 'page'. It is hard to find one clear theme, and turning it up gave rambling, off-topic text.
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.
-
of liquid as a bead, like a bead of sweat on your forehead. The earliest meaning of bead is "prayer bead," from the Old English gebed, "prayer."<|endoftext|>Solar geoengineering (SG) and CO2 emissions reduction
-
hard drive in the same PC.⏎Note: always be cautious when using dd, as it’s very easy to completely wipe out a drive, as we will show later in this article.⏎Creating a Drive Image⏎Boot up into the Ubuntu Live CD
-
directly behind his backyard in 1889.⏎Bailey can be seen sitting in that forcing house in the cyanotype photograph reproduced at the top of this page<|endoftext|>August 21, 2015⏎Station name board in
-
(92 mother tongues) The majority of the population are Hindu Adult Literacy Rate (6 and over): 53% Life expectancy: 68 years Home of Mt. Everest and the highest altitude lake in the world (Til
-
for Nurses in New York.⏎1916-1918: Mary Breckinridge’s newborn daughter, Polly, dies, followed two years later by the unexpected death of her four-year-old son, Breckie. The
-
blind, or homebound person in a hospital or in an extended-care⏎- Perform volunteer<|endoftext|>Nanocrystals May Provide Boost for Solar Cells, Solar Hydrogen Production⏎5 January 2006⏎Los Alamos National Laboratory scientists
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 think I was thinking about the same thing, but I was thinking of the same thing. So, I think that was something, but I was thinking of the same thing. I think I was thinking of the same thing. So,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2 Diabetes⏎Wheat bran cereal is one of the richest plant sources of magnesium. This mineral supports the activities of more than 300 enzymes, including enzymes involved in secretion of insulin and utilization of glucose. Studies have established that wheat bran
-
the U.S. Constitution.⏎In this year’s Religious Freedom Day proclamation, President Barack Obama writes,⏎Today, America embraces people of all faiths and of no faith. We are Christians and Jews, Muslims and Hindus, Buddhists and Sikhs
-
, which specifies where the runtime can locate an assembly (for more information, see Specifying an Assembly's Location).⏎<codeBase version="188.8.131.52" href="http://www.litware
-
children are naturally independent, with a strong need to express their own will. They test any boundaries set for them, including those set by their parents. These children can grow up to be rule-breakers, creative
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
- L3:294 suppresses
- L8:10 suppresses
- L8:7 suppresses
- L13:23 weakly suppresses
- L14:57 weakly suppresses
- L9:70 weakly excites
- L14:0 weakly suppresses
- L14:252 weakly excites
- L3:59 weakly suppresses
- L6:8 weakly excites
- L12:625 weakly excites
- L12:110 weakly suppresses
- L11:1536 weakly excites
- L12:940 weakly excites
- L12:1835 weakly suppresses
- L1:87 weakly excites
- L1:2990 weakly excites
- L7:62 weakly suppresses
- L12:2289 weakly excites 'of' after part-whole nouns
- L7:102 weakly excites
- L13:122 weakly suppresses
- L14:597 weakly suppresses unclear: document ends and list breaks
- L5:169 weakly suppresses
- L14:440 weakly suppresses
Feeds later layers
- L18:319 suppresses
- L16:309 weakly suppresses
- L19:2883 weakly excites
- L17:1985 weakly excites
- L18:2681 weakly excites
- L22:299 weakly excites
- L17:203 weakly suppresses
- L19:2542 weakly excites
- L20:2813 weakly excites
- L16:775 weakly suppresses
- L16:2599 weakly suppresses
- L16:412 weakly suppresses
- L18:1613 weakly excites
- L20:1979 weakly suppresses
- L23:1931 weakly excites
- L26:1803 weakly suppresses
- L21:1888 weakly excites
- L16:2074 weakly excites
- L25:184 weakly excites
- L16:1049 weakly suppresses
- L21:2678 weakly excites
- L16:1011 weakly suppresses
- L16:2486 weakly suppresses
- L18:549 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
- /light
- .neo
- /Internal
- НЕ
- ynet
- 说了
- ologue
- ""; ⏎
- kest
- egin
Suppresses
- orque
- urally
- isque
- oque
- incare
- arkadaş
- 队员们
- koş
- Separator
- ッ