the 'An' syllable starting a word
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. 23 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 the letters 'An' when they begin a longer word or name, such as Anhui, Anjou or Anuradha. When it was turned up, the model fell into repeating brackets and broken text instead of writing sentences.
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.
-
and China have projected. Bruce Fitt from the University of Hertfordshire in Hatfield, UK, and his teammates forecast levels of the disease in the Anhui and Hubei provinces from 2021-20
-
Charles II of Spain died in 1700 he bequeathed his throne and the Spanish possessions in the Netherlands, Italy and the Americas to Philip of Anjou, the grandson of King Louis XIV of France. The prospect of a
-
reference to the Roman practice of providing free wheat to Roman citizens as well as costly circus games and other forms of entertainment as a means of gaining political power. The Annona (grain dole) was begun
-
factory, which was a one-piece brass measuring 33.5 meters (109’) tall. This flagpole is still the tallest in service at Anitkabir, the mausoleum built in honor of Atatur
-
.⏎Richards’ doodle eventually made it into the hands of chemists at the University of Warwick. After several months olympicene was finally produced by Anish Mistry, from the David Fox group, using a multi-step synthesis.
-
defines a "family entity" as "a stable union between a man and a woman."2 The Catholic Church responded to the recent ruling with outrage. As Archbishop Anuar Battisti put it, the Supreme Court's decision marked a "frontal
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: The number of time it takes to read a page of 444444444444444444444444444444444444
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
postlude, or interlude, or any short instrumental work.⏎note2: Many symphonies are tonal works in four movements with the first in sonata form, which is often
-
system of the Nazi regime. Since the end of the war there has been a strong temptation to dismiss Nazism as an anti-western aberration, the exact antithesis of the values of the Enlightenment and liberal democracy. This reassuring diagnosis was
-
's blood (although we have to take two or more sample days or weeks apart and show that the RSV antibodies have increased). We can also use antigen-antibody tests to look for the virus: these tests give
-
integral part. In fact any happy occasion is seen as an excuse to imbibe in this "tonic water".⏎It is in this historical context that we, the Shirayuki
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
- L21:2093 suppresses
- L18:282 excites
- L20:1414 excites
- L17:391 excites
- L16:55 excites
- L12:2264 weakly excites
- L2:1923 weakly suppresses
- L20:1228 weakly excites
- L5:499 weakly excites
- L20:959 weakly excites
- L16:217 weakly excites
- L0:2545 weakly excites
- L21:2247 weakly excites
- L17:1082 weakly excites
- L7:287 weakly excites
- L5:1251 weakly excites
- L20:2206 weakly suppresses
- L22:1837 weakly excites
- L22:384 weakly suppresses
- L1:27 weakly excites
- L0:60 weakly suppresses
- L16:166 weakly excites
- L10:1490 weakly excites
- L0:3063 weakly suppresses
Feeds later layers
- L25:1700 excites
- L26:739 weakly suppresses
- L27:391 weakly suppresses
- L24:3051 weakly suppresses
- L25:2519 weakly suppresses
- L25:708 weakly excites
- L24:1571 weakly excites
- L24:1580 weakly excites
- L24:424 weakly excites
- L27:524 weakly excites
- L26:1075 weakly suppresses
- L25:1378 weakly excites
- L25:1584 weakly excites art and artist names in lists
- L27:884 weakly suppresses
- L26:1833 weakly excites
- L25:1154 weakly excites
- L24:1229 weakly excites
- L25:2478 weakly excites
- L24:770 weakly suppresses
- L24:196 weakly suppresses
- L24:2704 weakly excites
- L26:2368 weakly suppresses
- L27:1940 weakly suppresses
- L27:752 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
- h
- hair
- haft
- hra
- hm
- hist
- [
- hil
- essor
- hani
Suppresses
- pector
- eturn
- 集团股份
- BorderSide
- /display
- 全面发展
- -</
- arden
- onSelect
- mouseleave