the '0' in years like 2008
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 lights up on the zero in years that start with 20, such as 2004 or 2013. It goes quiet or negative on the 1 in years like 1889 or 1939. Turning it down made the model output strings of repeated 9s and 7s, and turning it up made the text wordier but did not clearly change its subject.
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.
-
environmental factors. Doctors have known about hoarding for centuries, although the disorder was only formally recognized by the psychiatric community as a distinct psychiatric illness in 2013. Perhaps the most famous person who had a hoarding disorder was Howard
-
BLS report released in 2011. So women earned 81 cents of every dollar earned by men, which has been typical since 2004.⏎What’s the difference? Unlike the measure of annual wages by the
-
the United States.⏎Medical Record Technician Career History⏎Medical practitioners have been recording information about their patients’ illnesses and treatments for hundreds of years. Before the 20th and 21st
-
prices and marginal speculation in food have done the same.⏎2011 also saw the highest average oil prices since 2008, at $94.81 per barrel. That is not
-
downloaded for free at www.archenergy.com/SPOT/index.html.<|endoftext|>Since the Supreme Court recognized an individual’s right to keep and bear arms in 2008 — but declined to say which types of arms — several lower courts have
-
president of the United States.⏎Minus the federal benefits, licenses for same-sex marriage were introduced. Same-sex marriage was officially legalized within the United States in 2015.⏎Serving in a presidency of exactly sixteen years, she also remains
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: the97697777777777777777777777777777777777777777777
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
booth has become something of a novelty. But grownups can’t be bothered.⏎The first public coin-operated pay phone appeared in Hartford, Conn., in 1889. The first phone booth debuted in the early 190
-
Inmarsat is a British company that provides telephony and data services to users worldwide. The company has developed a series of networks since it was founded in 1979. The withdrawal of RBGAN occurred because it was superseded by
-
the crystals were pure protein, and assembled by autocatalysis⏎) were incorrect. The first electron microscopical images of TMV were made in 1939 by Gustav Kausche, Edgar Pfankuch and Hel
-
has been used for the elimination of diseases in certain plants.⏎Tissue Culture History⏎The first person to attempt cell culture was G. Haberlandt in 1902. He was unsuccessful, but hypothesized that hormones were necessary for
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
- L21:1815 strongly excites
- L18:2542 excites
- L20:1396 excites
- L21:2497 suppresses
- L20:1520 excites
- L18:2103 suppresses
- L20:2625 excites
- L6:248 excites
- L20:642 weakly suppresses
- L6:13 weakly excites
- L14:1700 weakly excites
- L20:1355 weakly suppresses
- L8:457 weakly excites year digits in copyright and dates
- L11:2530 weakly excites
- L20:1020 weakly suppresses
- L21:741 weakly suppresses
- L22:332 weakly suppresses
- L19:2444 weakly excites
- L21:2903 weakly suppresses
- L17:2799 weakly suppresses
- L16:2242 weakly suppresses
- L0:121 weakly excites
- L6:932 weakly excites
- L6:2202 weakly suppresses
Feeds later layers
- L26:905 strongly suppresses
- L27:494 strongly excites
- L25:147 suppresses
- L27:893 suppresses
- L25:322 suppresses
- L27:205 excites
- L27:1087 suppresses
- L25:333 suppresses
- L26:976 suppresses the '1' starting a century
- L25:2314 suppresses the 1 starting a 1900s year
- L27:294 suppresses
- L27:1100 suppresses
- L25:2383 suppresses
- L27:785 excites
- L25:342 excites
- L25:58 weakly excites colon in clock times
- L27:132 weakly excites
- L24:453 weakly excites
- L27:1389 weakly suppresses
- L25:2029 weakly suppresses
- L26:880 weakly suppresses
- L26:898 weakly suppresses
- L24:1199 weakly suppresses
- L25:1087 weakly suppresses word right before a number value
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
- ADER
- amus
- 部长
- 占有
- 鳍
- 热水器
- ivalence
- dating
- 陵
Suppresses
- 9
- 九
- ninety
- 9
- nine
- Nine
- -nine
- 8
- 九十
- ninth