periods in abbreviations and the 's in 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. 6 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 punctuation inside names and abbreviations, such as the dot in "H.L." or "S.J." and the "'s" in "Parkinson's". Turning it up gave no clear change in what the model wrote.
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.
-
over the<|endoftext|>For more information, please visit our Functional Neurosurgery Program site or Center for Neuromodulation⏎What is Parkinson's disease?⏎Parkinson's disease (PD or, simply, Parkinson's) is the most common form of
-
and cons of gene technology for agriculture.⏎The Science Pub presentation begins at 6 p.m. in the Majestic Theater, at 115 S.W. Second St. in Corvallis. It is free and open to the public
-
.⏎McLaren, P., & Farahmandpur, R. (2001). Class, Cultism, and Multiculturalism.
-
than a⏎cabbage, concludes that it will also make better soup." - H.L. Mencken<|endoftext|>Ice shelf the size of Connecticut at risk of collapsing⏎- Grades
-
L., Allison, A. & James, R. 2016. Dorcopsis atrata. The IUCN
-
of the astronomy observatory in Vienna, even after the Suppression of the Jesuits. A lunar crater is named after him. Maximilian Hell, S.J. died 2
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 been thinking about the role of the 2016 United States v. United States v. United States v. United States v. United States v. United States v. United States v. United States v. United States
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ystone species in Florida’s xeric (dry) communities, meaning that its presence and behavior affects the survival of many other species.
-
was described as a shootout involving drug gangs.<|endoftext|>Osteogenesis imperfecta (OI), also known as brittle-bone disease, is a genetic (inherited) disorder characterized by bones that break easily without a specific cause. An estimated 2
-
look at the size and location of the thyroid or to find out if it is missing altogether.⏎Every day, you will give a synthetic (laboratory-made) form of thyroid hormone to your baby to replace the missing thyroid hormone. You’ll
-
the United States is different from the laetrile/amygdalin produced in Mexico. The patented Laetrile is a partly synthetic (man-made) form of amygdalin, while the laetrile/amygdalin made
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
- L4:212 excites
- L4:2436 weakly excites
- L4:951 weakly suppresses
- L2:1776 weakly suppresses
- L1:1430 weakly excites
- L3:1722 weakly suppresses
- L4:217 weakly suppresses
- L5:1360 weakly suppresses
- L3:699 weakly suppresses
- L5:1747 weakly suppresses
- L1:1987 weakly excites
- L5:2631 weakly suppresses
- L5:1372 weakly excites
- L5:73 weakly excites
- L3:187 weakly excites
- L0:1882 weakly excites unclear: 'high tech' and mid-word pieces
- L5:2840 weakly excites period after abbreviation in citations
- L4:1019 weakly excites
- L1:1768 weakly suppresses
- L1:1373 weakly suppresses
- L0:1667 weakly excites
- L0:421 weakly excites
- L4:68 weakly excites
- L3:1839 weakly excites
Feeds later layers
- L9:392 excites "and" inside organisation names
- L15:1554 excites
- L12:792 excites
- L9:1515 excites line break splitting a phrase
- L15:2449 weakly excites line breaks mid-sentence
- L7:2164 weakly excites period inside country abbreviations
- L8:333 weakly excites
- L17:146 weakly excites period after abbreviations and dates
- L15:316 weakly suppresses
- L11:810 weakly suppresses
- L7:712 weakly excites
- L8:1801 weakly excites
- L21:348 weakly excites mid-sentence line breaks
- L17:546 weakly excites
- L18:192 weakly suppresses
- L8:517 weakly excites
- L17:727 weakly suppresses
- L20:1248 weakly excites
- L14:493 weakly excites
- L16:1153 weakly excites
- L7:266 weakly excites
- L16:2137 weakly suppresses
- L9:119 weakly excites
- L15:1953 weakly suppresses unclear: document ends and number gaps
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
- ]:⏎⏎
- ]=[
- ]%
- 二字
- ]:⏎
- ]',
- include
- ']:⏎
- kbd
- ],
Suppresses
- ually
- uard
- ocratic
- emens
- 市政府
- eland
- pective
- nable
- xious
- ylum