end-of-text token after finished articles
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. 13 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 mostly fires at the point where one web page or article ends and a new unrelated one begins. It also fires on a few ordinary word breaks in news-style writing. Its negative side is less clear, showing up on words like 'their' or 'the' in the middle of a sentence.
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.
-
pain) and seizures in some instances<|endoftext|>Gauss Sum Walks⏎This Demonstration shows two types of pseudorandom
-
source of clean water, soap so they can stay clean, and a proper toilet to avoid contaminating water supplies. In many of Oxfam’s ongoing programs, our partners work on promoting good hygiene and sanitation to help people stay healthy even when
-
turtles from the wild and sell them.) Because of her size, she was rescued and sent to a Missouri farm, which is where the Newport Aquarium found her.<|endoftext|>The hustle and bustle of the back-to-school season is starting to wind down
-
York’s Metropolitan Museum of Art in 1983. The Designer died in Paris on June 1, 2008 from brain cancer.<|endoftext|>How do you get
-
the U.S. Congress to agree on a carbon tax or cap, his idea has serious merit and deserves a broad discussion, especially as Congress reassess its budget priorities. This includes billions of dollars of tax breaks and incentives for oil and other fossil
-
Unfortunately he died at a very young age, in 1938, which made it impossible for him to leave and extensive body of work behind. .<|endoftext|>|Siege of Jerusalem|⏎|Part of the First Jewish-Roman War|⏎
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 first thing in the question is the time, and the first thing is the time when it is, and the first thing is the time, and so, the first thing is the time, and the first is the time, and the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
better!)⏎Newspapers and magazines⏎Writing utensils (markers, pencils)⏎Glue or tape⏎- Gather and look at newspapers for layout inspiration. Consider their design, the placement of
-
and they had two children. Just a few years later, the young mother traveled to Rome to study under Maria Montessori. Impressed by Montessori’s revolutionary methods, Fisher worked to bring the Montessori Method to the United States
-
"?⏎Yes, curling has been compared to chess by fans and detractors alike. Fans emphasize the strategy involved. Detractors claim that both sports are similarly long and tedious for
-
such as a mobile or through your hotel.⏎- If you receive an email which appears to be from a family member or friend claiming to be in trouble look at the way the email is written, ask yourself whether the email sounds like it was written
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
- L12:3069 suppresses
- L10:1884 excites
- L11:1092 suppresses
- L12:2280 excites pipe before a table field label
- L9:941 excites
- L12:1794 excites
- L12:1913 excites
- L11:1532 suppresses
- L11:2765 excites
- L10:2293 suppresses
- L12:2414 suppresses
- L12:1824 weakly excites
- L6:623 weakly suppresses
- L12:2201 weakly suppresses
- L12:1518 weakly suppresses
- L12:1838 weakly excites
- L12:1121 weakly excites
- L10:331 weakly excites
- L12:1828 weakly suppresses
- L12:3032 weakly suppresses
- L1:171 weakly excites "but" after a negated clause
- L9:2721 weakly suppresses
- L10:2522 weakly excites
- L12:1432 weakly excites
Feeds later layers
- L14:252 suppresses
- L15:1369 suppresses
- L16:2823 suppresses
- L16:2988 suppresses
- L14:1664 suppresses
- L16:733 suppresses
- L15:209 suppresses
- L15:2265 suppresses
- L14:2813 suppresses
- L17:2565 suppresses
- L15:266 excites
- L16:3061 suppresses
- L17:2554 suppresses
- L15:1998 suppresses
- L18:2261 suppresses
- L16:2010 suppresses
- L14:2290 suppresses
- L18:2548 suppresses
- L17:2374 weakly suppresses
- L14:2356 weakly suppresses
- L18:2147 weakly suppresses
- L18:1358 weakly suppresses
- L16:3019 weakly suppresses
- L15:709 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
- itionally
- usk
- 得住
- alyzer
- ONGL
- enko
- 울
- otland
- createContext
- yn
Suppresses
- features
- feature
- features
- Latitude
- _feature
- eam
- Feature
- feature
- eras
- 封面