end-of-text token after document breaks
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. 7 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 most strongly on the hidden 'end of document' marker that separates one web page from the next. When it was turned up, the text drifted into odd, unrelated topics with no clear theme.
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.
-
959,<|endoftext|>Everyone's heard about an apple a day. As one of the most popular fruits
-
criteria.<|endoftext|>The haunting image of the fallen soldier, aged just 18, is one
-
:<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Aim High' printed from
-
:<|endoftext|>The hamstring muscles – the biceps femoris, semitendinosus,
-
the support leg; <|endoftext|>MAHAD SATYAGRAHA AND BABASAHEB AMBEDK
-
, and other vehicles.<|endoftext|>Most people remember the adage that having a hot breakfast will set the tone for
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 function g() is used to represent the operation of taking the function $ f(x) $ and applying it to get the result $ f(x) $. So, the question is: when should I use the function $ f
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
behavior. Here's a class called CarParts that declares a static inner class called Wheel. If we consider the real-world, we can find many objects
-
re toothed tiger. This adrenaline rush causes the memory to be imprinted into the amygdala, which is part of the limbic system. The amygdala holds the emotional significance of the event, including the intensity of emotion. The
-
oding – you’ve got to manage the energy you bring into a room and have a plan. With a plan, your brain is moving out of the reactive amygdala
-
counsel is grounds for automatic reversal of the conviction. Other mistakes are judged by the Harmless Error Rule. The Confrontation Clause of the Sixth Amendment gives the defense the right to cross-examine witnesses who testify about lineups, Showups,
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
- L5:1234 excites
- L6:1803 excites
- L5:1650 weakly suppresses
- L3:2635 weakly suppresses
- L3:1912 weakly suppresses
- L6:8 weakly suppresses
- L3:2023 weakly suppresses
- L6:12 weakly excites
- L3:533 weakly excites
- L2:1657 weakly excites
- L6:1821 weakly suppresses
- L3:1101 weakly suppresses
- L3:551 weakly excites
- L5:2191 weakly excites
- L5:1574 weakly suppresses
- L1:997 weakly excites
- L5:102 weakly excites
- L0:1954 weakly excites
- L5:1353 weakly suppresses
- L0:2242 weakly excites
- L5:50 weakly excites
- L6:587 weakly suppresses
- L4:1882 weakly excites
- L6:1355 weakly suppresses
Feeds later layers
- L11:1159 excites
- L16:2172 weakly excites
- L9:2376 weakly suppresses
- L10:2352 weakly suppresses
- L9:1115 weakly excites
- L8:1853 weakly suppresses
- L18:2186 weakly suppresses
- L13:1295 weakly excites
- L13:786 weakly suppresses
- L8:1603 weakly suppresses
- L10:1140 weakly excites
- L10:2284 weakly excites
- L8:1096 weakly suppresses
- L12:2109 weakly suppresses
- L25:2292 weakly suppresses
- L9:1862 weakly excites
- L21:3058 weakly suppresses
- L11:2517 weakly suppresses
- L24:1800 weakly suppresses
- L13:2629 weakly excites
- L24:1022 weakly suppresses
- L9:2830 weakly suppresses
- L9:838 weakly excites
- L10:1321 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
- Preferred
- Preferred
- preferred
- Armenian
- preferred
- tlement
- prefer
- ayrıl
- orderly
- ardown
Suppresses
- اع
- ity
- ceptor
- 乎
- ceptors
- 革
- ibu
- icus
- نو
- لاق