unclear: document ends and answer colons
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. 16 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 is hard to pin down. It fires most strongly at the invisible marker that separates one web page from the next, and it switches off hardest on the colon in 'Answer:'. Turning it up made the model ramble and repeat phrases, which does not show a clear theme.
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.
-
code<|endoftext|>Jet engines, or gas turbines, suck in air at the front with a fan
-
wrong<|endoftext|>External Web sites⏎Britannica Web sites⏎Articles from Britannica encyc
-
groups<|endoftext|>What happens when you mix vinegar and baking powder?⏎When baking soda is mixed with
-
officers<|endoftext|>World Science Day for Peace and Development 2019⏎Open Science,
-
discarding of<|endoftext|>from The American Heritage® Dictionary of the English Language, 4th Edition⏎
-
9<|endoftext|>Virtual Kid Quest Guide for an Outdoor Art Quest⏎In celebration of Earth Day,
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: In the case of the long term, the best of the world for the world's future to "be well with us under a name of a name" (as we often do) and we "to be better for us under a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
70%.⏎Answer: B. Humidity levels should be kept below 60 percent, ideally between
-
are parallel, perpendicular or neither?⏎Parallel lines are the lines that never intersect and perpendicular lines are the lines that intersect at 90 degrees.⏎Answer: Lines with the same slope are parallel and if the slope of one line is the
-
done.⏎Answer: True. There are hundreds of thousands of dust mites in our homes, from
-
i.e., "I don't want to see it again").⏎In #2 "saw" is past tense and neutral about the present. (see The<|endoftext|>In the field of meta-ethics, scholars have identified key problems in the
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
- L9:1130 strongly suppresses
- L10:1713 suppresses
- L5:2946 excites
- L3:163 excites
- L5:334 excites
- L9:1757 suppresses
- L6:1228 weakly excites
- L4:1521 weakly excites
- L9:511 weakly excites
- L10:330 weakly excites
- L11:2445 weakly suppresses
- L3:609 weakly excites
- L0:683 weakly excites
- L6:2746 weakly suppresses
- L11:1214 weakly suppresses
- L6:840 weakly suppresses
- L2:1869 weakly excites
- L11:723 weakly suppresses
- L3:1898 weakly excites
- L3:2984 weakly suppresses unclear: mid-word subword pieces
- L6:228 weakly suppresses comma after a filler word
- L3:2704 weakly suppresses
- L4:1396 weakly excites "make" or "do" before an object noun
- L4:2015 weakly excites
Feeds later layers
- L18:461 suppresses
- L17:1340 excites
- L21:282 suppresses
- L17:816 suppresses
- L23:200 suppresses
- L17:244 excites
- L18:1382 weakly suppresses
- L21:1042 weakly excites
- L17:1680 weakly suppresses
- L18:191 weakly excites
- L21:1856 weakly suppresses
- L18:581 weakly suppresses
- L18:589 weakly excites
- L25:928 weakly excites
- L21:1766 weakly suppresses 'in' in 'bear in mind' phrases
- L19:327 weakly suppresses
- L18:1512 weakly excites end-of-document boundary token
- L18:2028 weakly suppresses
- L18:632 weakly excites
- L25:113 weakly suppresses
- L19:1000 weakly suppresses
- L27:2551 weakly excites
- L20:1072 weakly suppresses
- L18:2566 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
- anship
- ,...⏎
- ,…
- ',//
- vows
- div
- iration
- inspirational
- ,...⏎⏎
- ','');⏎
Suppresses
- thems
- ptime
- cé
- PJ
- ij
- Buenos
- ("\\
- ["
- кри
- oma