the letter y starting a word
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. 26 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 a lone "y" that starts a word, such as "yawn", "yam" or "yells". Turned down, the model repeated the word "tinker" and similar filler, and turned up it wandered off-topic.
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.
-
medical history and do a physical examination.⏎Medical history questions may include:⏎- When did the excessive yawning begin?⏎- How many times do you yawn per hour or day?⏎- Is it worse in the morning, after lunch
-
aur, taught Achilles how to use this herb.⏎Whether or not that is true, the world will never know, but we do know that healers placed rolled yarrow leaves into the nose of those who were bleeding throughout medieval times. Achilles might
-
garden⏎In the first volume of “Deer in My Garden”, beloved local garden guru Carolyn Singer wrote, “’Grandma, why don’t you grow yucky flowers?’ ‘Why would I do that, Marcus?’ I asked, respectful
-
shipwrecks from the Byzantine Empire, eight of which are now described in a new report.⏎The shipwrecks were discovered at a site called Yenikapi, in Istanbul, in what was a port of the ancient city
-
. There, too, the government, along with billions from Europe and the US, has tried to give growers an incentive to change crops. But alternative crops like yuca or rice sell for much less and are harder to get to market than
-
provides. The highly nutritious oca root is also grown commecially in New Zealand where it has become a popular root vegetable and where it is commonly known as yam.⏎In North America and Europe, oca root is relatively unknown and not
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: what about t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t t
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. So closely entwined are their lives that a number of evolutionary biologists argue that dogs and humans have evolved together.⏎Over the past decade, a loose-knit band of scientists in laboratories in the United States and abroad have worked diligently
-
the first park set aside to protect man-made wonders?⏎Mesa Verde National Park, Colo.⏎What is the largest U.S. National Park?⏎Wrangell-St. Elias National Park and Preserve, Alaska (13.2 million acres)⏎
-
Gothic, Perpendicular Gothic choir, religious fervour, Roman city of Eboracum, Roman ruins, school visits, secret sign, skilled carvers, St Bosa of York, St William of York, stained glass, stonem
-
to retrieve the correct value from the data table above.⏎Have you ever built an app based on Excel?⏎Method 1: The easiest way (VLOOKUP & MATCH)⏎Probably the easiest – or at least the shortest – way of returning values
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
- L18:405 excites
- L20:612 excites
- L21:442 suppresses
- L5:779 suppresses
- L16:143 suppresses
- L9:449 excites
- L6:1826 excites
- L6:328 weakly suppresses
- L17:278 weakly excites 'as' in 'as well as' and 'as much as'
- L1:542 weakly suppresses
- L23:2833 weakly excites
- L18:249 weakly excites
- L16:17 weakly excites
- L22:2314 weakly excites
- L2:1795 weakly excites
- L20:174 weakly excites
- L2:1699 weakly excites
- L18:423 weakly excites
- L20:1120 weakly excites
- L18:131 weakly excites
- L18:168 weakly excites
- L16:7 weakly excites
- L20:97 weakly excites
- L17:378 weakly suppresses
Feeds later layers
- L27:654 weakly suppresses
- L27:1118 weakly excites
- L27:2583 weakly excites
- L27:2412 weakly excites
- L27:2799 weakly suppresses the letter d inside words
- L27:1436 weakly suppresses
- L27:1028 weakly excites
- L27:893 weakly suppresses
- L27:1920 weakly excites
- L27:2285 weakly suppresses
- L27:1594 weakly suppresses
- L27:1800 weakly excites
- L27:2932 weakly suppresses capital F starting a word
- L27:534 weakly suppresses
- L27:558 weakly suppresses
- L27:1771 weakly excites
- L27:1990 weakly excites
- L27:204 weakly excites
- L27:1042 weakly suppresses
- L27:679 weakly suppresses
- L27:1867 weakly excites
- L27:1866 weakly excites
- L27:1405 weakly suppresses
- L27:577 weakly excites
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
- Replies
- 議
- emi
- òng
- 泳
- メント
- letics
- Foot
- 一角
- ゃ
Suppresses
- modx
- keycode
- velte
- Sakura
- .INTER
- tal
- krat
- rup
- ibil
- VX