first letter m or oct fragments
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. 25 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 on the first piece of a longer word that has been split into pieces, most often the letters 'm' or 'oct' starting words like 'mussels', 'mackerel' or 'octinoxate'. Turning it up made the model output long runs of blank lines. The opposite side is about words like 'over' and 'design' with no shared pattern.
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.
-
hu”. Since all the devotees who performed pooja on this auspicious day will achieve Moksha (Eternal Bliss), this day is also called “Mokshothasava Day”. Since all thirty three crore Gods are relived
-
and refuges (coral reefs) and provide food for fish (krill, pteropods, shellfish larvae) and humans (oysters, mussels, clams). Recent research reveals just how threatened such species are by acid
-
and ultimately make their way into the oceans. This is how the beauty industry affects our ocean ecosystem threatening marine life. Chemicals, especially oxybenzone, and octinoxate, found in sunscreen contribute to coral bleaching.⏎The Emergence of
-
asi Tidhis in a year Suklapaksha Ekadasi is termed as “Vaikuntha Ekadasi”. It is also called as “Mukkoti Ekadasi”. Though Dhanurmasa occurs in Sour
-
launching its new laptops known as Chromebooks and apparently, these do not contain any installed software.<|endoftext|>Salmon and other fatty fish such as sardines and mackerel are the superstars of heart-healthy foods. That’s because they
-
he was still a babe in his cradle and he strangled the two serpents that were meant to kill him, as the morning sun dispersed the creeping mists.⏎If Zeus-Order had ruled the world, the right of the first
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: how to make a healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy healthy
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
outhern India frequently accessorize their traditional garment with a spare⏎scrap of cloth, called an angavastram. The garment is designed to be worn over⏎the shoulder. This garment will remain on the shoulders for the time being.⏎
-
⏎the results of the first stage. As of September 2008 the design teams⏎use 3D modeling software to generate CAD drawings from initial design⏎sketches. The designs are then prototyped using an in-house
-
vioglacial waters. Up until 1860 amber procuring methods were off⏎the beaches and with shallow diving. It was obvious that sea amber came⏎from strata beneath the sea and a larger supply could be obtained by⏎
-
made certain recommendations on the basis of which the Chief⏎Ministers in their Conference held in July, 1972, laid down the ‘National Guide lines’⏎that were to govern the ceiling<|endoftext|>Daddy Long Legs and The Inch
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
- L22:377 strongly excites
- L20:268 strongly excites
- L20:1277 strongly excites
- L22:165 strongly suppresses
- L18:528 strongly excites
- L19:1090 strongly suppresses
- L19:2773 strongly suppresses
- L22:751 strongly suppresses
- L9:808 strongly excites
- L7:74 strongly excites unclear: scattered mid-sentence tokens
- L8:380 strongly excites
- L16:442 strongly excites
- L21:579 strongly suppresses
- L20:515 strongly excites period right before a line break
- L15:643 strongly suppresses
- L18:702 strongly suppresses
- L17:696 strongly excites
- L19:973 strongly excites
- L16:3055 strongly suppresses
- L24:1619 strongly excites
- L21:1564 strongly excites
- L24:244 suppresses
- L14:568 suppresses
- L24:1434 excites sentence end before a line break
Feeds later layers
- L26:2953 excites
- L27:1017 excites
- L26:2873 excites
- L27:1536 excites
- L26:230 excites
- L26:757 excites
- L27:2557 excites
- L27:1681 excites
- L27:598 excites
- L27:116 excites
- L26:740 weakly excites
- L27:207 weakly suppresses
- L27:97 weakly suppresses
- L27:487 weakly excites
- L26:993 weakly excites
- L26:460 weakly excites
- L27:1801 weakly excites
- L26:758 weakly excites
- L26:167 weakly excites
- L26:2346 weakly suppresses
- L26:449 weakly suppresses
- L26:787 weakly suppresses
- L26:749 weakly excites
- L26:1018 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
- ⏎
- ⏎
- ⏎⏎
- ⏎
- " ⏎
- ' ⏎
- - ⏎
- ⏎ ⏎
- "⏎
- '⏎
Suppresses
- ниц
- .firebaseapp
- .BorderColor
- ño
- ymce
- roid
- ⍣
- lander
- ophile
- rement