'so' before much or many
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. 11 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 responds to the word 'so' when it means 'so much', 'so many' or 'so quickly', as in 'missed by so many news sources'. It is switched off, and even pushed the other way, when 'so' is followed by an adjective like 'so vast' or 'so covert'. Turning it up did not give a clear shared 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.
-
several common cosmetic products, the harmful ingredient, the purpose for incorporating the chemical, and the adverse effect. So why is this still going on today when we know so much more about the effects? Unfortunately cosmetics are a low priority because of they are
-
a tautology: US policy is democratic because America is free. One suspects that if there was more truth to the assertion it would not have to be repeated so frequently, but merely be understood.⏎But, let
-
CPU hours and producing a hundred terabytes of data. And yet, when you ponder the vast sweep of space and time being modeled, it seems a marvel that so much universe can be squeezed into such a small box<|endoftext|>Tupac Shakur
-
with more alert minds they should be more receptive to instruction.⏎The benefits are so obvious, in fact, it’s hard to understand how so many school officials were so easily motivated to do away with physical education classes in the first place.⏎After seeing
-
contrast, an `io soaring over the Panaewa farm lots bears no resemblance to another hawk. It must have arrived a long time ago to have changed so much. But how long ago was that?⏎The answers have come recently by a
-
. An expert at the Coastal Bend Bays and Estuaries Program (CBBEP) said things could have been much worse had the state agencies not responded so quickly.⏎Just north of North Beach, along Highway 181, is
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 is the value of the score of the current score in the game that I have built up to a point that it's almost 30% that I can't make up the time that I have made it up to now that I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
it sank with its vast cargo of precious stones, jewellery, gold, silver and tobacco still on board. In fact, the value of this ravaged fleet was so vast that the Spaniards spent a decade attempting to salvage the remains. They brought
-
- The Something is very different from the hunters and their dogs (the speaker knows all about them). The Something is far more covert. In fact, it’s so covert that no one sees or hears it make gaps in the wall.⏎- The
-
he leads his black soldiers to death,⏎he cannot bend his back.⏎- By the time the 54th infantry fights its last battle, the Colonel is so exhausted from fighting that he
-
one-quarter of the global average, and nearly 700 million of its citizens live in regions considered highly water-stressed. Meanwhile, groundwater depletion has been so significant in the areas around Beijing that parts of the city are falling into the earth
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
- L5:1364 excites
- L10:2106 suppresses
- L7:2914 suppresses
- L8:14 excites
- L5:2308 suppresses
- L4:656 excites
- L9:2328 suppresses
- L8:7 weakly excites
- L4:2688 weakly suppresses
- L7:1990 weakly suppresses
- L3:1953 weakly excites 'such' after 'no' or 'any'
- L1:2005 weakly suppresses
- L10:2935 weakly excites
- L7:2331 weakly suppresses
- L7:191 weakly excites
- L10:487 weakly excites
- L7:2216 weakly excites
- L7:460 weakly suppresses
- L2:2335 weakly suppresses
- L9:1423 weakly excites
- L8:1161 weakly suppresses
- L3:937 weakly suppresses
- L10:1477 weakly excites
- L4:3051 weakly excites
Feeds later layers
- L16:2782 strongly suppresses
- L17:2863 suppresses
- L15:2168 suppresses
- L16:1838 suppresses
- L14:1127 suppresses
- L20:1312 suppresses
- L14:2082 suppresses
- L16:2066 suppresses
- L15:906 suppresses
- L16:1802 suppresses
- L14:2132 excites
- L18:593 excites
- L13:958 suppresses
- L16:1645 suppresses
- L12:1668 weakly suppresses
- L12:631 weakly suppresses
- L13:706 weakly suppresses
- L17:626 weakly suppresses
- L13:2509 weakly suppresses
- L14:1960 weakly suppresses
- L21:881 weakly suppresses comma after 'if' clause in prose, braces in code
- L16:881 weakly suppresses
- L13:883 weakly suppresses
- L16:2975 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
- Tech
- ."),⏎
- 但这
- Tech
- ").⏎⏎
- Aid
- 但也
- 最大的
- ."));⏎
- .'),⏎
Suppresses
- {}{⏎
- ترا
- getParent
- inosaur
- _METADATA
- ={()=>
- iena
- .getProject
- <article
- ampa