word fragments before a name, 'ma' bias
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. 27 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 opening pieces of words split across tokens, often names or nouns, such as 'bul' in bulge or 'Sel' in Selma. Turned up strongly, the model began writing 'Ma Ma Ma', so it seems to push toward the sound 'ma'.
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.
-
Japanese scifi movie...Scientists have spotted stellar fossils in the center of the Milky Way, globular clusters orbiting in the central bulge which seem to have come from somewhere else. Like all fossils these are vital
-
Hungarian king Kalman, Jews were permitted to live only in cathedral towns and on bishops estates. In 1241 the Mongols (also known as Tartars) invaded Hungary, wreaking havoc and destruction. Jewish merchants made a major
-
content was digitised by double rekeying. All rights reserved.⏎The Plague of 1348—The Origin of the Charterhouse—Sir Thomas More there—Cromwell's Commissioners—Prior Houghton—The Depart
-
lle, o méle, lo méle, u mméla “honey”⏎lo latte, o latte, lə lattə “milk”⏎ru cane “dog”⏎ru peʃʃu “singular fish”⏎
-
case studies, demographics<|endoftext|>Tom Houck is a lifelong U.S. civil rights activist. In 1965, the white high school student traveled to Selma, Alabama, to participate in the march there, led by the Rev.
-
May morning to view the medows gay,⏎I met a pretty dairy-maid all in the new mown hay.⏎* 4 *⏎As I was
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: Can an algorithm be pointed out so that an algorithm can be pointed out again? 问题是问题是这个问题的问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题问题
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
which at the time of the Three States (Samguk 三國) was occupying the South-Western part of the Korean peninsula 3. The state of Paekche, because of the continuous pression exerted upon it by K
-
, “being Buddha nature and non-being Buddha nature.” I like the use of “being” in this reference. In the first sentence when referencing all tathagatas, he is putting forth the meaning of “non-being Buddha nature”. In
-
an cheese and roast in the oven for a crispy and flavorful side dish.⏎- Broccoli and mushroom omelette: Saute broccoli florets and sliced mushrooms and fold them into a fluffy egg omelette for a filling and nutritious breakfast.
-
Nature Beyond Conditions and a commentary by Rev. David Xi-Ken Astor, Sensei⏎BUDDHA NATURE BEYOND CONDITIONS⏎“All tathagatas are without Buddha nature, but at the same time, previously they have fully
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
- L26:789 excites
- L8:10 excites
- L26:1564 excites
- L25:2769 suppresses
- L16:17 weakly suppresses
- L22:859 weakly suppresses
- L3:2084 weakly excites
- L7:57 weakly excites
- L3:2630 weakly suppresses
- L23:2194 weakly excites
- L19:546 weakly excites
- L2:2142 weakly suppresses
- L22:1902 weakly excites
- L17:1140 weakly suppresses
- L8:7 weakly excites
- L25:1021 weakly suppresses
- L26:1624 weakly excites
- L15:984 weakly suppresses
- L8:13 weakly suppresses
- L0:2459 weakly suppresses
- L8:14 weakly excites
- L19:911 weakly excites
- L4:205 weakly excites the digit 2 inside citations
- L7:64 weakly suppresses
Feeds later layers
None: this is the last layer.
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
- Ma
- ma
- Ma
- ma
- MA
- MA
- sa
- (pa
- Pa
- pa
Suppresses
- 枝
- etcode
- onian
- olvers
- illator
- LICENSE
- ibil
- 问题是
- .PerformLayout
- IBLE