unclear: the word 'kind' and mixed words
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. 22 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 a grab bag of words, most often 'kind' (as in 'a kind gesture' or 'kind of funny'), plus pieces of 'Creative' and a few others. Turning it up or down did not produce any clear change in what the model wrote.
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.
-
a part of it. Open and hold doors for people. Let someone else go before you in a grocery store line just because it is a nice gesture. Act kind in front
-
electrical resistance,⏎- I is the electric current,⏎- V is the voltage.⏎(Note: this is not true where the impedance is complex)⏎|This page uses Creative Commons Licensed content from Wikipedia (view authors).|<|endoftext|>Of the moon that is
-
part as an image that "the deep heart of nature" (perhaps God) may look more kind on the errant Hester and her child than her Puritan neighbors do.⏎
-
Success in Cirrcuit lies⏎Too bright for our infirm Delight⏎The Truth's superb surprise⏎As Lightening to the Children eased⏎With explanation kind⏎The Truth must dazzle gradually⏎Or every man be blind---⏎1.
-
impulsively and irrationally. Salekin uses a laptop computer and LCD player to explain brain plasticity
-
uba diver trained to work in underwater installations and conduct, assist in, or be a subject of scientific research. Also called oceanaut.⏎from Wiktionary, Creative Commons Attribution/Share-Alike License⏎- n. An underwater explorer.⏎from
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: there is a 24 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
protector deity Penden Lhamo⏎- The Goddess of the White Umbrella (the Buddhist deity Ushnisha-sitatapatra)⏎- The Great mystic Virupa⏎- Thunderbolt and bell⏎- Prayer wheel⏎- Cabinet
-
old<|endoftext|>Immune cells in the brain have surprising influence on sexual behavior⏎Relatively uncommon mast cells appear powerful in development, study finds⏎Researchers have found a surprising new explanation
-
actual glimpse of its mystic history. From ancient buildings to the modern wonders Delhi offers anything one can desire
-
circle. Man today is not wiser than the ancient seers and sages and thinkers, not more spiritual than the great seekers of old, the first mighty mystics, not superior in arts and crafts to the ancient artists and craftsmen;
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
- L19:2402 weakly suppresses
- L19:107 weakly excites
- L20:601 weakly suppresses
- L17:148 weakly excites
- L19:2924 weakly suppresses
- L20:570 weakly excites
- L21:237 weakly suppresses
- L18:420 weakly suppresses
- L6:374 weakly excites
- L9:1213 weakly excites
- L14:1187 weakly excites
- L20:2446 weakly suppresses
- L19:277 weakly suppresses
- L11:2297 weakly suppresses
- L20:380 weakly excites
- L5:2738 weakly suppresses
- L21:2189 weakly suppresses
- L3:2866 weakly suppresses
- L12:819 weakly excites
- L0:2370 weakly suppresses
- L13:971 weakly excites
- L16:2063 weakly suppresses
- L11:865 weakly excites
- L21:16 weakly suppresses
Feeds later layers
- L23:520 weakly suppresses
- L27:2044 weakly suppresses
- L23:2965 weakly suppresses
- L24:2833 weakly suppresses
- L23:1882 weakly suppresses
- L24:561 weakly suppresses
- L23:2034 weakly excites
- L25:2661 weakly suppresses
- L27:1178 weakly excites
- L27:2810 weakly suppresses
- L27:908 weakly suppresses
- L23:2070 weakly excites
- L24:2463 weakly suppresses
- L24:2050 weakly suppresses
- L24:800 weakly suppresses colon in clock times
- L23:1210 weakly excites
- L25:1556 weakly excites
- L23:1831 weakly suppresses
- L27:1316 weakly suppresses
- L25:618 weakly suppresses
- L26:649 weakly suppresses
- L26:1994 weakly suppresses
- L24:1251 weakly suppresses
- L25:787 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
- eri
- eria
- rists
- adam
- alty
- ng
- Side
- emy
- edy
- Side
Suppresses
- agnitude
- ounding
- iqué
- icine
- 做一个
- ulaire
- obox
- uild
- less
- iciel