pieces of words split into 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. 2 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 small piece in the middle of a longer word that the model has split into parts, such as 'Jant' in 'Jantzen' or 'at' in 'insatiable'. Mostly it shows up in names and uncommon words. Turning it up did not produce a clear theme in the written text.
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.
-
in gene activity among individuals can give powerful clues for diagnosis, prognosis and intervention of different diseases. Senior author Emmanouil Dermitzakis, Louis Jeantet Professor at UNIGE, who led the study, emphasizes that today’s
-
30,000 guilders. During this period David Penner was a minister of the church of Bagdannen, whereas Salomon Jantzen seems to have been a deacon. The members of this Lithuanian
-
. The Art of Slow Reading. Portsmouth, NH: Heinamann, 2012.⏎And advocate to taking one’s time to
-
23, 2011⏎When two galaxies merge to form a giant, the central supermassive black hole in the new galaxy develops an insatiable appetite. However, this ferocious appetite is unsustainable.⏎The growth of super
-
of 1934, the Franks r…⏎E-pasta adrese, uz kuru nosūtīt darba saiti:⏎Saite uz darbu:<|endoftext|>Branch of biology that groups and names organisms based on studies
-
dharma of another. Nothing is ever lost in following one’s own dharma. But competition in another’s dharma breeds fear and insecurity.”⏎The word janma translates as “birth”. Ashtami (ush-tum-me) is the
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: I have a number of records. I have a number of records. I have a number of records. I have a number of records. I have a number of records. I have a number of records. I have a number of records
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, pickled pigs feet, lutfisk or dried codfish and variety of sweets.⏎A popular Christmas tradition is to
-
unkel⏎starring Jim Witter,⏎featuring Ian Tanner⏎“Time it was, and what a time it
-
land.⏎The cut comes into effect on December 19 and will apply across England and Wales.⏎According to Defra over 70% of the nitrates in Britain’s waterways can be traced back to agriculture. They cause twin
-
water contaminants discovered by water testing labs.⏎Nitrates are routinely discovered in tainted water supplies because they’re so prevalent in the environment. The primary source of nitrates in the
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
- L0:2257 suppresses
- L0:1301 weakly suppresses
- L0:2714 weakly suppresses
- L1:21 weakly excites
- L0:178 weakly suppresses
- L0:1224 weakly suppresses
- L1:2340 weakly suppresses
- L1:2469 weakly suppresses
- L1:2801 weakly excites
- L1:478 weakly excites
- L1:951 weakly excites
- L1:2505 weakly suppresses
- L1:874 weakly excites
- L1:1595 weakly suppresses
- L1:990 weakly suppresses
- L0:286 weakly excites
- L0:659 weakly suppresses
- L1:2683 weakly excites
- L0:2209 weakly excites
- L1:87 weakly suppresses
- L0:101 weakly suppresses
- L0:2919 weakly suppresses
- L1:2433 weakly suppresses
- L0:814 weakly suppresses
Feeds later layers
- L8:2323 weakly excites
- L4:230 weakly suppresses
- L10:2594 weakly excites
- L16:814 weakly suppresses
- L12:2283 weakly excites
- L17:81 weakly suppresses
- L24:2468 weakly suppresses
- L14:3021 weakly excites
- L15:1869 weakly excites
- L19:2525 weakly excites
- L12:467 weakly suppresses
- L14:605 weakly excites
- L6:1899 weakly excites
- L21:858 weakly excites
- L25:2353 weakly suppresses
- L20:321 weakly excites
- L24:2841 weakly excites
- L17:2965 weakly excites
- L19:673 weakly suppresses
- L9:1821 weakly suppresses
- L14:275 weakly excites
- L3:686 weakly excites
- L26:869 weakly excites word before him or he
- L13:921 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
- erg
- onium
- one
- erto
- rom
- erge
- 级
- 一口气
- eu
- rise
Suppresses
- stash
- arness
- xffffffff
- ��
- Utah
- AFX
- .LookAndFeel
- chatt
- .hashCode
- ätz