unclear: fragments inside longer 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. 23 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 is hard to pin down. It fires on pieces of longer words or names, such as the middle of 'Polk', 'Polaris' or 'volcano', and on a few Latvian words. Turning it up or down gave repetitive or unrelated text, so it doesn't show a clear job.
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.
-
100 it gets 1st grade and is very good.⏎If the sheep gets<|endoftext|>|Republic of Latvia⏎|Anthem: Dievs, svētī Latviju!⏎God Bless Latvia!⏎and largest city⏎|Eth
-
landers, whose cloth-making industry relied on England for wool. Early in the 14th century, Flemish artisans rose up in a series of bloody revolts against the aristocratic cloth dealers who had long monopolized power. The Count
-
, the other three districts have swung between control of both parties in recent years.⏎And at the legislative level, the focus will be on suburban districts like those in Polk County, where Democrats hope booming population will lead to more favorable districts. Population
-
drained by Little Antietam Creek. The major portions of the caves are developed along strike joints, with bedding exerting an influence on passage shape and speleothem occurrence. The general dip is 25º to the east away from
-
with metal alloys for aircraft. He had the results of seven years' labour in a portfolio with him and rather than let the research fall into German hands, (Poland having been invaded by Germany), he has hidden it.<|endoftext|>BURKE TE
-
confer, like the ability to develop uncanny expertise in an area of interest. This year's class includes specialists on supervolcanoes and medieval weaponry.⏎''Look at Jack,'' Justin pointed out. ''
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: [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] [] []
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
products. In the spring of 1934, the Franks r…⏎E-pasta adrese, uz kuru nosūtīt darba saiti:⏎Saite uz darbu:<|endoftext|>Branch of biology that
-
Old Ind. grīva- “neck,” grīvā- “the back part of the neck
-
it gets 1st grade and is very good.⏎If the sheep gets<|endoftext|>|Republic of Latvia⏎|Anthem: Dievs, svētī Latviju!⏎God Bless Latvia!⏎and largest city⏎|Ethnic groups (
-
(70 million). Other Major ones include Wu (90 Million) Min (70 million) in total 13 “sub languages” (ISO 639-3) While the Chinese character would be in
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
- L21:1138 strongly excites
- L19:1404 suppresses
- L21:1519 excites
- L10:1921 weakly suppresses
- L4:1261 weakly suppresses
- L22:1105 weakly excites
- L0:989 weakly suppresses
- L19:1944 weakly excites
- L12:2740 weakly excites
- L1:382 weakly excites
- L11:1594 weakly excites
- L9:1990 weakly excites
- L1:2536 weakly suppresses
- L8:1174 weakly excites
- L20:1798 weakly suppresses
- L12:101 weakly excites
- L15:524 weakly excites
- L20:2166 weakly excites
- L8:7 weakly excites
- L1:136 weakly excites
- L1:3061 weakly excites
- L2:24 weakly suppresses
- L21:792 weakly excites
- L10:1083 weakly excites
Feeds later layers
- L25:427 weakly suppresses
- L26:408 weakly suppresses
- L24:534 weakly suppresses
- L27:715 weakly suppresses
- L25:2415 weakly excites
- L24:1995 weakly suppresses
- L24:1695 weakly excites
- L26:1944 weakly excites
- L25:962 weakly excites
- L26:1982 weakly excites
- L25:1029 weakly excites
- L26:1082 weakly excites
- L26:1288 weakly suppresses
- L25:1742 weakly suppresses
- L24:1279 weakly excites
- L24:44 weakly excites
- L24:1931 weakly excites
- L24:466 weakly suppresses
- L25:2423 weakly excites
- L26:2676 weakly suppresses
- L27:1253 weakly excites
- L26:112 weakly suppresses
- L25:475 weakly suppresses
- L24:667 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
- longing
- lycer
- ăng
- hare
- pherd
- agus
- lington
- 娘家
- veillance
- мес
Suppresses
- []
- []
- ([]
- [][]
- ([]
- [][]
- []{
- []⏎
- [] ⏎
- icolor