digits inside numbers
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. 20 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 digit that sits in the middle of a longer number, such as the 1 in 144 or the 0 in 2048, and on the slash or dash between numbers. Turning it up did not produce text about numbers, so its effect on writing is unclear.
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.
-
Permanent Magnet or Ferromagnetic Material. The Stepper Motors are manufactured with steps per revolution of 12, 24, 72, 144, 180, and 200, resulting in
-
pictures to your PC⏎maximum image size supported for upload is 2048x2048 pixels,⏎be sure to set camera to this or lower⏎- Hand
-
of the Zodiac and of the hours of the day (the Chinese measured the day in units equivalent to two Western hours, which when multiplied by twelve makes up a twenty-four hour day).⏎In an astonishing irony, in 1860 the
-
But how do you divide that by 3.2?⏎I’d recommend moving both of the dots to the right to get $2400 \div 32$. And you know your 32
-
that contains a description of the layout of the file.⏎|Program||Format||File Header||Block Header||Parameters|⏎|Varian||Serial||32/28||24||File Header and Param File|⏎|NMRView
-
ships," which means that they were propelled primarily or entirely by sail. The length of the round ships ranged from 26 to 48 feet (8 to 14.7 meters), and they were between 8 and 1
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: the problem of a student with a different home and in a different environment of the future in terms of a difference of language and environment. For example, in a project of a student in a different environment, the same student in the future in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is located 250 kilometers (155 miles) from Villahermosa, the capital of Tabasco.⏎
-
per year — for a potential savings of $625 a year (based on 12,000 miles of driving per year and $3.79 a gallon, nationwide average as of 5/19/2
-
gent plate margin between the African Plate and the Eurasian Plate. It is the tallest active volcano in Europe, currently standing 3,329 m (10,922 ft) high, though this varies with summit eruptions
-
Highest point||top of flag staff|⏎|- elevation||15.75 ft (4.8 m) over stone base|⏎July 3, 1929
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
- L19:2271 excites
- L19:2794 weakly suppresses
- L18:1902 weakly excites
- L16:2997 weakly excites
- L6:1406 weakly suppresses
- L8:295 weakly excites
- L16:1057 weakly suppresses
- L7:693 weakly excites
- L8:2062 weakly excites
- L13:56 weakly excites
- L15:1909 weakly suppresses blank slot before a number
- L19:251 weakly excites
- L19:1607 weakly suppresses
- L19:731 weakly excites
- L8:1890 weakly suppresses
- L7:134 weakly excites
- L17:1905 weakly suppresses
- L1:1372 weakly suppresses
- L10:220 weakly suppresses hyphen inside compound words
- L1:500 weakly suppresses
- L12:1924 weakly suppresses
- L1:2638 weakly suppresses
- L17:1100 weakly excites
- L15:12 weakly excites
Feeds later layers
- L22:1859 strongly excites
- L21:2396 strongly suppresses
- L21:2069 strongly excites
- L24:2534 excites
- L22:1099 excites
- L22:150 excites opening parenthesis before an acronym
- L22:1562 excites
- L22:1135 excites
- L22:1684 suppresses
- L23:569 excites
- L21:2801 weakly suppresses
- L22:2223 weakly suppresses
- L24:1898 weakly suppresses
- L23:2256 weakly suppresses
- L23:2964 weakly excites
- L21:2637 weakly suppresses
- L22:2986 weakly suppresses
- L22:1193 weakly suppresses
- L27:1605 weakly excites
- L27:450 weakly excites
- L23:1964 weakly excites
- L24:3063 weakly excites
- L23:2672 weakly excites
- L22:2584 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
- 作者所有
- �
- ypes
- ய
- acs
- 体系
- 的通知
- 蝽
- ーン
- 内容简介
Suppresses
- ハン
- unner
- rowave
- pais
- Gamer
- WATCH
- bose
- ULATE
- reate
- unic