split word pieces in names and years
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. 9 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 middle pieces of longer names and words that get chopped into fragments, like the 'ople' in Constantinople or the 'ka' in Topeka. It also reacts to digits inside years. Turning it up did not produce a clear theme in the writing.
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.
-
were rivalry between the two brothers, and El Cid had supported Sancho. But Sancho died in 1072,
-
a separate peace with Germany and declared itself the enemy of all bourgeois states, France and Britain prominent among them.⏎What made you want to look up "Constantinople Agreement"? Please share what surprised you most...<|endoftext|>Your healthcare provider will ask about
-
Aid Company; School buildings; Schools; Shawnee County, Kansas Territory; Topeka, Kansas Territory⏎Letter, E. [Ephraim] Nute
-
President Boris Yeltsin. Like the Americans, the Russians use a mandatory 10 minute launch window (the time needed for a submarine-la<|endoftext|>Constantinople AgreementArticle Free Pass⏎Constantinople Agreement, (March 18,
-
with the example of the Blessed Virgin, St. Joseph, the Shepherds, and, who could forget the Adoration of the Magi. The popular devotion as we know it has been particularly strong in Spain since the
-
rix of the parabola. [Adopted from iCoachMath]⏎Focus:Focus of a parabola lies on the axis of symmetry.⏎Directrix:Directrix is a line that is perpendicular to the axis of symmetry of a
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 been told that there are 1000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
id. Source: Twitter⏎1929 motor coach no.1 at Palma station. Source: Twitter⏎Sóller station in 2019, with the Puerto tramlines on the left. Source: Twitter⏎The
-
general and first president, respectively.⏎These two 1847 stamps hold their value through market fluctuations and through thick times and thin.⏎The 2020 Scott Specialized Catalogue of United States
-
in equivalent purchasing power in 2010). So the picture has never been quite so pretty as the image that is so often laid out for us when we are told of the halcyon past in which
-
until its demolition in the mid-1960s. The oval dining room was removed from the mansion and recreated in the Baltimore Museum of Art where it remains a part of the American Decorative Arts wing.⏎This demolition sparked a renewed awareness
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
- L8:764 excites
- L7:448 excites
- L7:305 excites
- L8:418 excites
- L8:474 excites
- L6:2881 suppresses
- L8:800 weakly excites
- L6:113 weakly suppresses
- L6:1717 weakly suppresses
- L8:2096 weakly excites
- L6:1667 weakly suppresses
- L5:276 weakly suppresses
- L2:2692 weakly suppresses
- L8:2719 weakly suppresses
- L8:3069 weakly excites
- L7:122 weakly suppresses
- L5:158 weakly excites
- L8:1432 weakly excites
- L3:139 weakly excites
- L6:2356 weakly excites
- L3:238 weakly suppresses
- L3:596 weakly excites
- L1:1464 weakly suppresses
- L7:113 weakly suppresses
Feeds later layers
- L10:375 suppresses
- L14:403 suppresses
- L19:1026 suppresses
- L13:1926 excites
- L10:2351 excites science news body text
- L13:2480 suppresses
- L11:464 suppresses
- L10:3044 excites
- L10:429 suppresses
- L13:2491 weakly excites
- L12:1970 weakly suppresses
- L12:1713 weakly suppresses
- L12:61 weakly suppresses
- L16:546 weakly suppresses
- L20:2388 weakly suppresses
- L11:2896 weakly excites
- L23:1047 weakly excites
- L10:1517 weakly suppresses
- L10:2196 weakly suppresses
- L11:2736 weakly excites
- L10:716 weakly excites
- L21:2819 weakly excites
- L11:1636 weakly suppresses
- L17:254 weakly suppresses unclear: word before a twist or contrast
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
- ewidth
- еств
- BaseActivity
- leness
- (show
- 玒
- NCY
- vens
- красот
- zach
Suppresses
- unks
- -⏎⏎
- 升
- Kn
- ...⏎⏎
- reet
- 割
- kn
- -⏎⏎
- 吸