comma after a full date
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 very consistent: they mostly 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 lights up on the comma in dates like 'Jan 09, 2017' or 'July 12, 2017', the spot right before the year. It goes quiet, or flips negative, on the middle digits of years written inside text, such as the '3' in 1380. Turning it up mostly made the model fall into loops of blank lines, so the steering results do not tell us much more.
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.
-
Jan 09, 2017 08:57 AM EST⏎MIT Professor
-
direction you would get more water pressure (voltage) because the pumps are helping each other<|endoftext|>Canine fear-related aggression (Proceedings)⏎Nov 01, 2009⏎CVC IN SAN DIEGO PROCEEDINGS⏎
-
pale little girl with a thin neck, and spindling legs half-hidden in the folds of Aunt Frances’s skirts. But she didn’t<|endoftext|>Dec 6, 2023⏎Airbus has successfully passed the Preliminary Design Review
-
York Times, Project Glass, technology, turn by turn directions, UI, User interface, video chat, Virtual reality, Virtual reality glasses⏎January 18, 2012 • 11:07 am
-
with something familiar for scale pic.twitter.com/OdQA07qhBb⏎— Adrian Luckman (@adrian_luckman) July 12, 2017⏎You may imagine that the breaking off of the iceberg
-
every bit of opinion you might start writing, think of including an “and…” statement.⏎— Danielle Degelman (@deedegs) November 26, 2014⏎One approach if giving criticism is to put it inside
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 family of a family with a big family, and a family with a small family. If you have a large family, there is a big family with a big family, and a small family with a large family. The pattern is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, is thought to be the name or epithet of a Kartid amir or military commander. Destroyed a second time by Timur’s army (1380), Qal’a-ye Ikhtiyar al-Din was
-
the pope's favor forced him to leave Rome and work for King Alfonso of Naples (1452-1455), but he resumed his position as papal
-
emerged as the leader of local Turks in the fight against the tottering Byzantine Empire. The final conquest of the Byzantines was not achieved until 1453 with the fall of Constantinople (now Istanbul), but by that date
-
Colov.⏎Later, the tower had a defensive character as well, especially in the time of the attack of Omer Pasha on Montenegro in 1862. At the end of the 19th century this building was
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
- L18:3021 excites
- L14:2049 weakly excites
- L20:151 weakly excites
- L17:1333 weakly excites
- L19:1272 weakly suppresses
- L19:1308 weakly suppresses
- L9:42 weakly suppresses
- L21:1042 weakly excites
- L7:1564 weakly excites website boilerplate and navigation text
- L12:1443 weakly excites opening parenthesis after a name or term
- L12:1145 weakly excites
- L10:80 weakly suppresses
- L2:1049 weakly excites
- L12:1140 weakly excites
- L14:3017 weakly excites
- L18:901 weakly excites
- L5:2915 weakly suppresses
- L12:668 weakly excites
- L9:2659 weakly suppresses
- L15:23 weakly suppresses
- L1:2161 weakly suppresses
- L4:460 weakly suppresses
- L5:1081 weakly suppresses
- L14:1811 weakly excites word before an evaluative adjective in essays
Feeds later layers
- L23:3020 excites
- L23:1093 excites
- L26:555 excites
- L27:1779 excites
- L23:233 suppresses
- L23:337 weakly excites
- L25:487 weakly suppresses
- L25:1667 weakly excites
- L23:910 weakly excites
- L24:930 weakly excites
- L25:465 weakly excites
- L23:550 weakly excites
- L24:274 weakly excites unclear: mixed technical tokens
- L24:1601 weakly suppresses
- L24:2744 weakly excites
- L25:459 weakly suppresses
- L27:1452 weakly excites
- L23:61 weakly suppresses line breaks mid-sentence
- L26:2078 weakly excites
- L27:112 weakly suppresses
- L27:454 weakly excites
- L27:825 weakly excites
- L27:2187 weakly excites
- L23:113 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
- (⏎⏎
- (⏎⏎
- [⏎
- ь
- (⏎
- ьев
- [⏎⏎
- URN
- 之人
- (⏎
Suppresses
- MSC
- Rnd
- ]={
- ={
- ={}⏎
- @register
- ContextHolder
- CEL
- NSE
- →