"the" in nickname titles
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. 6 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 fires on the word "the" when it comes between a person's name and a nickname, like "Rosey the Riveter", "Yaroslav the Wise" or "Robert the Bruce". Turning it up did not produce any clear change 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.
-
1000-piece puzzle of World War II propaganda posters tells the story from the Home Front. Who wouldn't recognize Rosey the Riveter, the woman, sporting a red bandana, her bicep
-
grand prince. This succession pattern was generally followed through the reigns of Svyatopolk (1015–19); Yaroslav the Wise (1019–54); his sons Izyaslav
-
tiny bit safer.⏎Cyril The Lonely Cloud by Tim Hopgood⏎It’s a bright and lovely day, the perfect day for a picnic, until Cyril the lonely cloud shows up. Everybody agrees he is a bore and a spoilsport
-
want to be told something like ‘I’m going to clean you,’ and then the robot cleans. But the opposite was true,” Chen said.⏎Image: Cody the robot touches one of 56 study participants. (Georgia Tech)⏎- Robot
-
to King Edward I of England during the latter's brief conquest of Scotland. However, eight years later, he was appointed Warden of Stirling Castle by Robert the Bruce of Scotland to whom he had willingly pledged allegiance. And yet Edward of England
-
to a 40-day siege of the Church of the Nativity in Bethlehem. This is probably the oldest Christian church in the world, built by Constantine the Great and dating from AD330. A member of the girl’s family
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, and, and. The, and. The, and. The, and. The, and. The, and. The, and. The, and. The, and. The, and. The, and. The,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
she travels to Wyoming and shares them with her new family. Here are two articles about Swedish Christmas traditions.⏎Swedish Christmas Traditions:⏎from Christmas World/WorldofChristmas.net⏎In Sweden, Christmas begins with Saint Lucia ceremony on 13
-
"wielding his machete until the foe had<|endoftext|>|Theme||Visible||Selectable||Appearance||Zoom Range (now: 0)|⏎Well-to-well tracer tests contribute significantly to the reservoir description, which is essential in determining the best
-
Walter Lord, author of the classic "A Night to Remember", describes the process vividly in his 1986 sequel, "The Night Lives On". He compares the ships of Titanic's day to the first great liner, the
-
n), composed between 650-550 BC, that is believed to be one of the oldest.⏎The Homeric Hymn to Demeter⏎However, the episode beforehand is as important as the story itself. It starts
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
- L5:1623 weakly suppresses
- L2:407 weakly suppresses
- L3:2962 weakly excites
- L0:2526 weakly excites
- L1:2751 weakly excites
- L4:2659 weakly excites
- L4:758 weakly suppresses
- L5:1658 weakly excites
- L4:2964 weakly suppresses
- L2:2337 weakly excites
- L3:682 weakly excites
- L4:2373 weakly excites
- L5:38 weakly suppresses
- L4:2656 weakly suppresses
- L1:2357 weakly suppresses
- L3:2043 weakly excites
- L2:2966 weakly suppresses
- L3:568 weakly suppresses
- L0:1318 weakly excites
- L5:201 weakly excites
- L3:108 weakly suppresses
- L4:672 weakly suppresses
- L1:188 weakly suppresses
- L1:1385 weakly suppresses
Feeds later layers
- L7:26 suppresses
- L9:758 excites
- L10:217 suppresses
- L10:457 suppresses
- L9:1607 suppresses
- L13:1550 suppresses
- L7:6 excites
- L14:282 excites
- L18:93 excites
- L10:105 excites
- L16:323 excites
- L9:845 excites
- L19:413 suppresses
- L10:1748 weakly excites
- L19:223 weakly excites
- L15:1601 weakly suppresses
- L10:121 weakly excites
- L15:181 weakly excites
- L19:711 weakly suppresses
- L19:211 weakly excites
- L20:398 weakly suppresses
- L15:286 weakly excites
- L9:1026 weakly excites
- L9:401 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
- 书
- intelligent
- odable
- luck
- 多元
- Jud
- eta
- av
- 进步
Suppresses
- acement
- cies
- erties
- /close
- (compare
- exels
- :NS
- �数
- rish
- comparer