the word 'or' between alternatives
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. 15 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 word 'or' when it offers a second option, like 'a piece of white cardstock or poster board'. Turning it up did not produce a clear theme in the generated text.
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.
-
Dane breed comes from German descent without any known associations to the country of Denmark, at least that’s what canine historians say.⏎The breed might be four centuries old or even more, but the actual origins of the breed are murky. Great Danes
-
; ribs 5, acute or round-obtuse; vittae 1 in each furrow, 2 on commissure. Seed face plane or slightly concave. Carpoph
-
edges of hedgerows and woodland, and by the roadside. Wild chervil blooms from mid spring to early summer. It is an herbaceous biennial or short-lived perennial plant in the family Apiaceae (Umbellifera
-
2012.⏎Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.3 or any later version published by the Free Software Foundation; with no Invariant Sections,
-
a life-like silhouette.⏎In a darkened room, have an individual stand sideways against a wall. Behind⏎the person, tape a piece of white cardstock or poster board on the wall.⏎The board should start about mid-chest and go
-
a record of a David de Boivil witnessing a charter as early as 1164.<|endoftext|>For the questions below, use the most efficient Excel formula or procedure to find your answer. Calculations by hand (or calculator) should only
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 same set of algorithms are being used to calculate the same set of data in the same set of set of data set, which the same set of data set, the same set of set of data set, the same set of set of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
murex shellfish, known as Tyrian purple. This color was, in many cultures of ancient times, reserved for the use of royalty, or at least nobility.⏎It was often attacked by Egypt, besieged by Shalman
-
appear: cataract and AMD (Age-Related Macular Degeneration) are the most limiting. But it is possible to prevent them or, at least, to
-
Cinderella a folktale or fairy tale?⏎Cinderella is considered a fairy tale, not a folk tale. Folk tales have a basis in reality or something which happened in history.⏎What is a common event in a folktale<|endoftext|>
-
voted down.⏎Jay founded the New-York Society for Promoting the Manumission of Slaves, and Protecting Such of Them as Have Been, or May be Liberated or the New York Manumission Society, and became its first
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
- L10:135 strongly excites 'or' in alternative requirements
- L4:108 excites "or" offering a choice in web text
- L11:273 suppresses 'and' joining clauses in instructions
- L7:1475 excites 'or' joining alternatives in lists
- L13:1208 excites
- L14:2065 excites
- L13:553 suppresses
- L12:836 suppresses
- L10:901 suppresses
- L1:1710 suppresses
- L3:1200 excites
- L13:411 excites
- L7:1646 suppresses
- L11:406 suppresses
- L14:2014 excites
- L12:1094 excites
- L11:1861 weakly suppresses
- L6:733 weakly suppresses
- L9:2010 weakly suppresses
- L12:1452 weakly suppresses
- L12:2393 weakly excites
- L0:198 weakly excites
- L10:891 weakly suppresses
- L9:352 weakly suppresses
Feeds later layers
- L16:925 excites
- L18:1138 excites
- L17:2166 excites
- L16:1782 excites
- L16:935 suppresses
- L19:1343 excites
- L16:1959 excites
- L16:2387 suppresses
- L16:1913 suppresses
- L16:1068 suppresses
- L16:2289 excites 'or' linking two alternatives
- L16:1181 weakly excites
- L16:1735 weakly excites
- L21:1974 weakly excites
- L16:2818 weakly suppresses
- L19:2597 weakly excites
- L16:420 weakly excites
- L22:266 weakly excites
- L16:678 weakly suppresses
- L18:2197 weakly excites
- L24:1184 weakly excites
- L16:2261 weakly excites
- L16:1492 weakly suppresses
- L20:2429 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
- 까
- other
- ichte
- 别的
- else
- simply
- /dd
- 仅仅
- IDs
- ENDING
Suppresses
- seriousness
- GridBagConstraints
- ynchron
- occupies
- 有大量的
- -Length
- ї
- Loves
- 走下去
- -'+