line breaks after numbers or fragments
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. 2 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 line break that comes after a number or a short, cut-off fragment at the end of a line, often in lists, page numbers or headings. Its negative side is weaker and falls on words like 'day' and 'boating'. Turning it up did not produce a clear theme in the written 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.
-
1,2⏎Whether it be driving the kids to extracurriculars, getting up at
-
University, USA⏎"TEEB brings a rigorous economic focus to bear on the problems of ecosystem degradation
-
ahead of formal qualifications’. 1⏎Research by the Institute of Leadership and Management with Ashridge Business School suggests that over
-
half, and half again, to⏎make four smaller rectangles. How many different ways can you fold⏎This problem focuses
-
On the other side, they⏎should write one sentence about Abraham Lincoln's life. After the penny is⏎created
-
3⏎Scientists learning what it takes for seals to live in the wilds of Antarctica⏎
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 I have a lot of information, and I have been told that I have a lot of information. What is the answer to this? - I have been told that I have a lot of information. What is the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
<|endoftext|>For most of his writing career, Vladimir Nabokov was frustrated by both the Freudian and behavorialist schools of psychology that reigned in his day, seeking a richer and more complex view of the mind than either provided. In
-
was replaced by the St. Lawrence Seaway which opened in 1959. In 2002, Lachine was reopened as a boating recreational area. An environmental reclamation project continues to clean up the polluted canal.⏎
-
would do<|endoftext|>Printed on a large, single sheet of paper (thus ‘broadside’), these cheap, everyday publications were the popular songs of their day, reaching the height of their popularity in the 1660s.
-
sheriff’s department, registry of deeds, and courthouses.⏎Portland harbor is a working waterfront serving tankers, tourists, and fisheries, along with recreational boating from year-round and summer residents. Ferries connect the mainland to the many Cas
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
- L0:37 weakly excites
- L0:886 weakly excites
- L0:1224 weakly suppresses
- L0:592 weakly excites
- L1:1861 weakly excites
- L1:2791 weakly excites
- L0:1329 weakly excites
- L1:1277 weakly suppresses
- L0:178 weakly suppresses
- L0:2270 weakly suppresses
- L1:1190 weakly excites
- L0:268 weakly suppresses
- L1:1125 weakly suppresses
- L0:2432 weakly excites
- L1:2066 weakly suppresses
- L0:1243 weakly excites
- L0:267 weakly excites
- L1:1574 weakly excites
- L1:1537 weakly excites
- L0:532 weakly excites
- L1:1249 weakly suppresses
- L1:1511 weakly suppresses
- L0:1564 weakly excites
- L1:2986 weakly excites
Feeds later layers
- L19:902 excites
- L3:1353 weakly suppresses
- L4:795 weakly suppresses
- L18:1217 weakly suppresses
- L12:2055 weakly suppresses
- L27:387 weakly suppresses 'a' right after 'played' or 'play'
- L3:75 weakly suppresses
- L22:357 weakly excites
- L25:1644 weakly suppresses
- L18:2851 weakly suppresses
- L24:272 weakly suppresses
- L22:152 weakly excites
- L10:170 weakly excites
- L23:754 weakly suppresses
- L7:1960 weakly suppresses
- L4:547 weakly suppresses
- L27:2799 weakly suppresses the letter d inside words
- L21:2311 weakly excites
- L26:1063 weakly suppresses
- L27:1970 weakly suppresses
- L8:3064 weakly suppresses
- L7:2871 weakly suppresses
- L27:1749 weakly suppresses hyphen joining word parts
- L21:995 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
- imagine
- ide
- "`⏎⏎
- ment
- мен
- ife
- _Err
- Apache
- 供图
- owl
Suppresses
- ila
- quina
- enty
- esus
- →pr
- illard
- ByUsername
- venience
- าย
- ILA