blank slot before a number
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. 4 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 an empty spot right before a number, like the gap in 'pages 1-2' or 'Published - 2002'. When it was turned up, the model's writing got stuck repeating ideas, so steering adds little.
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.
-
elements of a successful speech. Not only do they get to become better-informed...⏎7th - 11th Social Studies & History⏎President Obama's State of the Union Address
-
caused<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally
-
thisDataSet.Tables ' Get a specific table by name. Dim table As DataTable = tablesCollection("Suppliers") Console.WriteLine(table.TableName) End Sub⏎Available since 1.1<|endoftext|>A musical trip through the 60’s featuring the songs
-
declining concentration of tetracycline. (C) 2002 Elsevier Science B.V. All rights reserved.⏎|Publication status||Published - 2002|⏎- whole-cell biosensors⏎- antibiotic resistance<|endoftext|>Latin
-
to facilitate children’s science exploration through everyday kitchen substances and experiences such as mixing, pouring, heating, and cooling. Full text» [PDF, 162KB]⏎Marvelous Explorations Through Science and Stories (MESS
-
igación con células madre embrionarias humanas revela los primeros pasos en el desarrollo humanoContributed by: AdminStaff1 · Views: 1,130⏎Contributed by: AdminStaff1 · April
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: of the content, I have been thinking of the matter. The question is: What is the purpose of the matter? It seems the content, I have been thinking of the matter. The question is: What is the purpose of the matter
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Earth. We will learn a lot more with Sutter's Mill."<|endoftext|>For people with disabilities, technology has become an invaluable tool to help level the playing field in
-
that both the plaintiff and the defendant enjoy a level playing field during their trial. In other words, bias deprives a party of the constitutional
-
incorporate an arts program that has as much integrity as the science or math programs that already exist.⏎And the list goes on: education in the arts levels the playing field for disadvantaged children<|endoftext|>Philosophers, psychologists, and ordinary people are all interested
-
zip or compress.⏎Step 2: To compress or zip multiple files into one, use the selection rectangle to drag all the files and select them together.⏎Step 3
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
- L3:626 excites
- L0:602 weakly excites words right before a number
- L1:2340 weakly suppresses
- L2:549 weakly suppresses
- L0:414 weakly excites blank after a dash in number ranges
- L0:1506 weakly excites
- L2:75 weakly suppresses
- L3:272 weakly excites
- L3:1000 weakly excites
- L3:2300 weakly excites
- L3:230 weakly suppresses
- L1:217 weakly suppresses
- L2:422 weakly excites
- L2:510 weakly excites
- L0:1635 weakly excites
- L2:801 weakly suppresses
- L0:484 weakly excites
- L2:2328 weakly suppresses
- L2:1307 weakly suppresses
- L3:171 weakly excites
- L1:2677 weakly suppresses
- L3:421 weakly suppresses
- L3:691 weakly excites
- L1:275 weakly suppresses comma closing a main clause
Feeds later layers
- L10:220 excites hyphen inside compound words
- L16:471 excites blank spot before a year
- L17:285 excites
- L18:377 excites number right after 'about' or 'less than'
- L16:256 excites number right after 'age of' or 'ages'
- L19:581 excites blank slot before a number
- L22:824 excites blank before an expected number
- L22:372 excites empty spot before a number or "3D"
- L23:77 excites the space right before a number
- L18:1534 excites space between month and day number
- L21:339 excites spot just before a year
- L8:111 excites the space before a year in dates
- L13:923 excites word right before a number or year
- L27:283 excites the spot right before a number
- L7:229 excites first digit of a split number
- L19:26 suppresses
- L25:422 excites
- L5:386 excites
- L21:175 excites
- L27:170 excites unclear: spaces in mixed-language text
- L17:1722 excites number right after a product or name
- L26:2003 excites hyphen joining two word parts
- L15:368 excites
- L19:298 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
- antino
- <=$
- noon
- unanimous
- 日讯
- скоп
- latin
- 下周
- gether
- 工作报告
Suppresses
- s
- اء
- sr
- rel
- rze
- gre
- gende
- sur
- New
- sb