text boundary after a document break
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. 9 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 responds to the invisible marker that separates two unrelated web pages, especially right where a new page starts. Turning it up didn't produce a 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.
-
made many<|endoftext|>NSCAW, No. 19: Risk of Long-term Foster Care
-
959,<|endoftext|>Everyone's heard about an apple a day. As one of the most popular fruits
-
,0<|endoftext|>Guidelines for Restoring Cultural Landscapes⏎Identifying, retaining and preserving
-
7Availability: In StockCBD Stock No: WW265132<|endoftext|>Ophthalmologists work in hospitals and clinics on land and aboard ships.⏎Average Military
-
718<|endoftext|>Women who have kids later on in life may live longer, according to the findings
-
0240.<|endoftext|>By 2003, plenty of data suggested that pesticide drift was harming
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 want to add to the user's request. The user's request is to build a new application, and create a new website. The user has asked to create a new website, and I want to build a new application. So,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
soon after), it’s time the real experts divulge the dirty, little secret of Lincoln studies: We keep finding new evidence about the life of the 16th president, and some of it can be kind of shocking.⏎It’s worth
-
less need for water, fertilizer and pesticides. Also, native plants are less likely to create environmental havoc should they "escape" from your garden, unlike aggressive invasives.⏎Native plants have an important role in the local ecosystem (community
-
along the road to recovery. The rebound follows cuts to global emissions of chlorofluorocarbons and halons, which have been destroying the gase
-
scientific and industrial communities have collaborated to find safe and economical replacements for chlorofluorocarbons (CFCs), the manmade chemicals implicated in
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
- L5:2 strongly excites
- L8:2395 strongly suppresses
- L8:257 strongly excites
- L8:67 strongly excites
- L8:50 suppresses
- L8:291 suppresses
- L3:5 suppresses
- L8:341 excites
- L5:135 excites
- L3:63 suppresses
- L7:3 excites
- L3:551 excites
- L8:2765 suppresses
- L7:1902 excites
- L8:136 excites
- L0:9 weakly excites
- L3:604 weakly suppresses
- L7:2770 weakly suppresses
- L3:133 weakly excites
- L4:382 weakly excites
- L3:139 weakly excites
- L6:1800 weakly suppresses
- L1:853 weakly excites
- L2:322 weakly suppresses
Feeds later layers
- L20:96 strongly excites
- L20:116 strongly suppresses
- L27:118 strongly suppresses
- L22:156 strongly suppresses
- L19:552 strongly suppresses
- L10:32 strongly suppresses
- L20:429 strongly suppresses
- L26:120 strongly suppresses
- L26:222 strongly suppresses
- L26:312 strongly suppresses
- L19:721 strongly suppresses
- L21:270 strongly suppresses
- L22:115 strongly excites
- L21:269 strongly excites
- L26:266 suppresses
- L18:109 suppresses
- L25:323 suppresses
- L25:251 suppresses
- L26:595 suppresses
- L26:412 suppresses
- L27:104 suppresses
- L26:29 suppresses
- L10:70 suppresses
- L16:270 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
- orally
- Catalonia
- agreeing
- Carn
- Maced
- BED
- arresting
- contacted
- Hampton
Suppresses
- 字段
- ユーザー
- User
- backend
- user
- Object
- cache
- ⏎
- データ
- 程序员