end of document token
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. 19 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 marker the model uses to show that one web page or article has ended and a different one begins. It is strongest right where a text stops. Turning it up or down just made the model's writing collapse into repetition or nonsense.
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.
-
and to support supplement programs aimed at boosting the quality of education in urban settings.⏎Under Chief Justice Robert Wilentz (1979-199<|endoftext|>Tobacco Industry Influence on the U.S. Military⏎Location(s): United
-
to be debate in the party about the way forward, but its leaders have shown few signs that they are willing to give up, or even share, power.<|endoftext|>The Federal
-
by so many news sources.⏎Michelle Kaffenberger is a former research assistant at The Heritage Foundation and is currently a graduate student in Economic Development at Vanderbilt University.<|endoftext|>As a gastrointestinal surgeon and critical care specialist, John Alverdy has treated many
-
status||Published - 17 Oct 2008|⏎- age-related changes⏎- equid skulls⏎- horse use⏎- work-related changes<|endoftext|>Make your Arduino play the French Can Can using a buzzer. No pliers
-
melanoma.⏎This fact sheet from the National Cancer Institute explains the subject thoroughly. It also lists a number of the clinical trials currently underway on experimental therapeutic vaccines.<|endoftext|>What Are PK Boosters?⏎Just like with anything, gardening and⏎PK Booster
-
and Emerging Technologies Flagship competition launched by the European Commission, the other one being the “Graphene” project led by Swedish theoretical physicist Jari Kinaret.<|endoftext|>
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 a 4444-4555-5555-5555-5555-5555-5555-5555-5555
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
civil transport aircraft designed by a UK team is leading a pack of three concepts undergoing preliminary studies for the three-year, €7 million ($9 million) European hypersonic research project, Long Term Advanced Propulsion Concepts and Technologies (LAPCAT).⏎
-
2006, 2nd ed. ) by Patriia C.⏎John Winthrops City on a hill EssayJohn winthrops city on a hill John Winthrops 1630 speech not only contributed an enduring metaphor
-
Eating disorders most commonly affect women and occur in adolescence and young adulthood, though they can occur in anyone, of any gender, at any age.”⏎Eating disorders often occur alongside other mental health disorders such as obsessive compulsive disorder, mood disorders
-
heavy-ion collisions at the Relativistic Heavy Ion Collider (RHIC), a facility that is dedicated to studying nuclear physics. These heavy strange baryons, containing at least one strange quark, still cannot be observed directly,
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
- L16:562 suppresses
- L17:212 excites
- L5:46 suppresses
- L17:91 suppresses
- L17:495 excites
- L2:135 excites
- L15:219 suppresses
- L3:59 suppresses
- L17:2843 weakly suppresses
- L18:257 weakly suppresses
- L15:159 weakly excites
- L15:2311 weakly excites
- L17:183 weakly suppresses
- L18:260 weakly suppresses
- L17:102 weakly excites
- L14:397 weakly suppresses
- L3:115 weakly excites
- L8:2577 weakly suppresses
- L17:1527 weakly excites
- L6:7 weakly excites
- L17:339 weakly excites
- L5:57 weakly excites
- L15:2073 weakly suppresses
- L16:2508 weakly excites
Feeds later layers
- L27:727 weakly excites
- L20:628 weakly excites
- L21:106 weakly suppresses
- L21:1227 weakly suppresses
- L27:897 weakly excites
- L27:2298 weakly excites
- L20:681 weakly excites
- L22:213 weakly suppresses
- L27:1487 weakly excites
- L20:897 weakly excites
- L21:348 weakly excites mid-sentence line breaks
- L26:460 weakly suppresses
- L26:1518 weakly excites
- L27:1568 weakly excites
- L27:181 weakly excites
- L27:1118 weakly excites
- L27:2899 weakly excites
- L20:252 weakly suppresses
- L27:2038 weakly suppresses
- L27:2352 weakly excites
- L27:1767 weakly excites
- L26:380 weakly excites
- L20:294 weakly excites
- L27:2756 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
- ...⏎⏎⏎⏎
- 词条
- nâ
- —⏎⏎
- .…⏎⏎
- �能
- â
- ��
- ries
- […]⏎⏎
Suppresses
- gaz
- .scalajs
- urdy
- Mc
- Rule
- ⏎
- BASEPATH
- ownt
- ট
- <<