end of text after an abrupt cut
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 fires on the invisible end-of-document marker, mostly when the text before it stops in the middle of a thought or right after a number. Turning it up or down did not change the output in any clear way.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
tried to do this multiple times but I'm really confused about the equations and when to apply them.<|endoftext|>Plant Holiday Amaryllis Bulbs⏎If you had a blooming am
-
to review it. We're adding the slope at a single point to the slope over a big area, which is something I can't quite work out visually.<|endoftext|>The world produces more than enough food to feed everyone, yet each night more than
-
, 2012<|endoftext|>Some weather conditions are simply not fun to ride in. As cyclists, windy days
-
65-84 age group.⏎<|endoftext|>An unstructured genre that blends high aesthetic standards with nonfiction commentary, the journalistic
-
,0<|endoftext|>Guidelines for Restoring Cultural Landscapes⏎Identifying, retaining and preserving
-
24.12 million.<|endoftext|>Dr. Smith, a family medicine physician, is preparing for a full day of
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: "It was a special time of the day, and the time of the day and the time of the day of what it would be, and the time the time of the day, the time of the day, and the time of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
What is the highest overlook on the rim of the Grand Canyon?⏎Point Imperial, 8,
-
What is the difference between 3-axis and 5-axis?⏎Three-axis machining is processed by
-
4⏎1 T T F T⏎2 T T T F⏎3 F T T F⏎4 T F F T<|endoftext|>July 7, 2016⏎Experimenting with Post-Quantum Cryptography⏎Quantum computers
-
experiences right at home. In this article, we will look at five strategies for motivating students and keeping their brains engaged over the summer.⏎
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
- L10:507 suppresses
- L8:808 weakly excites
- L6:254 weakly suppresses
- L6:1911 weakly suppresses
- L0:3047 weakly excites
- L12:141 weakly excites
- L7:2936 weakly excites
- L6:64 weakly suppresses
- L14:2348 weakly excites
- L10:271 weakly suppresses
- L6:1273 weakly suppresses open parenthesis in citations
- L6:492 weakly excites
- L11:1325 weakly suppresses
- L11:954 weakly suppresses
- L4:2843 weakly excites
- L6:1089 weakly suppresses
- L9:2578 weakly suppresses
- L12:1705 weakly excites
- L12:1600 weakly suppresses
- L7:302 weakly suppresses
- L8:148 weakly excites
- L0:1369 weakly suppresses
- L10:1095 weakly suppresses
- L12:178 weakly suppresses
Feeds later layers
- L22:2857 weakly suppresses
- L24:2207 weakly suppresses
- L27:1764 weakly suppresses
- L23:3039 weakly suppresses prepositions in how-to location phrases
- L24:2498 weakly suppresses unclear: word before an event or name
- L22:1552 weakly suppresses
- L23:2691 weakly suppresses
- L20:1989 weakly excites
- L20:1047 weakly suppresses
- L24:730 weakly suppresses
- L18:400 weakly suppresses
- L17:1072 weakly excites
- L22:2377 weakly suppresses
- L23:520 weakly suppresses
- L21:55 weakly excites
- L16:1709 weakly excites
- L23:1381 weakly suppresses
- L18:1137 weakly suppresses
- L23:2539 weakly suppresses
- L22:2348 weakly suppresses
- L21:2464 weakly suppresses
- L20:240 weakly excites
- L20:1076 weakly suppresses
- L20:14 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
- ertz
- migrate
- Frage
- ette
- ')]⏎
- iker
- esthetic
- udit
- edia
- дер
Suppresses
- 县委
- cxx
- ifornia
- actories
- 社会各界
- AppComponent
- Belgi
- 同志们
- 干警
- 昼