'as' in Incas, 'te' in intelligence
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. 1 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 middle pieces of certain words, such as the 'as' in 'Incas' and the 'te' in 'INTELLIGENCE'. Steering changed the text but did not point to one clear theme.
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.
-
the rising sun. Both mountains overlook the entire site, down to the river below, which puts the Incas· great achievement of construction into perspective.⏎The Incas worshipped the sun god Inti, so the summer solstice was the most
-
between two giant peaks. Near-vertical slopes drop away on either side, down to a massive bend in the Urubamba River. What could have motivated the Incas to undertake such a construction at this remote location in the Andean cordillera
-
the Intelligence made of neutron particles, cannot be destroyed. In the INTELLIGENCE OF HUMAN BEING IS THE TRUE INDENTITY OF THE PERSON,
-
them to experience things through the five senses of : TOUCH, SMELL, TASTE, HEAR, and SEE. At the time of death both, INTELLIGENCE, and Consciousness leaves the body. Consciousness being made of Electron
-
fold formula is the master pattern of which living beings on earth have been made.. A human being is a threefold being united into one: 1. INTELLIGENCE. 2. CONSCIOUSNESS. 3. PHYSICAL
-
in solving the edge loading problem due to excess misalignment and deflection in aerospace actuation gearing is to localize tooth-bearing contact by crowning the teeth. Irrespective of the applied load, if the misalignment and/or deflection are large
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 a student at a university, and I have a degree. I am currently working in a research position. I am a professional, and I have a full degree. I have a degree in science, and I am currently a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Cyberbullying⏎Bullying is part of the cultural landscape for adolescents today, even as it has been in previous generations. Some teenagers are still bullied in person–a 2014 study found that while in-school bullying among teens
-
istic features exists, the government plays a dominant role in controlling the economic activities and in the social welfare of the people.⏎2.A welfare state is socialistic in nature. It is based on the principles of equality and is keen to provide equal opportunity
-
in the world. Many colleges and universities offer online courses and programs designed specifically for those with disabilities, allowing them to pursue a degree or certification without attending classes in person.⏎Whether they hope to climb the corporate ladder or start their own business, they
-
and being able to positively present yourself in writing and in person. Understanding the importance of professional networking and showing how you would leverage a developed network
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
- L0:1301 suppresses
- L0:1224 suppresses
- L0:178 suppresses
- L0:781 weakly excites
- L0:743 weakly excites
- L0:387 weakly excites
- L0:3038 weakly excites
- L0:2411 weakly excites
- L0:181 weakly excites
- L0:723 weakly excites
- L0:210 weakly suppresses
- L0:39 weakly suppresses 'the' right after 'On' in 'On the other hand'
- L0:367 weakly excites
- L0:2579 weakly suppresses
- L0:2615 weakly excites
- L0:962 weakly suppresses
- L0:2543 weakly suppresses
- L0:2009 weakly suppresses
- L0:2817 weakly suppresses
- L0:1042 weakly suppresses "This" starting a labelled sentence
- L0:356 weakly suppresses
- L0:1034 weakly suppresses
- L0:514 weakly suppresses
- L0:1610 weakly suppresses
Feeds later layers
- L16:915 weakly suppresses
- L2:2406 weakly suppresses
- L25:845 weakly suppresses
- L14:2135 weakly suppresses
- L22:492 weakly suppresses
- L5:417 weakly suppresses
- L17:1897 weakly suppresses
- L14:2621 weakly suppresses
- L20:1134 weakly suppresses
- L7:2414 weakly suppresses
- L19:223 weakly excites
- L11:1780 weakly suppresses
- L4:1530 weakly suppresses 'see' in 'let's see' and 'to see if'
- L12:901 weakly excites
- L22:255 weakly suppresses
- L10:289 weakly suppresses
- L3:2520 weakly suppresses
- L16:656 weakly suppresses
- L15:756 weakly suppresses
- L17:1269 weakly excites
- L23:1246 weakly suppresses
- L3:1204 weakly suppresses
- L24:498 weakly excites
- L9:2563 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
- 们的
- ي
- s
- 特别是
- sayılı
- يش
- erno
- 们
- ylation
- chein
Suppresses
- жу
- ئة
- .readString
- 흡
- structors
- ifu
- 祖父
- nett
- iquement
- +#+#+#+#+#+