word pieces inside names and terms
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. 23 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 a middle chunk of a longer name or word that the model has split into pieces, like the 'l' in a surname or the 'elt' in a place name. When it was turned down, the model kept coming back to the words 'objective' and 'goal'. Turned up, the writing just wandered off topic.
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.
-
large open air theater built in a semicircular pattern, with ascending rows of seats built into a hillside<|endoftext|>Posts tagged education⏎Education Week writer Larry Ferlazzo reached out to me and three other Education & Mental Health professionals to help respond
-
from taking her place in the sky. What might she have accomplished had she lived to a ripe old age?<|endoftext|>Sixteen years ago, Patricia L. Gerbarg, a Kingston, N.Y., psychiatrist, contracted Lyme disease. But as
-
1416 and initially worked as a Greek scribe. His study under the famous schoolmaster Vittorino da Feltre gave him a mastery of Latin rare among Greek emigres, and by
-
Arctic Circle.⏎The ICEHOTEL in Swedish Lapland is celebrating its 20th year.<|endoftext|>Source: Video and Images Created by Amanda Soderlind⏎Welcome to this lesson today on autosomal recessive traits and disorders.
-
“In the West, we’re not used to the idea that one drug — or herb — can have a broad range of physical and mental health benefits,” Gerbarg explains. But when you consider that stress has been linked to a host of
-
icit that human-like behavior?”⏎Pierson is joined on the paper by Daniela Rus, the Viterbi Professor of Electrical Engineering and Computer Science; Sertac Karaman, associate professor of aeronautics and astronautics;
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 an objective and an objective of the goal. I have an objective of the objective. I have an objective of the objective. I have an objective of the objective. The object of the objective is the object of the objective. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
they should be conducting⏎- How to apply the results to create effective, high performing designs⏎The session will be in English language.<|endoftext|>Statement of Purpose⏎THE CIVIL WAR CENTER⏎The Civil War was arguably the most important event in
-
Jepson Interchange (more information)⏎©Copyright 1993 by the Regents of the University of California⏎AND IS MAINTAINED FOR ARCHIVAL PURPOSES ONLY⏎Perennial to trees, from membranous bulb, fib
-
pollution.⏎WiseGeek has one of the clearest definitions of Cap and Trade: A cap and trade system is a method for managing pollution, with the end goal of reducing the overall pollution in a nation, region, or industry. Many
-
Seven years ago Albert Moore died."⏎Scanned image and text by George P. Landow⏎[This image may be used without prior permission for any scholarly or educational purpose.]<|endoftext|>Tough Task for High School Students: Not all of us are
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
- L22:1483 strongly excites
- L21:2060 strongly excites
- L19:1767 strongly suppresses
- L18:1028 excites
- L22:2793 excites
- L20:2026 excites
- L21:2211 suppresses
- L21:2168 excites
- L22:1725 suppresses
- L19:2676 excites
- L12:1633 suppresses
- L5:2027 suppresses
- L20:1602 excites
- L18:2659 suppresses
- L22:2856 excites
- L20:2589 suppresses
- L13:1753 suppresses
- L6:1737 weakly suppresses
- L12:1477 weakly suppresses
- L11:2117 weakly excites
- L16:2095 weakly excites
- L22:2630 weakly excites
- L18:2747 weakly suppresses
- L12:2450 weakly suppresses
Feeds later layers
- L27:2772 strongly suppresses
- L25:2948 strongly suppresses
- L27:1039 strongly suppresses
- L27:1845 strongly suppresses
- L24:1675 suppresses
- L27:2053 suppresses
- L26:2714 weakly suppresses
- L24:2784 weakly suppresses
- L25:3019 weakly suppresses
- L25:1403 weakly suppresses
- L25:2611 weakly suppresses
- L26:2629 weakly suppresses
- L27:3050 weakly suppresses
- L25:953 weakly suppresses
- L25:1602 weakly suppresses
- L24:1282 weakly suppresses
- L26:1969 weakly excites
- L25:1473 weakly suppresses
- L25:2563 weakly excites
- L26:1731 weakly suppresses
- L25:1601 weakly excites
- L24:2077 weakly excites
- L24:1963 weakly suppresses
- L24:2807 weakly suppresses
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
- шив
- @$
- pond
- iless
- tục
- kar
- 湾区
- ру
- CodeAt
- ollapsed
Suppresses
- 目的
- objective
- goal
- Objective
- purpose
- Objective
- objectives
- objective
- aim
- 目标