word-start 'acc' fragments
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 fairly consistent: several of them 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 opening 'acc' of words like accident, accurate and accompanied. Turning it up made the model produce strange repeated text, such as a long string of zeros, which does not reveal much about the neuron.
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.
-
tisation.⏎ALBERT HOFMANN AND LSD⏎Swiss research chemist for Sandoz⏎First created LSD in 1938⏎Accidentally absorbed LSD through fingers in 1943⏎Took deliberate higher
-
that can lead to drowning.⏎It also leads to an attitude of accountability and consequences. Blaming bodies of water and warm weather would lead to an attitude of “accidents happen”.⏎Icy roads are no more a cause of crashes than deep
-
. In his own words, Captain Laudonniere states: “Having reconnoitered the river, I landed to talk […]⏎AN ARTIST WHO ACCOMPANIED THE FRENCH EXPEDITION TO FLORIDA UNDER LA
-
For the most part, nonetheless, a writer retains the single right to the title, characters, story as well as storyline of the work and also is complimentary to accredit any of those elements to others. In publishing, an author gives the author
-
Terzan 5 and others like it were once small galaxies in their own right.⏎It seems the Milky Way's massive bulge (and others like it) accrete over time as dwarf galaxies erode and are eaten up by passers-by
-
recent decision by the accountancy firm Ernst and Young who decided to drop degree classification thresholds for graduate<|endoftext|>Synchronize and Simplify Barometric Compensation Across Entire Site⏎Accurate Barometric Compensation⏎Leveloggers measure absolute pressure (water pressure + atmospheric
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: the user is a young, perhaps not very young, 6-8-year-old, who is a child of both of their young, both of their 5-6ths, and is not quite too young to be on the fr
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 size of Source and Destination IP address in IP header?⏎A. 4 bits⏎B. 8 bits⏎C. 16 bits⏎D. 32 bits⏎6-7
-
is doubled?⏎A) 2 F⏎B) F / 2⏎C) F / 4⏎D) 4 F⏎E) F /
-
. This website works best with modern browsers such as the latest versions of Chrome, Firefox, Safari, and Edge. If you continue with this browser, you
-
:<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Aim High' printed from http://nrich.math
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
- L2:1661 strongly excites
- L5:149 strongly suppresses
- L18:88 strongly excites
- L7:342 strongly excites
- L17:241 strongly suppresses
- L10:269 strongly excites joining word in a list
- L1:11 strongly suppresses
- L17:437 strongly excites
- L0:1010 strongly excites
- L18:440 strongly excites
- L17:1216 suppresses
- L16:374 suppresses
- L1:50 excites
- L1:340 excites
- L0:2910 excites
- L1:53 suppresses
- L0:1996 excites
- L18:288 excites
- L7:674 suppresses
- L15:114 excites
- L4:969 excites the letter C in vitamin C
- L14:643 suppresses
- L6:944 excites
- L0:2939 suppresses
Feeds later layers
- L22:607 strongly excites
- L22:1155 strongly excites
- L20:665 strongly excites
- L21:2906 strongly excites
- L21:590 excites
- L22:2898 excites
- L21:320 excites
- L26:2677 excites
- L20:314 excites
- L23:619 excites
- L27:2987 excites letter c inside a split word
- L20:613 excites
- L20:582 excites
- L26:2459 excites letter k starting a unit or fragment
- L22:2025 excites
- L25:2319 excites
- L22:1086 suppresses
- L26:1729 excites
- L27:2597 excites
- L26:1536 excites
- L27:2867 excites
- L24:2013 excites
- L22:2170 excites
- L26:2504 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
- illion
- YLES
- rypt
- 슭
- rust
- ipient
- istani
- opper
- пас
- .Headers
Suppresses
- unre
- enticate
- 国有资产
- rec
- kim
- ocr
- 出来了
- ⏎
- 沱
- %p