parentheses and code 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. 11 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 responds to opening brackets and punctuation in technical text, such as '(CDC)' after an organization name and parts of code. It is mixed, and the strongest hit is a bullet mark in a forum post. Turning it up or down did not give a clear story.
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.
-
in the field from August 2 – August 20.<|endoftext|>Genetics as it applies to evolution, molecular biology, and medical aspects.⏎10 posts • Page 1 of 1⏎Hope somebody can help, I've got an
-
11). http://www.cdc<|endoftext|>|This article is about the assassin of pharaoh Cleopatra. For other uses, see Aya or Amun.|⏎Prior to establishing the Hidden Ones, A
-
age of 25 in 2006.⏎Researchers T.J. Mathews and Brady E. Hamilton of the Center for Disease Control and Prevention’s (CDC) National Center for Health Statistics, Division of Vital Statistics, Reproductive
-
a charge of bad taste,<|endoftext|>8 edition of Principles of Biblical Interpretation found in the catalog.⏎April 1, 2
-
Released: Nov 9, 1993.⏎- Full Report:⏎Pursuant to a congressional request, GAO reviewed the Department of Energy's (DOE) plans
-
, December 2, 2012.⏎The contest is a partnership with The Planetary Society in Pasadena, California; the Massachusetts Institute of Technology’s (MIT) Lincoln Laboratory in Lexington, Massachusetts; and the University of Arizona in Tucson
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 am not going to be in a position of need to use my resources for the benefit of a possible crisis of need..." I am not going to be in a position of need to use my resources for the benefit of a possible crisis
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
-- Important numbers!⏎xs = [6, 9, 13, 42]⏎-- Without library⏎f :: Bool -> [Int]⏎f b = if b then xs else ⏎-- With library⏎g :: Bool -> [
-
found by quadrature and is⏎To use MATLAB we first construct an M-file that defines the equation.⏎% filename rhs.m⏎function ydot=rhs(t,y)%compute the rhs of eqns.⏎%for any input values t, y
-
parent graph of the cotangent function f(x) = cot x,⏎by using the same techniques you use to find the tangent
-
conuts and banana leaves<|endoftext|>const void * memchr ( const void * ptr, int value, size_t num );⏎void * memchr ( void * ptr, int value, size_t num );⏎Locate character in block of memory⏎
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
- L8:1610 suppresses
- L10:406 excites
- L10:392 excites
- L5:272 excites
- L10:1187 excites
- L7:248 excites
- L6:276 suppresses
- L8:288 suppresses
- L7:457 excites
- L1:993 weakly suppresses
- L2:296 weakly suppresses
- L10:779 weakly suppresses
- L10:1425 weakly excites
- L7:718 weakly excites
- L7:513 weakly excites
- L8:1334 weakly suppresses
- L10:1251 weakly suppresses
- L10:1625 weakly suppresses
- L3:1634 weakly excites
- L7:158 weakly excites
- L8:144 weakly suppresses
- L4:2907 weakly excites
- L1:1535 weakly excites
- L10:681 weakly suppresses
Feeds later layers
- L15:1775 suppresses
- L16:1856 suppresses
- L12:1163 suppresses
- L16:2881 suppresses
- L12:855 suppresses
- L15:2582 suppresses
- L15:2454 suppresses
- L15:487 suppresses
- L13:568 suppresses
- L18:653 suppresses
- L12:829 suppresses
- L13:1463 suppresses
- L13:1987 suppresses
- L17:946 suppresses
- L13:711 suppresses
- L16:375 suppresses
- L14:726 suppresses
- L21:1347 weakly suppresses
- L13:2612 weakly suppresses
- L13:845 weakly suppresses
- L12:1298 weakly suppresses
- L16:1402 weakly suppresses
- L13:1283 weakly suppresses
- L12:443 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
- .side
- side
- idence
- stown
- continent
- 高的
- aan
- RoleId
- estone
- missible
Suppresses
- 濯
- ilos
- REST
- rest
- proj
- ldr
- /array
- 项目的
- 援
- 缇