opening parenthesis before an acronym
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. 22 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 open bracket right before a short form or abbreviation, such as '(SpaceX)' or '(SAE)'. When turned up, the model's output became unrelated text with no clear link to acronyms, so the steering evidence does not support the pattern.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
Space Exploration Technologies (SpaceX)⏎Due to the recent retirement of the space shuttle, NASA⏎now
-
thread’s major diameter, minor diameter, pitch diameter, pitch (in case of metric threads) or threads per Inch (TPI in case of Standard American Equivalent (SAE) threads), flank angle, etc. Normally, a GO / NO
-
types of external beam radiation, such three-dimensional conformal radiation therapy (3D-CRT) and intensity modulated radiation therapy (IMRT) are
-
scanning (SEM) or transmission electron microscopy (TEM) coupled with energy dispersive x-ray spectroscopy (EDX). This technique can provide not only information on the shape of aerosol
-
and resource requirements. Developments in GIS have provided agencies with the ability to collaborate more efficiently and effectively. Emergency managers resolve natural disaster challenges using the Comprehensive Emergency Management (CEM) approach. CEM is broken down into four phases: mitigation,
-
While computers can already fly a passenger jet better than a trained human pilot, people are still faced with the dangerous yet tedious task of driving automobiles. Intelligent Transportation Systems (ITS) is the field that focuses on integrating information technology with vehicles and transportation infrastructure
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 a 3-1-1-3-3-3-3-3-3-3-3-3-3-3-3-3-3-3-3-3-3-3-
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
74,820 per year for a semi-private room in a nursing home.⏎- $6,965 per month, or $83,580 per year for a private room in a nursing home.⏎- $
-
health officials.⏎To date in 2014, thirteen animals have tested positive for rabies in Erie County (five raccoons, five bats, and one each of a cat, skunk and fox).⏎On June 11
-
in French in the French part) are available with closed-captioning. Users will also find 164 film clips (80 in English; 84 in French in the French part).⏎These 120 films are
-
which means that they were propelled primarily or entirely by sail. The length of the round ships ranged from 26 to 48 feet (8 to 14.7 meters), and they were between 8 and 16 feet
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
- L21:1073 excites
- L21:2591 suppresses
- L20:1845 excites digits inside numbers
- L21:569 excites
- L21:1359 suppresses
- L21:2789 excites
- L21:1053 suppresses
- L21:1384 excites
- L15:187 weakly suppresses
- L13:1079 weakly excites
- L16:2864 weakly suppresses location or target word before a noun phrase
- L21:2069 weakly excites
- L20:1906 weakly suppresses
- L20:2944 weakly excites
- L2:2044 weakly excites
- L19:1008 weakly excites
- L21:1154 weakly suppresses
- L6:1205 weakly excites
- L4:119 weakly excites
- L4:2364 weakly excites
- L8:591 weakly suppresses
- L6:1198 weakly suppresses
- L0:2287 weakly suppresses
- L6:6 weakly suppresses
Feeds later layers
- L23:471 weakly suppresses
- L24:2602 weakly excites
- L23:107 weakly suppresses
- L23:54 weakly excites
- L23:133 weakly suppresses
- L25:2545 weakly excites
- L23:1359 weakly excites
- L23:1259 weakly excites
- L23:2047 weakly suppresses
- L23:710 weakly excites
- L26:532 weakly excites
- L26:834 weakly excites
- L25:2208 weakly excites
- L23:671 weakly suppresses
- L23:231 weakly excites
- L23:2394 weakly excites
- L25:1718 weakly suppresses
- L24:78 weakly suppresses
- L27:2696 weakly excites
- L23:663 weakly excites
- L25:1401 weakly suppresses
- L24:2775 weakly excites
- L24:2707 weakly suppresses
- L25:828 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
- lm
- 屈
- PT
- clouds
- (single
- INCLUDING
- 克斯
- .ib
- (filters
- representing
Suppresses
- ials
- atabase
- ernel
- ality
- 蛄
- adow
- chosen
- ethod
- занят
- charge