opening parenthesis before an abbreviation
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. 17 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
It fires on the open bracket that comes right after a full name or phrase, just before its short form, like the "(" in "Multiple sclerosis (MS)". It goes quiet on the letters inside an acronym. Turning it up made the model drift into odd lists of capital letters like "A. A. A.", so the link to the trigger is only loosely visible in the output.
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.
-
ant drugs together on rates of venous thromboembolism (a blood clot in the veins) and stroke (both consequences of under-ant
-
for radioisotope thermoelectric generator (RTG) assembly operations, focusing on: (1) where DOE plans to
-
concern surrounding these chemicals is the hazardous substances contained in them. Alkylphenol ethoxylates (APEs), corrosive substances and volatile organic compounds (VOCs) top the list.⏎Alkylphenol ethoxyl
-
a person encounters biotoxins, such as mold, fungi, volatile organic compounds (VOCs), particulates, inflammagens, and other gram-negative bacteria,
-
Mina MJ, Leng Y, Elledge SJ, Niebuhr DW, Scher AI, Munger KL and Ascherio A⏎Multiple sclerosis (MS) is a chronic inflammatory demyelinating disease of the central nervous system
-
high-frequency signals, describing anything from AM radio between 535 kilohertz (kHz) and 1605 kHz to computer local area networks (LANs) at 2.4 gigahertz (GHz). However,
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 to ICI for I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in I in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
drought responses of these two tree species to a long-term and repeated drought in monoculture and mixed plantings, the “Kranzberg Roof Project” (KROOF, kroof.wzw.tum.de) was initiated in 2
-
boards: the Woman's Board of Missions, located in Boston (WBM), the Woman's Board of Missions of the Interior, located in Chicago (WBMI), the Woman's Board of Missions for the Pacific, located in San
-
700.⏎Graphics by Lynn Waterman, ©2004<|endoftext|>ABOUT THE AMERICAN SOCIETY OF MECHANICAL ENGINEERS⏎WHAT IS ASME⏎Founded in 1880 as
-
The excavations being carried out at Khor Rori in Salalah [Credit: Times of Oman]⏎"A team of researchers from Kerala Council for Historical Research (KCHR) is set to join the team involved in the excavations at Khor
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
- L12:1340 excites
- L15:2982 excites
- L14:2419 excites
- L11:651 suppresses
- L16:722 suppresses
- L16:2770 excites
- L9:2517 suppresses
- L7:765 suppresses
- L6:1912 excites
- L14:1972 excites
- L11:1635 suppresses
- L13:1945 excites
- L11:258 suppresses
- L8:1796 excites
- L16:2560 suppresses
- L14:3067 suppresses
- L10:268 excites
- L15:687 suppresses
- L0:2077 excites
- L6:1478 suppresses
- L6:7 suppresses
- L13:2576 weakly suppresses
- L11:1639 weakly excites
- L14:954 weakly suppresses
Feeds later layers
- L19:1829 suppresses opening bracket before an acronym
- L18:1803 suppresses
- L20:1080 suppresses opening paren before an acronym
- L20:2479 suppresses
- L21:2345 suppresses
- L18:609 suppresses
- L18:170 excites
- L18:1540 suppresses
- L21:857 suppresses
- L21:144 weakly excites
- L19:813 weakly suppresses
- L22:823 weakly excites
- L18:1448 weakly suppresses
- L18:496 weakly suppresses
- L22:110 weakly excites
- L23:853 weakly excites
- L23:1657 weakly excites
- L19:822 weakly excites
- L20:1306 weakly suppresses
- L18:1697 weakly excites
- L20:109 weakly suppresses
- L25:1459 weakly suppresses
- L20:405 weakly suppresses
- L18:695 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
- ось
- cents
- zerbai
- spo
- LAY
- atos
- ungs
- ющихся
- уж
- Dag
Suppresses
- braco
- MMC
- arest
- springfox
- GetHashCode
- aced
- Ljava
- acing
- uce
- grass