the colon in web addresses
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. 12 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 '://' right after 'http' or 'https' in a web address, and on the '.gov' part of a site name. Turning it up made the model's output wander into unrelated text, so the steering does not clearly match this.
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.
-
Undersea Features) (Wunderman, R., ed.). Bulletin of the Global Volcanism Network, 32:1. Smithsonian Institution. https://doi.org/10.5479/si.GVP.BGV
-
Water Resources Division ;⏎Available from the Earth Science Information Center, Open-file Reports Section,<|endoftext|>Source: California Invasive Plant Council⏎URL of this page: http://www.cal-ipc.org/ip/management/plant_profiles/Rumex_acet
-
8-2-2002. Pro-poor Markets for Environmental Services Carbon Sequestration and Watershed Protection. Nature Conservancy. http://biodiversityeconom ics.org/incentives/0208
-
Research Library and Cultural Center, where a large solar canopy shades outdoor exercise equipment.⏎For more information about the FPL SolarNow program, visit this website: https://SolarNow.FPL.com/<|endoftext|>The proposed new social standards by The NC Department
-
available?⏎Dried chrysanthemum flowers can be found<|endoftext|>Copyright © University of Cambridge. All rights reserved.⏎'Trick or Treat' printed from http://nrich.maths.org/⏎Emily and Tristan from St James' School⏎answered
-
PDF […]⏎The World As It Could Be Curriculum and Resource Guide for Teachers & Organization Leaders is now available through the Education Resources Information Center (ERIC) at http://www.eric.ed.gov/ERICWebPortal/detail?accno=ED5
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 World Wide Web and the Internet of Things, which I have not seen me use as much as the rest of the world around me, I don't know what's missing." I have not seen me use it as much as the rest
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ography, July 2, 2014. DOI: 10.1007/s13524-014-0308-7<|endoftext|>Designs and simulation results are given for two
-
the true dimensions of the Larsen C iceberg, here's an animation based on real data, with something familiar for scale pic.twitter.com/OdQA07qhBb⏎— Adrian Luckman (@adrian_luckman) July
-
. As well as I will create a user who has full priviliges to access drupal database:⏎mysql> create database database_name;⏎mysql> CREATE USER user_name@localhost IDENTIFIED BY '
-
Dead<|endoftext|>Note: Supplemental materials are not guaranteed with Rental or Used book purchases.⏎- ISBN: 9781908049490 | 1908049499⏎- Cover
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
- L11:2558 suppresses
- L11:102 excites
- L8:687 excites
- L9:31 excites
- L9:280 suppresses
- L10:2362 suppresses
- L10:86 suppresses
- L10:59 suppresses
- L9:11 suppresses
- L11:3008 excites
- L11:2834 suppresses
- L9:1825 suppresses
- L9:1189 suppresses
- L10:1592 weakly suppresses
- L3:171 weakly excites
- L10:2975 weakly excites
- L9:151 weakly excites
- L11:189 weakly excites
- L8:2184 weakly suppresses
- L9:1931 weakly suppresses
- L0:1082 weakly suppresses
- L7:2760 weakly excites
- L1:2925 weakly suppresses
- L7:572 weakly suppresses
Feeds later layers
- L18:175 excites
- L13:532 excites
- L13:370 suppresses
- L14:1289 excites
- L16:825 excites
- L14:780 excites
- L15:524 excites
- L16:2579 excites
- L14:1136 excites
- L14:432 excites
- L13:1707 excites
- L13:2129 suppresses
- L16:1385 excites
- L22:374 excites
- L19:972 excites
- L13:2241 excites
- L17:280 excites
- L14:1972 excites
- L19:1102 excites
- L15:1962 excites
- L13:543 excites
- L20:806 excites
- L17:775 excites
- L15:2194 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
- etcode
- iasi
- urls
- ulp
- iat
- iated
- Urls
- fires
- exist
- 荚
Suppresses
- ))-
- elligence
- vers
- resher
- ]<<
- verting
- alo
- lombok
- แก
- 骀