zeros inside long number codes
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. 21 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 lights up on digits inside long number codes such as ISBN numbers, postal codes, journal ISSNs and DOIs, especially the zeros. It goes quiet on ordinary words, and it goes quiet on the word 'Opt' in phrases like 'opt-in'. Turning it up made the model's output drift into unrelated text, so the steering results don't show a clear meaning.
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.
-
book is published by Red Fox, an imprint of Random House. (see www.kidsatrandomhouse.co.uk) ISBN 978-0-0
-
induced changes in gene flow patterns among Tasmanian devil populations, Biological Conservation, 165 pp. 69-78. ISSN 0006-320
-
provider:|| Postal: Katajanokanlaituri 6B, 00160 Helsinki|⏎Web page: http://www.wider.unu.edu
-
23(2), 1976 pp. 125–135. doi:10.1112/S0025579300008718.<|endoftext|>
-
Northwest, on the Touch for map. Marker is at or near this postal address: 3048 R Street Northwest, Washington DC 20007, United States of
-
from Wiktionary, Creative Commons Attribution/Share-Alike License⏎- n. An underwater explorer.⏎from WordNet 3.0 Copyright 2006 by Princeton University. All rights reserved.⏎- n. an underwater swimmer
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: if you have a square and you need to go to a point in the square, how do you get to that point in the first and second and third and fourth and ... etc. and so on. So, first, I think of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Meditation and Prayer raise our frequency by 15 MHz⏎20 Ways to Increase Your Frequency & Optimize Your Health⏎As certain patterns of vibration are associated with good health, one
-
to increase walking, bicycling and using public transportation. The plan will engage the public and partners across the region to identify the region’s principal active transportation network.⏎Opt-In poll for active transportation⏎Metro recently conducted an Opt-In poll on active transportation
-
I had been shooting whistlers (golden-eye ducks) over decoys.”⏎“The software can also automatically detect and remove red - and golden-eye defects.”⏎“OptiML Red-Eye uses embedded image processing automatically detects and corrects red-
-
conditioner's ability to remove heat and lower temperature<|endoftext|>Last month, the National Academies of Science, Engineering, and Medicine hosted a 3-day workshop: "Optimizing care systems for people with intellectual and developmental disabilities." During this workshop,
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
- L18:474 strongly suppresses
- L19:906 strongly excites
- L18:112 strongly suppresses 'over' meaning time or duration
- L10:1052 strongly suppresses
- L2:1795 strongly suppresses
- L5:1546 strongly suppresses
- L16:177 strongly excites zero digit inside dates and codes
- L17:213 strongly suppresses
- L0:44 suppresses
- L1:15 suppresses
- L0:1906 suppresses
- L17:106 excites
- L20:121 excites
- L17:1046 suppresses
- L5:138 suppresses
- L1:178 suppresses
- L4:498 suppresses
- L0:1241 excites
- L10:715 suppresses
- L1:1082 excites
- L7:2095 suppresses
- L7:1581 suppresses the last zero in large numbers
- L16:75 excites
- L2:657 suppresses
Feeds later layers
- L27:2618 suppresses
- L23:1773 suppresses
- L23:2981 suppresses
- L26:1712 suppresses single letter starting a split word
- L24:2973 suppresses
- L26:2206 weakly suppresses
- L26:2017 weakly suppresses
- L24:2537 weakly suppresses
- L25:2782 weakly suppresses
- L26:2888 weakly suppresses
- L24:2265 weakly excites
- L25:1060 weakly suppresses
- L27:3015 weakly suppresses 'F' starting a word in a split
- L23:1608 weakly suppresses
- L23:2515 weakly suppresses
- L23:1865 weakly suppresses
- L27:2187 weakly suppresses
- L23:2766 weakly suppresses
- L27:3054 weakly suppresses
- L26:1196 weakly suppresses
- L22:1892 weakly suppresses the digit 6 in years 19xx
- L25:2342 weakly suppresses
- L23:1366 weakly excites
- L24:2801 weakly suppresses
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
- 掌声
- ftware
- orks
- ости
- ��
- ость
- blank
- ulates
- dge
- olist
Suppresses
- entimes
- course
- lahoma
- posing
- portun
- imum
- erves
- ngoing
- 凿
- course