the digit 3 in day-of-month dates
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. 25 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 '3' in dates like 'June 30' or 'Feb. 3', which is the first digit of a day of the month. It is turned off by digits that start ages like '17' or '18'. Turning it up made the model talk about 'points' and repeat itself, and turning it down made it repeat the digit 8.
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.
-
atorsA student-built pollinator garden is one of Muhlenberg College’s newest sustainability initiatives.⏎By: Grace Oddo '22 Wednesday, June 30, 2021 02:03 PM⏎Professor
-
report go to the NOAA news website www.noaanews.noaa.gov⏎– Erich Luening⏎Print this page<|endoftext|>Today is Sunday, Feb. 3, the 34th day of 2013. There are
-
part of a series of features highlighting freshman seminars offered at Princeton this semester. Other features:⏎Exploring the science and nuance of facial perception⏎Posted October 31, 2011; 12:00 p.m
-
stakeholder groups.⏎Disseminate results to stakeholders and seek funding to develop a technology-based prototype⏎July 1, 2008 - June 30, 200
-
lower-income or urban schools on average already have a higher body mass index (BMI) and consume more soda. Another new study by Massachusetts General Hospital released June 3, indicates the pattern of fast food and obesity starts before they even begin school.⏎
-
- proper n. The northern parts of Scandinavia and Finland with the Kola Peninsula in Russia, traditionally inhabited by the Sami people.⏎from WordNet 3.0 Copyright 2006 by Princeton University
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:777688777777777777777777777777777777777777777777
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Illinois Senate.Biography Roberts was born August 20, 1866 in Decatur, Michigan. He graduated from high school at 17 and became a teacher. He then chose to take Ph.D. coursework at
-
applauds the action by the FDA to monitor e-cigarettes, cigars, hookah tobacco and pipe tobacco, including barring the sales of the products to teens under 18 years old. This is an important step in regulating all products containing tobacco and
-
: a minimum age of 16 for a learner's permit; a waiting period of at least six months before
-
-fold, according to a study by researchers at RTI International and the Florida Department of Health.⏎The study, published in Pediatrics, found that youth (age 12-17) exposure to electronic cigarette advertisements increased by 256
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
- L22:332 suppresses
- L13:33 excites
- L15:12 weakly excites
- L7:229 weakly excites first digit of a split number
- L19:258 weakly suppresses
- L8:2009 weakly suppresses
- L4:448 weakly excites
- L3:651 weakly excites zero in dates inside URLs
- L22:517 weakly excites
- L20:18 weakly excites
- L19:2148 weakly excites
- L21:1732 weakly suppresses
- L11:1852 weakly excites the digit 1 in numbers
- L8:20 weakly suppresses
- L23:2230 weakly suppresses
- L7:8 weakly suppresses
- L22:2783 weakly excites
- L23:2924 weakly suppresses
- L23:2135 weakly excites
- L23:2338 weakly suppresses
- L24:371 weakly suppresses
- L13:2570 weakly excites
- L12:1080 weakly excites
- L23:600 weakly suppresses
Feeds later layers
- L27:132 excites
- L27:1196 suppresses
- L26:905 suppresses
- L27:678 suppresses
- L26:580 suppresses
- L26:848 weakly excites
- L27:205 weakly excites
- L27:391 weakly suppresses
- L27:763 weakly suppresses
- L27:496 weakly suppresses
- L27:1202 weakly excites
- L27:435 weakly suppresses
- L27:153 weakly suppresses the spot just before a historical number
- L27:1022 weakly excites
- L27:1100 weakly suppresses
- L26:2551 weakly excites
- L27:1153 weakly excites
- L26:2809 weakly excites
- L27:992 weakly suppresses
- L26:103 weakly suppresses blank space before a number
- L27:241 weakly excites
- L27:352 weakly suppresses
- L27:238 weakly suppresses
- L26:667 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
- ändig
- addElement
- ioned
- Cad
- @admin
- .outer
- fried
- eldorf
- 蛘
- airie
Suppresses
- 6
- 7
- 8
- 4
- 5
- 7
- 6
- 8
- 4
- 5