year digits in copyright and 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. 8 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 responds to the digits of years, especially in copyright lines and dates such as "Copyright 2014" or "Last edited 2020". Turning it up made the model write repeated lines about years and numbers, which fits the pattern.
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.
-
may be able to formulate alternatives that may satisfy those interests.⏎The Questioning Process During Bargaining Interactions By Charles B. Craver⏎Copyright © 2014 Charles B. Craver⏎Copyright © 2014 The
-
San Antonio vied with each other for trade, commerce and business.⏎George R. Woolfolk⏎Prairie View A&M College⏎Rights© 2021 Prairie View A & M University⏎This work is licensed under a Creative
-
Delivery: Yes⏎- Store Delivery: Yes⏎Prices are for internet purchases only. Prices and availability in WHSmith Stores may vary significantly⏎© Copyright 2013 - 2017 WHSmith and its suppliers.⏎WHSmith
-
⏎Publication Type: Peer Reviewed Journal⏎Publication Acceptance Date: February 14, 2008⏎Publication Date: May 16, 2008⏎Citation: Davis, L.M., Spackman
-
under general anesthesia in<|endoftext|>BBB - A+ Rating⏎- the best there is⏎An excerpt from www.HouseOfNames.com archives copyright © 2000 - 2014⏎Origins Available: English, Irish
-
friendly Cite/link Email Feedback|⏎|Title Annotation:||British Empire|⏎|Article Type:||Brief Article|⏎|Date:||Jan 24, 20
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 most common way of dealing with a problem is to use a mirror, but even if you can, it's still the problem of using a mirror. It's the same as a child's first move. But why? Well, the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ezers or a special tick-removing tool. Grasp the tick as close to the skin as possible. If you don’t have tweezers, protect your fingers with a tissue.⏎- Pull the<|endoftext|>Listen to radar echoes from satellites
-
nests, but the Son of Man has no place to rest."⏎American Standard Version⏎And Jesus said unto him, The foxes have holes, and the birds of the heaven have nests; but the Son of man hath not where to lay his
-
child and/or the apprehension of a suspect.<|endoftext|>In 1906, Italian economist Vilfredo Pareto made the famous observation that 20% of the population owned 80% of the property in Italy, later
-
to use caution at all times. These cats’ stealth and adaptability allow them to survive well in a wide environment.⏎They’ve been known to jump 20 feet in one leap. The puma creeps up on its prey until it
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
- L6:932 excites
- L4:1291 excites second digit of years 20xx
- L1:1675 excites
- L1:1862 weakly excites
- L0:1529 weakly excites
- L5:236 weakly suppresses
- L4:973 weakly suppresses
- L5:2476 weakly excites
- L2:1334 weakly suppresses
- L0:329 weakly suppresses the digit 1 in years
- L2:2289 weakly suppresses
- L4:318 weakly excites the digit 9 in years 1900s
- L6:1747 weakly excites technical word fragments, splicing and isotopes
- L1:3021 weakly suppresses
- L5:881 weakly suppresses
- L5:871 weakly suppresses
- L5:60 weakly excites
- L0:982 weakly excites
- L1:1960 weakly excites
- L5:1207 weakly suppresses 'of' after abstract nouns
- L5:1051 weakly excites
- L7:3031 weakly excites
- L3:2549 weakly suppresses
- L0:2572 weakly excites
Feeds later layers
- L18:1551 excites middle digit of years in the 2000s and 2010s
- L13:2285 excites
- L18:2542 excites
- L9:349 excites the spot before a year
- L18:1865 excites
- L18:2103 excites
- L12:1264 excites gap before a year in dates
- L10:220 excites hyphen inside compound words
- L14:1700 excites
- L17:1476 excites
- L14:1253 excites
- L10:152 excites
- L10:456 weakly excites
- L13:2971 weakly excites
- L16:256 weakly excites number right after 'age of' or 'ages'
- L22:1279 weakly excites
- L14:395 weakly excites
- L10:1784 weakly excites
- L9:466 weakly excites
- L17:126 weakly excites
- L11:300 weakly excites
- L12:2691 weakly suppresses
- L9:811 weakly excites
- L15:368 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
- Sept
- sept
- go
- ника
- lets
- /tinyos
- ikt
- Sept
- 机能
- yleft
Suppresses
- ancode
- tres
- verständlich
- mant
- weil
- earer
- urname
- egen
- ultipart
- .addField