stray bracket markers from web pages
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. 24 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 odd bracket symbols that scraped web pages leave behind, such as a "[edit]" or "[Update]" tag stuck in the middle of a sentence. It was weakest on copyright lines. When it was turned up, the model produced strings of parentheses or random text.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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.
-
is another temple which is located about 500 meters away from the Phnom Da temple in the southwest direction. This temple is made of sandstone.⏎[SinglePic not found]Every year thousands of local and international tourists visit the Ph
-
8, with the capitulation of Koguryŏ, was able, for the first time in history, to unify under its rule all the Korean peninsula.⏎[page 2 of the original]⏎Still today in Japanese the state of Paek
-
, international humanitarian law, computer data, targeting.⏎Facts[edit | edit source]⏎[F1] Neutrality International (NI), an
-
any other living scholar during the 1980–1992 time period. He was the eighth-most cited scholar in any time period.⏎Politics[change | change source]⏎In the 1960s he criticized the
-
Next when you are done.⏎[Update: 3/21/2009] If you are
-
eness in which people try to seek and give confirmation and support, and to reach consensus. Life is a community, a struggle to preserve intimacy and avoid isolation.⏎[If intimacy says, “We’re close and the same,” and independence says,
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: ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( (
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the water, were the major factors shaping phytoplankton communities.⏎- Abstract views: 1777⏎- PDF: 842⏎Copyright (c) 2009 Silvia TAVERNINI, Daniele
-
. 3, No. 8⏎Copyright © James B. Jordan 1991⏎Chronologies and Kings (
-
dietary habits, optimization of management of diabetes and cardiovascular diseases, and conducting continuous surveillance to monitor the trends in NCD prevalence, their risk factors, and treatments.⏎Copyright © 2020 The Royal Society for Public Health. Published by Else
-
16, this study identified that the impact of religious education on Taiwanese students' civic values was low, with the main focus being personal and spiritual life development. Copyright © 2021 selection and editorial matter, Kerry J. Kennedy and
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
- L19:1048 excites
- L3:207 suppresses
- L23:353 excites
- L19:251 excites
- L11:171 weakly excites
- L20:2699 weakly suppresses
- L17:1386 weakly excites
- L18:866 weakly suppresses
- L9:2579 weakly suppresses
- L4:38 weakly suppresses
- L21:2905 weakly excites
- L6:547 weakly excites
- L17:663 weakly suppresses
- L11:1909 weakly excites
- L18:240 weakly suppresses
- L7:1423 weakly suppresses
- L16:109 weakly suppresses
- L4:169 weakly excites
- L13:1818 weakly excites
- L14:43 weakly excites
- L16:1071 weakly excites "any" in questions and negatives
- L15:644 weakly suppresses
- L7:268 weakly excites
- L19:2919 weakly suppresses
Feeds later layers
- L26:13 excites
- L26:1374 suppresses
- L25:360 weakly suppresses
- L26:234 weakly suppresses
- L27:218 weakly suppresses
- L25:478 weakly excites
- L27:1007 weakly excites
- L26:1976 weakly excites
- L26:2383 weakly excites
- L25:1058 weakly suppresses
- L25:2737 weakly excites 'a' before 'few' or 'couple'
- L25:1792 weakly excites the modal word 'may'
- L27:1885 weakly suppresses
- L26:1999 weakly excites
- L25:179 weakly suppresses
- L25:1675 weakly excites
- L26:335 weakly suppresses
- L25:3050 weakly suppresses unclear: word pieces in medical terms
- L26:912 weakly suppresses
- L25:2959 weakly suppresses
- L26:1566 weakly excites
- L26:1677 weakly suppresses
- L26:1446 weakly suppresses
- L26:770 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
- uition
- ienza
- NECTION
- ám
- oise
- 诿
- bstract
- ßer
- aseline
- ourcing
Suppresses
- composer
- terior
- (subject
- fare
- (rep
- atters
- entai
- adar
- lator
- Passage