closing bracket after an acronym
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. 3 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
It fires on the closing bracket that ends a short name or abbreviation written in brackets, such as the ')' in '(WWF)' or '(WHO)'. Turning it up did not bring out this pattern in the sample outputs.
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.
-
.<|endoftext|>A random variable X has the cumulative distribution function⏎Calculate the variance of X.⏎The answer specifies that the density function is⏎I got the f(x) = x-1 part, and I got how to calculate the variance after you
-
is based⏎Two NGOs -- the World Wide Fund for Nature (WWF) and a consortium led by Greenpeace -- have put forward blueprints that could be
-
Treaties and Documents⏎1949 Conventions and Additional Protocols, and their Commentaries⏎Historical Treaties and Documents⏎Convention (III) relative to the Treatment of Prisoners of War. Geneva, 12 August
-
) 1999 RQ36.”⏎The asteroid was discovered in 1999 by the Lincoln Near Earth Asteroid Research (LINEAR) survey at MIT’s Lincoln Laboratory. LINEAR is part of NASA’s Near Earth
-
problems and other post-COVID complications.⏎The new method developed by KAUST — known as Deep-Lung Parenchyma-Enhancing (DLPE) — overlays
-
truth trials, but there’s a lot left to do for women,” said Julissa Mantilla Falcón, Interamerican Commission for Human Rights (CIDH) Commisioner, told The Herald in an interview. “Even after the trials
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 effect of a previous decision to be in the same time I have been thinking about lately... the effect of a previous decision to be in the same time I have been thinking about lately... the effect of a previous decision to be in the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
1989 (Aptian), G. crenulata (Matsumoto, 1938; Albian–lower Cenomanian) and G. maedai sp. nov. (middle to ?
-
drug in the world, but did its advent really change society in Britain and even the way we eat?⏎In 1965 something lurking under the meniscus of British society punctured the surface.⏎A man named
-
an optical show in the city. By analyzing this case, I will attempt to illustrate the visual novelty, the level of acceptance and then great popularity of the stereoscope compared to other entertainment in this period. Finally, I will present new data concerning
-
of the shock of jumps and landings and helps to distribute joint fluid evenly to lubricate and nourish the knee.⏎In the United States, a torn meniscus is the most common reason for knee surgery.⏎The symptoms of a torn men
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
- L1:341 suppresses
- L0:2146 weakly excites
- L2:359 weakly suppresses
- L1:86 weakly suppresses
- L0:870 weakly excites
- L0:1424 weakly excites
- L1:2535 weakly suppresses
- L2:70 weakly excites
- L1:1506 weakly suppresses
- L1:1384 weakly excites
- L1:1674 weakly excites
- L1:367 weakly suppresses
- L2:75 weakly suppresses
- L1:360 weakly excites
- L2:1320 weakly suppresses
- L0:70 weakly excites
- L2:455 weakly suppresses
- L2:1657 weakly excites
- L1:1029 weakly excites
- L0:2179 weakly suppresses
- L1:124 weakly suppresses
- L2:780 weakly excites
- L1:899 weakly suppresses 'which' after 'find out' or 'choose'
- L1:151 weakly excites
Feeds later layers
- L8:577 excites
- L11:293 excites
- L7:6 suppresses
- L15:316 suppresses
- L13:2620 weakly excites
- L11:55 weakly excites
- L14:756 weakly excites
- L4:1013 weakly suppresses
- L5:38 weakly suppresses
- L15:616 weakly excites closing parenthesis ending an aside
- L21:1024 weakly suppresses
- L10:860 weakly excites
- L13:685 weakly excites
- L8:1789 weakly excites
- L6:37 weakly suppresses
- L8:455 weakly excites
- L10:679 weakly excites
- L17:2516 weakly suppresses
- L17:382 weakly excites closing bracket after a fact in parentheses
- L14:2097 weakly excites
- L4:1394 weakly suppresses
- L12:187 weakly suppresses
- L9:29 weakly suppresses opening parenthesis before an acronym
- L14:2816 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
- reta
- aye
- sx
- eczy
- =wx
- gx
- ẫ
- maries
- ayne
- Raqqa
Suppresses
- ories
- toll
- Pipe
- 产出
- тро
- Product
- Install
- �
- Installer
- hog