opening parenthesis starting an aside
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. 13 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 an opening bracket that starts an explanation or example right after a noun or phrase. When it was turned up, the model's text got stuck repeating words instead of making sense.
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.
-
volume of the container it is in. Despite all of the constituent particles being charged, typically the plasma itself has no overall charge. However, some plasmas (non-neutral) can be created with an overall charge (either positive or negative)
-
carcinomatosis—are cancers that are frequently referred to as peritoneal cancers. Whether peritoneal cancer starts in the peritoneum or spreads from somewhere else (colon⏎Ovarian Epithelial, Fallopian Tube, and Primary
-
enjoy their colorful visuals, whimsical themes, and, in some cases, varied functions (for instance, pop-up, or moving-part valentines). Displayed singly
-
). This may be further exacerbated by few harmful chemicals (proteins) such as tumor necrosis factor alpha (TNF-α
-
a savings account.⏎Once the “long-term savings” envelope reaches a certain dollar amount (you can agree upon this amount with your child beforehand), it’s time to open
-
the plasma conversion step started with two pilot plants: the first, to test the gasification and plasma conversion; the second, to test the methanation process (where
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 day I was going to go to the sun (I'm going to go to the sun (I'm going to go to the sun (I'm going to go to the sun (I'm going to go to the sun (I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is Noether's theorem, which states that there is a one-to-one correspondence between each one of them and a differentiable symmetry of nature. For example,
-
of the operating principle of relationship between components. Also knows as a Schematic or Diagrammatic Drawing.⏎13. Perspective Drawing⏎Descriptive three-quarter view produced using a perspective drawing technique. Created using line only without the application of color or
-
System (DNS) is a global database that translates domain names (such as www.internetsociety.org) to Internet addresses that
-
amayani (Hindi:कामायनी) (1936) is a Hindi epic poem (Mahakavya) by Jaishankar Prasad (1889–1937
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
- L9:1414 suppresses opening parenthesis after copyright or year
- L5:330 excites opening parenthesis after a noun
- L7:110 weakly excites
- L9:1370 weakly suppresses
- L8:2184 weakly excites
- L11:2400 weakly excites
- L12:2654 weakly suppresses
- L10:270 weakly excites
- L6:1391 weakly suppresses
- L11:257 weakly excites
- L6:304 weakly excites
- L10:1756 weakly excites
- L12:1557 weakly suppresses comma after an introductory phrase
- L12:1248 weakly excites
- L12:878 weakly suppresses
- L7:25 weakly excites
- L8:81 weakly suppresses
- L5:2548 weakly excites
- L2:1982 weakly suppresses
- L4:2540 weakly excites
- L11:1855 weakly suppresses
- L2:2703 weakly suppresses
- L7:2239 weakly excites
- L10:166 weakly suppresses
Feeds later layers
- L14:1616 suppresses
- L16:1643 suppresses
- L16:530 suppresses
- L15:113 excites
- L15:2956 weakly excites
- L15:2222 weakly suppresses
- L16:1128 weakly suppresses
- L16:2093 weakly suppresses 'of' in 'piece of' phrases
- L17:1229 weakly excites
- L16:2705 weakly suppresses
- L15:1528 weakly excites
- L15:261 weakly excites
- L20:1012 weakly suppresses
- L18:3033 weakly excites
- L16:659 weakly excites
- L14:2797 weakly suppresses
- L15:497 weakly suppresses
- L15:1156 weakly excites
- L14:1928 weakly suppresses
- L16:1394 weakly suppresses
- L25:880 weakly excites comma after a cause or effect clause
- L17:1516 weakly suppresses
- L15:2928 weakly suppresses
- L14:2022 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
- inee
- den
- potentially
- possibly
- aghan
- chool
- 偓
- xm
- brace
- _MAY
Suppresses
- dern
- perimental
- /Application
- 迸
- olatile
- نتج
- .getFont
- .FONT
- 纯
- ,true