comma after a clause subject
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. 2 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 a comma that comes right after the subject of a sentence, just before the rest of the sentence continues, such as 'When a recycled material, rather than...'. When it was turned up, the model's text did not visibly change in a clear way.
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.
-
adeption.⏎After Edward IV’s death, Elizabeth tried to mediate between the factions on the king’s council. At the news that her brother and her son, Richard Grey, had been arrested by Richard Duke of Gloucester, she at first
-
little isle of Oronsay, close by Islay; and here for a moment they thought their future abode was to be. But when Columba, with the early morning, ascending the highest ground on the island, to take what
-
economic benefit?⏎A: In a broad sense, recycling is part of an ethic of resource efficiency – of using products to their fullest potential. When a recycled material, rather than a raw<|endoftext|>Toss it. Pitch it. Dump it. Chuck
-
0, and it won’t arrive at the Jezero Crater until February 18, 2021. But if the helicopter soars, it’ll provide never-before-seen views of the Red Planet.⏎Popayán
-
rich blend of cultural and traditional customs. All Americans are renowned for their openness and friendliness to visitors. As a rule, dress is casual; although smart restaurants, hotels and clubs insist on suits and ties or long dresses. Smoking is becoming increasingly
-
academic disciplines that contribute to HCI research. The second part of the book focuses on the application of HCI to Information Systems research, and reviews ways in which HCI techniques, methodologies and other research components have been used to date in the IS field. The
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: "I have a lot of time, and the last I have been thinking about lately. This is a fun way to learn about, and the last I have been thinking about lately. I have been thinking about the times and the last I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
focuses on all delinquent youth in the community, while the second examines only delinquent youth placed in juvenile pretrial detention centers when
-
Trail and ‘The Chief’ viewing platforms.⏎Take a virtual tour of the Sea-to-Sky viewing platform…⏎To find out more about the Britannia Mining Museum please visit: http://www.britanniaminemuseum.ca/⏎To find out
-
to create the next-generation processor, while maintaining backward compatibility and ease of replication; that is, Intel had to be able
-
⏎MRI scans can detect very small irregularities and produce a very high quality picture on its own, while the operator may also give the patient a contrast agent that can make the image even
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
- L1:546 excites
- L1:171 suppresses "but" after a negated clause
- L1:383 weakly suppresses
- L1:105 weakly excites
- L1:505 weakly excites
- L1:1231 weakly suppresses
- L1:89 weakly excites
- L1:211 weakly excites
- L0:27 weakly suppresses
- L1:125 weakly excites
- L1:644 weakly suppresses
- L1:807 weakly excites
- L0:2877 weakly suppresses
- L1:698 weakly suppresses
- L1:1149 weakly suppresses
- L1:453 weakly excites
- L1:603 weakly excites
- L1:694 weakly excites
- L0:2582 weakly suppresses
- L1:1344 weakly suppresses 'who' after a person noun
- L1:1001 weakly excites
- L1:1715 weakly suppresses
- L1:2677 weakly suppresses
- L0:585 weakly suppresses
Feeds later layers
- L3:182 excites
- L4:80 suppresses
- L3:227 suppresses
- L3:524 excites
- L4:44 suppresses
- L3:47 suppresses
- L3:569 excites unclear: mid-phrase function words
- L5:420 suppresses
- L6:352 suppresses
- L4:69 weakly suppresses
- L4:317 weakly excites
- L3:518 weakly excites
- L4:59 weakly suppresses
- L24:2951 weakly excites
- L4:305 weakly suppresses
- L22:1702 weakly excites
- L5:1598 weakly suppresses
- L15:157 weakly suppresses
- L11:1285 weakly suppresses
- L17:27 weakly suppresses
- L11:2917 weakly excites
- L8:2210 weakly suppresses
- L4:539 weakly suppresses
- L20:82 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
- 量
- ",
- imony
- tin
- )":
- ibrate
- TRA
- ivities
- いで
- طا
Suppresses
- MPL
- ukkit
- jp
- _PHP
- 길
- /slick
- =path
- ilenames
- ingular
- ActiveSupport