comma after a sentence 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. 17 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
The neuron fires on a comma that comes right after the subject of a sentence, such as a person or group name, just before an extra phrase like 'Rama, along with his brother'. Its negative side is a small 'by' or 'for' in phrases like 'by far'. The steering outputs did not show a clear theme.
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.
-
Hebrides that Columba turned his face when he accepted the terrible penance of Molaise. He bade farewell to his relatives, and, with a few monks who insisted on accompany him whithersoever he might
-
built except the roof," Thomas replied. So Gundaphorus sent him gold and silver to roof the palace, and Thomas, thanking<|endoftext|>The Internet and copyright⏎Choose a topic:⏎Internet technology is developing faster
-
killed the ten-headed demon Ravana, in Satyug. Ravan had abducted Rama’s wife Sita. Rama, along with his brother Lakshmana, follower Hanuman and an army of monkeys
-
Biblical law, Herod did not want to offend the masses (or the Temple priests) by placing his portrait on the kingdom’s coins. Instead, Herod, like his Hasmonean predecessors, decorated his coins with depictions of common
-
world.⏎The GIUK Gap forms the principal choke point between Russia’s great Northern Fleet and its strategic interests in the North Atlantic and all points south. The Russians, as a resurgent power, have modernized their military forces, but they still
-
om-Tzion had recently died and her sons, Hyrcanus II and Aristobulus II, divided against each other in a terrible civil war.⏎The Romans, as usual, were in the midst of fighting a war. In 64
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: that, that, that, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
at low wind strengths:⏎The DualWingGenerator is self-optimising and can adapt itself to different wind conditions. In terms of its efficiency, it is by no means inferior to conventional small wind power stations and features amazing benefits
-
,190 km2 (459 sq mi) with a basal circumference of 140 km. This makes it by far the largest of the three active volcanoes in Italy, being about two and
-
rit for Sickness reliever), Azadirachta indica (Neem), belonging to the Meliaceae family, from prehistoric Indian subcontinent, for its multitudinous medicinal properties (Tanguturi et al. 2
-
and set the paragraph to be double-spaced. Do not simply hit the Enter button twice at the end of each line, it will become an editing nightmare and when the sorting is finished, it will put all of the blank spaces at the top
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
- L10:470 excites
- L16:1223 excites
- L15:2971 suppresses
- L16:1442 suppresses
- L7:1098 excites
- L15:2835 excites
- L14:2402 suppresses
- L11:237 suppresses
- L13:2366 suppresses
- L13:921 excites
- L16:1115 suppresses
- L16:1621 excites
- L15:3052 suppresses
- L14:728 suppresses
- L14:2416 suppresses
- L3:232 excites 'on' after a noun like 'impact on'
- L15:804 suppresses
- L13:1232 excites
- L15:824 suppresses
- L7:890 excites
- L15:280 excites
- L16:1556 excites
- L10:1855 suppresses
- L5:2747 weakly excites
Feeds later layers
- L19:2707 excites
- L20:458 excites
- L18:1401 suppresses
- L24:148 excites
- L25:1596 excites
- L25:620 excites
- L19:2403 excites comma after a full clause
- L20:2020 excites comma before a participle phrase
- L24:140 excites
- L19:3037 weakly excites
- L26:252 weakly excites
- L21:2114 weakly excites comma before an added phrase
- L22:1425 weakly excites
- L26:686 weakly excites
- L20:349 weakly excites
- L19:1995 weakly suppresses comma after a subject phrase
- L19:2498 weakly excites
- L27:1237 weakly excites
- L19:1463 weakly excites
- L20:586 weakly suppresses
- L18:843 weakly suppresses
- L19:2535 weakly excites
- L18:1881 weakly excites
- L19:818 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
- reflecting
- minating
- ''' ⏎
- 簃
- ''' ⏎
- ') ⏎
- apon
- ując
- ifr
- erot
Suppresses
- 控
- сел
- ц
- 腾
- lical
- 访
- stand
- 引
- 插
- colors