'on' after a noun like 'impact on'
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
This neuron responds to the word "on" when it follows a noun and introduces what that noun affects, such as "the impact on" or "the effect on". Turning it up gave repetitive, unrelated text.
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.
-
biological materials, specifically their ability to provide multiple functions despite severe limitations in the quality of available building blocks<|endoftext|>The OED “is widely regarded as the accepted authority on the English language. It is an unsurpassed guide to the meaning, history,
-
1961 (called the Principal Act) which⏎received the assent of the President on the 15th April 1962 and was published in the
-
natural disasters are essential to reduce loss of life and property by reducing the impact on populations both within the U.S. and abroad. The Federal Emergency Management Agency (
-
the Americas, especially in the movement of tobacco and sugar. The industries of Scotland produced cotton, coal and iron which were exported. Shipbuilding became a major industry on the Clyde, building many famous ships. By the end of the 19
-
More judicious prescribing of antibiotics for childhood respiratory infection has not increased the number of episodes of peritonsillar abscess or rheumatic fever. The effect on mastoidectomy is unclear, but a clinically important increase seems unlikely. (Level
-
and underscored the critical need for effective cooperation on border biosecurity measures.⏎Biosecurity is critical to protecting public health, ecosystems and
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: it is a question that is a question that is a Which one is correct? A) the A) the A) the A) the A) the A) the A) the A
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to a judge from the Salem Witch Trials, which was his Great-Great Grandfather John Hathorne. Hathorne was the only judge who did not express atonement for his crimes,
-
made atonement for the dead, so that they might be set free from their sin.”
-
It was declared, with a sigh, that most Indians would be removed due merely to the<|endoftext|>Editorial: Team effort needed to combat cyberbullying⏎Updated On: Sep 21 2012 11:49
-
.⏎The Hasinai group of the Caddo tribe built a village in the area about
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
- L1:1104 suppresses
- L1:906 suppresses
- L1:1194 suppresses
- L1:1 suppresses
- L2:1669 weakly excites
- L0:1280 weakly excites
- L2:1323 weakly suppresses unclear: fragments of rare nouns
- L1:2567 weakly excites
- L1:2296 weakly excites
- L1:2740 weakly suppresses word fragments inside non-English words
- L2:1942 weakly suppresses
- L2:2615 weakly suppresses
- L2:200 weakly suppresses
- L1:1179 weakly suppresses 'in' after 'result' (results in)
- L2:2552 weakly suppresses
- L2:691 weakly excites
- L2:2567 weakly excites
- L1:2014 weakly excites
- L2:2307 weakly excites
- L2:297 weakly suppresses
- L2:750 weakly excites
- L0:1660 weakly suppresses
- L1:2439 weakly excites
- L2:1215 weakly excites
Feeds later layers
- L17:303 excites
- L10:132 excites
- L17:385 excites
- L5:763 excites
- L18:293 excites
- L20:283 excites
- L19:1464 excites
- L17:2619 excites
- L4:16 suppresses
- L4:4 excites
- L16:2666 excites
- L27:168 excites
- L17:3068 excites comma after a sentence subject
- L13:132 excites "of" after a noun
- L6:613 excites
- L16:2120 excites
- L17:2302 excites 'in' after a noun
- L18:2908 suppresses
- L18:2887 excites
- L6:846 excites
- L22:489 excites
- L6:2379 weakly excites
- L4:1500 weakly excites
- L23:1185 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
- geb
- 的就是
- ervisor
- 年由
- HONE
- 任何一个
- Mastery
- ñana
- ⏎⏎⏎
- .heroku
Suppresses
- andid
- eselect
- filmer
- 带走
- .removeAll
- vice
- saç
- VIN
- idity
- wake