"which" after a noun of "the X of which"
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 fires on the word "which" when it comes after a phrase like "the origin of" or "the building of", where it starts a relative clause. Turning it up made the text wander into repeated, unclear phrasing.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
as Āmū, Chāch, Xīwa, Samarqand, Chaghāniyān and some others, the origin of which till now remains unknown. Name of Buxārā, which is
-
viruses cannot, in general, spread via Web site access.⏎Of course, the above remarks do not always apply. By relaxing the security settings (the details of which depend on both operating system and browser), it is possible to create an environment in
-
. So, if you are relying on government, you are in fact relying on your fellow countrymen who work hard each day to earn an income the portion of which would be taken ‘without prior consent’ by the government which would in turn use
-
their satellites.⏎In the face of these multiplying instances of treaty-breaking, of violence, and of open warfare, the United States followed a policy the successive stages of which are summed up here.⏎During the years preceding the outbreak of war in Europe in
-
the long-awaited and often-troubled advent of the railroad in the state, the building of which shifted from private to public efforts and back again, amid tumultuous social, business
-
supported by Tannen’s research and my personal observation and experience, is that there is a difference in communication style and objective between the genders – the understanding of which can aid our effective partnership in the kingdom. This difference in style is a reflection
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 am a person who is a good and honest person, who is always ready for the future, who is always ready for the future, who is a good and honest person, who is always ready for the future, who is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in clinical settings (note, there are now phase II clinica trials of a small molecule activator of cardiac myosin for congestive heart failure;Malik et al., Cardiac Myosin Activation: A Potential Therapeutic Approach for
-
66 million years ago - best-known for the extinction of the dinosaurs, the Cretaceous-Paleogene extinction event killed off three-quarters of all species. It's
-
historic mammal hair discovered in Cretaceous amberJune 16th, 2010 - 1
-
the journal Palaeontology.⏎The South Pyrenean area is rich in dinosaur egg sites, most of which communicate to sauropod eggs from the upper Cretaceous, dating back more than 70 million years ago.⏎During that period
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
- L0:1224 weakly suppresses
- L2:2692 weakly excites
- L0:1301 weakly suppresses
- L0:178 weakly suppresses
- L1:2305 weakly excites
- L2:70 weakly excites
- L0:1 weakly suppresses
- L1:1464 weakly excites
- L2:49 weakly suppresses
- L2:2328 weakly suppresses
- L1:1321 weakly excites
- L1:2910 weakly suppresses
- L0:1589 weakly suppresses
- L1:1630 weakly suppresses
- L1:2452 weakly suppresses
- L1:2683 weakly excites
- L2:1036 weakly excites
- L2:1829 weakly suppresses
- L2:2583 weakly excites
- L1:168 weakly suppresses
- L1:635 weakly excites 'by' after a past participle
- L1:1841 weakly suppresses
- L1:2444 weakly suppresses
- L2:92 weakly excites
Feeds later layers
- L4:2517 excites
- L7:1103 weakly suppresses 'as' in 'not as ... as' comparisons
- L4:117 weakly excites
- L5:2103 weakly excites
- L5:2417 weakly excites
- L12:2103 weakly suppresses
- L16:2390 weakly excites
- L16:2213 weakly excites
- L6:994 weakly excites
- L16:302 weakly suppresses
- L22:1962 weakly excites
- L15:1319 weakly suppresses
- L15:2309 weakly suppresses
- L23:2050 weakly excites
- L6:1751 weakly suppresses
- L11:65 weakly suppresses
- L7:1129 weakly suppresses
- L9:2402 weakly excites
- L25:1815 weakly excites
- L13:159 weakly suppresses
- L18:1803 weakly suppresses
- L15:721 weakly excites
- L4:1523 weakly excites
- L6:1240 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
- 在生活中
- sometimes
- urch
- иров
- erm
- Lambert
- osoph
- iscrim
- positories
- happen
Suppresses
- чная
- vens
- 深交
- Kun
- 珠江
- ocht
- 下一
- Hv
- rrha
- plans