'of' inside institution names
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. 9 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 reacts to the word 'of' in names of organisations and titles, like 'Annals of Internal Medicine' or 'Ministry of Culture'. Turning it up made the model drift into unrelated, rambling 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.
-
in children (from allergic rh<|endoftext|>The study was carried out by researcher Rosa Binimelis of the UAB Institute of Environmental Science and Technology. Binimelis is working on the European project AL
-
every state. The findings are published in Annals of Internal Medicine.⏎Researchers from the University of Washington studied death records and Census data to
-
women aged 40 to 49.⏎It said the risks outweigh the small potential benefit in younger women, but a new study in the "Annals of Internal Medicine"
-
strategy 2014-2020, the Ministry of Infrastructure, Transports and Networks, the National Telecommunications and Post Commission and the Ministry of Education co-la<|endoftext|>Student or Learner⏎Brass: the persons (or
-
couture can be used to describe any garment that's handmade and one-of-a-kind, haute couture is a special designation created and approved by the French Ministry of Industry.⏎Nowadays, it is very easy to find information on any subject.
-
large number of local manifestations of the element are inscribed on the National List of Intangible Cultural Heritage administered by the Department of Intangible Cultural Heritage of the Ministry of Culture.⏎These videos (and many more) can also be consulted through the UNESCO
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 don't know what I am doing in the world of the movie industry." This is the first sentence of the dialogue between the characters in "The World of the Movies," the first sentence of the dialogue between the characters in "The
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
; let’s say we want to see what has changed between Commit 1 and Commit 3.⏎The diff will compare the differences between the two commits, showing what
-
North Pacific Gyre.⏎Twice the size of Texas – though some scientists have argued that its mass has extended to twice the
-
University of Hamburg in Germany, in collaboration with colleagues from the University of Aarhus in Denmark, have synthesized a new superconducting material by growing
-
and trade in services.⏎Content and Structure⏎The course consists of the following modules:⏎• Module I – The Concept of Trade Negotiations: An Overview⏎• Module II – General Introduction to WTO Negotiations⏎• Module III – Multilateral Trade
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
- L8:1009 suppresses
- L8:283 suppresses
- L8:320 excites
- L8:350 suppresses blank space before a year
- L6:351 excites
- L8:1544 excites
- L8:298 weakly suppresses
- L5:2835 weakly suppresses
- L4:1374 weakly suppresses
- L8:134 weakly excites
- L8:928 weakly suppresses
- L8:455 weakly suppresses
- L8:668 weakly suppresses
- L8:421 weakly excites
- L8:187 weakly suppresses
- L8:272 weakly excites
- L6:345 weakly excites
- L5:1372 weakly suppresses
- L8:99 weakly suppresses
- L1:1484 weakly suppresses
- L7:2046 weakly suppresses
- L8:1101 weakly suppresses
- L5:2869 weakly excites
- L7:805 weakly suppresses
Feeds later layers
- L11:264 excites
- L10:959 suppresses
- L15:487 weakly suppresses
- L10:152 weakly excites
- L12:2033 weakly excites
- L11:2393 weakly excites
- L16:1402 weakly suppresses
- L11:341 weakly excites
- L16:2280 weakly excites
- L23:510 weakly excites
- L10:217 weakly excites
- L19:1141 weakly excites
- L16:1635 weakly excites
- L24:1098 weakly excites
- L11:1381 weakly excites
- L12:795 weakly excites
- L17:163 weakly suppresses
- L18:2835 weakly excites first piece of a split name or term
- L18:2078 weakly excites
- L15:1601 weakly excites
- L12:129 weakly excites unclear: mixed, mostly everyday health and advice prose
- L11:219 weakly excites
- L10:452 weakly excites
- L24:2767 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
- cle
- orro
- -FIRST
- are
- ь
- atr
- 卡拉
- hart
- ча
- arra
Suppresses
- intage
- relent
- 粱
- links
- phasis
- iasi
- missive
- 甑
- redit
- Genetics