'of' after nouns like anniversary or onset
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 fires on 'of' right after a noun such as 'anniversary', 'onset', 'undersides', 'issue' or 'use'. It is a word that leads into the thing being described. Its low side is about phrases like 'get the most out' and 'during the life'. Turning it up gave repetitive or 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.
-
Communist Party⏎- 27 June 2011⏎- From the section Asia-Pacific⏎China is preparing to mark the 90th anniversary of the founding of the ruling Communist Party. The BBC's Michael Bristow in
-
-polluting nations such as<|endoftext|>Learning disabilities are neurologically based conditions that interfere with the acquisition, storage, organization, and use of skills and knowledge. They are identified by deficits in academic functioning and in processing other
-
may have had trouble finding them this spring.”⏎Disease symptoms include wilting, yellowing and bloom loss. Homeowners may notice a white growth on undersides of leaves, especially during humid conditions, rainy
-
midnight. Auroral observations by the Wideband Imaging Camera on the Imager for Magnetopause-to-Aurora Global Exploration spacecraft are used to determine the onset times of substorms. It is shown that the magnetic phenomena at the earthward site of
-
being processed by steam, a container is normally selected from one of the first containers going into the bottom layer of the first retort crate. This container is then set aside where it is not subject
-
other colonies, territories and states.⏎In late July of 2019, there will be a large celebration commemorating the 400th anniversary of the first legislative assembly held in North America when Governor Sir George Yeardley,
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: "The Power of the Ocean, the Power of the People of the Sea of Nations (POC) of the Future of Humanity" (POC) in the United States and the South Africa of the South. This is the time to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
up in order to get the most out of your battery.⏎Lithium-Ion Battery: They charge faster than other
-
selected the Lockheed Martin Corp. to build the Joint Strike Fighter (JSF). The contract had a potential total value of over $200 billion during the life of the program. The U.S. planned to buy nearly 3,0
-
. Bruce Weinstein, better known to some as The Ethics Guy, listed five of them in his book Ethical Intelligence:⏎Do No Harm⏎Make Things Better⏎
-
healthy eating as defined by community members themselves.⏎This study includes three of the most deprived electoral wards in Stoke-on-Trent. In each of these areas
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
- L8:2257 suppresses
- L8:502 excites
- L6:1735 excites
- L8:1003 suppresses "the" or "of" before a noun phrase
- L7:23 suppresses
- L8:447 excites
- L8:25 suppresses
- L8:853 suppresses slot right before a quantity
- L5:1721 excites
- L8:2095 suppresses
- L8:706 suppresses
- L6:10 excites
- L6:823 suppresses
- L8:892 weakly excites
- L8:9 weakly excites
- L8:433 weakly suppresses
- L5:2890 weakly excites
- L4:1686 weakly excites
- L3:1096 weakly suppresses
- L8:500 weakly suppresses 'to' after verbs like continue or help
- L6:1602 weakly excites unclear: hyphens and size comparisons
- L6:2217 weakly suppresses
- L8:913 weakly suppresses
- L8:879 weakly suppresses
Feeds later layers
- L10:584 suppresses
- L10:40 suppresses
- L20:1788 excites
- L13:2065 excites
- L13:471 weakly excites
- L11:219 weakly excites
- L10:67 weakly excites
- L15:487 weakly suppresses
- L13:376 weakly excites
- L21:2342 weakly excites
- L12:653 weakly excites
- L13:2469 weakly suppresses
- L21:1073 weakly excites
- L15:1811 weakly excites
- L12:74 weakly suppresses
- L15:1168 weakly excites
- L13:1062 weakly excites
- L17:2302 weakly excites 'in' after a noun
- L21:1866 weakly suppresses
- L12:1741 weakly excites
- L19:2926 weakly excites
- L12:1685 weakly suppresses
- L24:2657 weakly excites
- L11:1461 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
- large
- rebellion
- =com
- borough
- создан
- gibi
- себе
- lastName
- Rebel
- собой
Suppresses
- -piece
- iations
- inue
- Material
- ideos
- 跄
- 物
- AMENT
- inqu
- unist