'and' joining paired words
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. 20 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 'and' when it links two parallel words or actions, like 'trouble and strife' or 'read and write'. It goes strongly negative on 'but' and on 'and' in a list of steps. Turning it up made the model repeat 'and' many times.
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.
-
wonder how to subtract mixed numbers? Do you know how to handle unlike denominators or how to regroup or to borrow with mixed numbers? Take a deep breath and let us take it one step at a time. Click and begin.⏎GCF
-
are nearby? What type of boundary are the quakes on?⏎Your paper should meet the following requirements:⏎3-4 pages in length (does not include title and reference page)⏎1-2 outside sources⏎Why Work with CheggAnswers⏎
-
⏎Screw Conveyors designs and manufacturer in Coimbatore⏎Screw Conveyors designs and manufacturer in Kerala⏎Screw Conveyors designs and manufacturer in Tamil Nadu⏎Screw Conveyors suppliers in Coimbatore
-
here is what I’ve learned along the way:⏎1. Encourage discourse.⏎It may feel a little contrived, but I’ve found that providing discourse stems and sample questions helps students figure out how to leave and respond to comments.⏎2.
-
flinching, and insists on pursuing such joy as he can find.⏎Resilience is his middle name.⏎Here circumstances are presented as clouds that help trouble and strife to germinate and proliferate.⏎‘This is the world, and we
-
illing’ advisedly because he says it really brought home to him just how fragile the Earth is. “It really brought it home to us that you cannot go and live on Mars, you can’t go and live on Titan, you can’t
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 so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so so
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
instance, if you are thinking of the number 1, it would be "A". Number 2 would be "B". 3 is "C", and so on.⏎- Now, think of a country that starts with the letter you
-
ternut Squash can do for you⏎It’s a great hydrator. One serving of butternut squash is roughly 87 percent water, which can help keep you
-
and is known only to the Butonian people on the Pulau Makassar. The study used the symbolic interaction
-
grams of dietary fiber⏎- 3 grams of sugar⏎- 1.4 grams of protein⏎- 6 milligrams of sodium⏎One serving of butternut squash for vitamins and nutrients is filled with:⏎- Over 10
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
- L13:131 strongly excites 'and' joining parallel phrases
- L16:376 suppresses 'or' joining alternatives
- L4:57 suppresses
- L9:92 suppresses "and" or "or" joining two items
- L10:74 excites
- L18:868 suppresses 'and' joining two nouns in a pair
- L15:727 excites
- L19:2080 suppresses ampersand joining two items
- L1:1250 suppresses
- L0:313 excites "and" joining paired nouns
- L18:2345 suppresses
- L19:547 weakly suppresses
- L4:108 weakly excites "or" offering a choice in web text
- L9:2010 weakly suppresses
- L17:572 weakly excites
- L12:51 weakly excites
- L4:659 weakly excites
- L0:1656 weakly suppresses
- L14:154 weakly suppresses
- L11:406 weakly suppresses
- L18:1107 weakly excites
- L5:548 weakly excites
- L17:2113 weakly excites
- L17:1637 weakly excites
Feeds later layers
- L26:347 excites the 're' in self-reliant and similar
- L22:632 excites
- L25:6 suppresses
- L24:205 excites
- L21:478 excites
- L25:1694 suppresses 'and' joining paired names or fields
- L27:568 excites
- L23:379 excites
- L22:198 suppresses
- L27:289 excites
- L21:2949 suppresses comma closing 'the more X, the more Y' clauses
- L23:2015 suppresses
- L26:693 suppresses
- L26:65 suppresses
- L25:120 suppresses
- L24:398 suppresses
- L25:1008 suppresses
- L21:739 weakly excites 'both X and Y' pairing words
- L25:247 weakly suppresses
- L21:191 weakly excites
- L22:689 weakly suppresses comma closing an intro clause
- L26:1232 weakly excites
- L26:1007 weakly suppresses
- L27:948 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
- but
- 性和
- 但
- But
- but
- 但现在
- 但他
- 但他们
- 和支持
- 但仍
Suppresses
- vice
- /or
- IGIN
- vice
- /OR
- rog
- amp
- 蔼
- rew
- yne