'and' or 'or' inside noun lists
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. 2 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 lights up on the word 'and' or 'or' when it joins a list of things, like 'surveillance cameras, access controls and secure data centres'. When it was turned up, the model's writing fell into loops of 'the' and 'and' and stopped making sense.
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.
-
such as spiders, heights, flying or crowded places, even when there is no. For example, you may know a
-
nature, processes and methods of science…“⏎This needs to be embedded within our teaching of the
-
, including surveillance cameras, access controls and secure data centres. These measures safeguard against physical threats like theft, fire, and
-
, health and wellbeing considerations have matured into technical building components like indoor air quality, temperature and humidity. These improvements come with their own financial upsides. MIT’s Real Estate
-
of plants and animals). Other plants, animals and/or insects may use them for habitat or food. When natural areas are lost or
-
Avocado plays an important role in women’s health. It helps in estrogen production and reproductive hormonal balance.⏎- Nutrients like folic acid, potassium, fiber and other minerals make it super food for good health.⏎Turmeric or golden spice is
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 student who used to be a bit of a long time ago. Now that is, I am a student who used to be a bit of a long time ago. Now that is, I am a student who used to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
points help you find the general shape of your graph so that you have a nice place to start.⏎To sketch the full parent graph of cotangent, follow these steps:⏎Find the vertical asymptotes so you can find the domain.⏎Because cotangent
-
by visiting ruins of sugar plantation sites and rock carvings.⏎Saint Croix is the largest of three main U.S. Virgin islands. This island is known for its historical sites, rainforest, beaches and hills. There are plantation villages of
-
not every native plant will thrive in every location. It's important to know your site conditions and the requirements of the native plants when making a selection
-
will inform you on how to perform the proper set-up of your AutoCAD settings. This set-up is crucial in order for you to produce a drawing that makes sense and is visually attractive.
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
- L1:1100 weakly suppresses
- L0:1190 weakly excites
- L1:686 weakly suppresses
- L0:2808 weakly suppresses
- L0:21 weakly suppresses last digit of years like 1970
- L1:304 weakly excites
- L0:839 weakly excites
- L1:108 weakly suppresses
- L0:2529 weakly excites
- L1:2992 weakly excites
- L1:494 weakly suppresses
- L0:988 weakly excites
- L0:861 weakly suppresses
- L1:2100 weakly excites
- L1:2119 weakly excites
- L0:2535 weakly excites
- L0:2834 weakly suppresses
- L0:126 weakly excites
- L0:1053 weakly suppresses
- L1:1355 weakly suppresses
- L1:149 weakly excites
- L1:387 weakly excites "to" after "it's important"
- L0:1613 weakly suppresses
- L0:1626 weakly suppresses
Feeds later layers
- L3:177 suppresses
- L4:121 suppresses
- L6:930 weakly suppresses middle digits inside numbers
- L8:347 weakly suppresses
- L9:1073 weakly suppresses
- L8:516 weakly suppresses
- L6:2438 weakly suppresses
- L11:627 weakly excites
- L15:2979 weakly suppresses
- L8:168 weakly suppresses
- L17:1712 weakly suppresses
- L13:1147 weakly excites
- L7:292 weakly excites
- L3:747 weakly suppresses
- L11:3037 weakly suppresses
- L6:237 weakly suppresses
- L21:641 weakly excites
- L8:870 weakly suppresses "Disease" in Centers for Disease Control
- L5:22 weakly suppresses
- L8:976 weakly suppresses
- L13:2145 weakly excites
- L16:1484 weakly excites
- L15:3012 weakly excites
- L16:814 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
- 'l
- '
- status
- ië
- '^
- ijkl
- .setStatus
- '|
- lul
- "},"
Suppresses
- -mouth
- trot
- 跑了
- -spacing
- -transitional
- (mapped
- transparent
- scape
- Convertible
- )*/⏎