'and' joining two parallel subjects
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. 14 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 matching things, like two countries or two groups of people. The strongest negative hits are on commas in numbers and a few unrelated words. Turning it up gave repetitive or unclear text.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
Mozambique in 2008 and in Ghana, Nepal and Lao PDR in 2009. The pilots in Lao PDR and Nepal specifically focused on the information and communication needs of Indigenous Peoples and were implemented by
-
to seven hours a week for science, technology, history, geography and creative arts; and two to three hours a week for sport and other physical activities.⏎English and maths, which dominate the proposed mandatory outcomes, should make up 45 to
-
using a nonlinguistic tone anchoring (adaptation) task and a linguistic selective adaptation task in children and adults with and without dyslexia. Children and adults with dyslexia had elevated tone-frequency discrimination thresholds, but both groups benefited
-
the FCAT?⏎The FCAT includes multiple-choice, gridded-response (fill in the blanks) and performance tasks (such as essays). The multiple-choice and gridded-response questions are machine scored. Each performance task test is scored by
-
, California has a coalition supporting its statewide GND platform, and Illinois and others have been starting to build their support.⏎Some U.S. cities, like Providence and Portland, have also passed climate
-
new footbath is not suitable for treatment of active painful lesions, but as an effective measure to prevent the spread of infection from cow to cow.⏎Sheep and cattle are prone to bacterial hoof infections that can be debilitating. A common condition in
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 problem. I don't know, the problem, the one that we know, the one that, I don't know. the problem, the one that, the one that, the one that, the one that, the one,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Curtiss and purchase a plane for $7,000 ($157,000 in today’s money). While waiting for it to be built,
-
9,490. Hourly rates range from a low of $13.50 to a high of $49.05 with a median hourly wage estimated at $23.80. Most mediators are employed
-
generally much weaker than tornadic spouts, and tend to be more short-lived. Tornadic waterspouts start in the cloud layer and extend downwards but fair weather waterspouts start at water level and extend upwards. Fair weather watersp
-
more on the type of calories you are consuming, than on the amount of calories you are consuming.⏎2. Choosing only low-fat or diet food⏎Processed, low-fat, or “diet” foods are often considered to be the ideal choice when
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
- L13:2451 excites
- L12:105 excites
- L13:1363 excites
- L13:1103 excites
- L13:619 suppresses
- L13:2908 suppresses
- L13:116 suppresses
- L13:2662 suppresses
- L13:300 suppresses
- L11:1251 excites
- L11:558 excites
- L9:454 excites
- L1:1464 suppresses
- L8:1175 weakly suppresses
- L10:782 weakly suppresses
- L13:651 weakly suppresses
- L10:3052 weakly suppresses 'the' after 'indicate' or 'is'
- L11:644 weakly suppresses
- L11:379 weakly suppresses
- L13:1562 weakly excites unclear: words that set up a qualifier
- L13:2540 weakly suppresses
- L12:1514 weakly suppresses
- L11:1094 weakly suppresses unclear: scattered web and commerce text
- L2:2280 weakly suppresses
Feeds later layers
- L15:1209 strongly suppresses
- L16:2935 suppresses
- L15:2334 suppresses
- L15:1492 excites
- L15:669 excites
- L16:473 suppresses
- L15:1067 excites
- L15:346 suppresses
- L15:752 suppresses
- L15:2050 excites
- L15:185 excites
- L18:2918 suppresses
- L15:772 excites
- L16:669 suppresses
- L15:2071 excites unclear: verbs before an object
- L15:1080 suppresses
- L15:2328 suppresses
- L15:474 excites
- L15:1917 suppresses
- L15:175 suppresses
- L15:1504 weakly suppresses
- L15:2229 weakly excites
- L16:2329 weakly excites
- L15:2333 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
- iaz
- Joint
- ough
- z
- зу
- 社
- Named
- ※
- 为企业
- ordial
Suppresses
- theless
- adays
- ONES
- elsif
- ethyst
- stdClass
- остей
- ELSE
- whales
- んだけど