'and' joining two actions
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. 17 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 actions in a row, like 'packed up our gear and selected food'. Turning it up made the model produce repeated filler words, so the steering does not clearly show this role.
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.
-
Albemarle County was subsequently enlarged, and in 1696 the area south of Albemarle Sound was removed from Albemarle and made into a new county, named Bath
-
own language. By the 8th century, the Cham script had outgrown Sanskrit and the Cham language was in full use. Most preserved manuscripts focus on religious rituals,
-
potential gentrification that often comes with neighborhood investment in disadvantaged communities.⏎The Urban Green<|endoftext|>Every winter, California newts leave the safety of their forest burrows and travel as far as three miles to mate in the pond where they were born.
-
surface-to-air missile.⏎Powers released his parachute and floated down to the surface where he was detained by KGB agents and interrogated
-
thousands of Japanese American internees. One of the first families to arrive was the Wakatsukis, who were ordered to leave their fishing business in Long Beach and take with them only the belongings they could carry. For Jeanne Wakatsuki
-
are off limits (unless a dieter chooses to put the risk of severe foodborne illness aside and eat raw meat). For this reason, the raw food diet typically resembles a vegan
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: What if the one thing that is, in the one, that one, that one, that one, that one, that one, that one, that one, that one, that one, that one, that one, that one,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
up to 100 times faster than Map Reduce for iterative algorithms or interactive data mining.⏎- Learn how it provides in-memory cluster computing for lightning fast speed and supports Java, Python, R, and Scala APIs for ease of development.⏎-
-
Germany occurred in 1966, but it was the oil crisis of 1973/74 that ended the "economic miracle" once and for all. By the time Herbert Grünewald succeeded Kurt Hansen as Management Board
-
pupils are taught and governed, are what it mainly embraces, and the teacher is successful in the Art of Teaching, who trains children to be symmetrically developed men and women, in a physical, intellectual and moral sense. Certain modes of procedure,
-
transported soil from the mainland, tucking it into crevices on this otherwise barren rock in hopes of cultivating flowers and vegetables. For a few months in summer and early fall, the rock became a colorful garden in the
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
- L10:155 excites
- L15:184 excites
- L13:44 suppresses inside web addresses and citations
- L13:131 excites 'and' joining parallel phrases
- L16:376 suppresses 'or' joining alternatives
- L11:223 excites
- L8:328 excites
- L4:76 excites
- L15:59 suppresses
- L16:1184 excites
- L7:92 excites
- L14:1639 suppresses "and" joining two nouns
- L10:74 excites
- L2:24 suppresses
- L5:548 excites
- L12:158 suppresses
- L16:771 suppresses
- L9:92 weakly suppresses "and" or "or" joining two items
- L15:830 weakly excites
- L4:57 weakly suppresses
- L10:370 weakly suppresses
- L15:1260 weakly excites
- L10:1377 weakly suppresses
- L15:2765 weakly excites
Feeds later layers
- L19:1769 excites
- L21:299 excites
- L19:271 excites
- L20:1214 excites
- L19:1608 excites
- L25:824 excites
- L18:2045 excites
- L21:535 weakly excites
- L18:472 weakly excites
- L18:715 weakly excites
- L24:398 weakly excites
- L23:566 weakly excites
- L18:529 weakly excites
- L20:1124 weakly suppresses
- L18:1533 weakly excites
- L19:385 weakly suppresses
- L19:652 weakly excites
- L18:2393 weakly excites
- L20:2815 weakly excites
- L20:1399 weakly excites
- L18:2216 weakly excites
- L18:740 weakly excites
- L19:504 weakly suppresses
- L18:597 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
- 堃
- ędzi
- øy
- azio
- readcrumbs
- dependencies
- VERAGE
- charts
- schläge
- 朴
Suppresses
- them
- issent
- They
- loro
- ihnen
- .Next
- →dc
- Next
- 家都知道
- "].(