unclear: 'decre' and split word fragments
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. 25 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 most strongly on the start of the word 'decree' (written as 'decre' before its ending) and on a few other word pieces in historical or legal text. Turning it up made the model wander off into unrelated writing.
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.
-
sentences. The UN Working Group on Enforced or Involuntary Disappearances (WGEID) stated that while Chile has
-
only way to ensure that the knowledge that leads to long life, and peace and prosperity in a nation is passed to each generation. “These are the commands, decrees and laws the Lord your God directed me to teach you to observe in the
-
870)|⏎|Convoked by||Paul III|⏎about 255 during the final sessions⏎Documents and statements⏎Seventeen dogmatic decrees covering then-disputed aspects of Catholic religion⏎|Chronological list of ec
-
his city's role in the Trans-Atlantic slave trade. He acknowledged that London was "tainted" by it<|endoftext|>Nearly three months after the Treaty of Paris was signed ending the American Revolution, the last British soldiers withdraw from New York City,
-
Poles turned from the Roman Church, winning a measure of religious freedom. Some church buildings were put to Protestant use. Laws forbade the execution of Roman Church decrees
-
housing, health and educational needs as well as economic, social and cultural rights.⏎The United Nations Convention on the Rights of the Child, to which New Zealand is a signatory, says that social services is also about ensuring that in all that we
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 *red* *and* *red* *and* *red* *and* *red* *and* *red* *and* *red* *and* *red* *and* *red*
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
loss of enthusiasm.⏎33. Failure will never overtake me if my determination to succeed is strong enough.⏎34. Failure doesn
-
one by one, they find themselves overwhelmed by guilt and contrition over their support of a murderous regime. He shows how their awakened moral reckonings and higher interests overrode lifetime habits and disciplines on the road to “the wolf’s lair
-
place on the American streets. It never belonged to the streets like during former days of jazz. This fusion of music was styled in a positive manner that made it overtake rap in early ‘90s (Keyes 245).
-
-bedded poorly sorted sand and gravel. The aquifer is composed of hydraulically connected alluvial and terrace deposits that unconformably overlie the Permian-age Formations.⏎The aquifer boundaries are from a ground
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
- L20:1074 weakly excites
- L3:1101 weakly suppresses
- L24:370 weakly suppresses
- L0:2774 weakly suppresses
- L9:833 weakly excites
- L19:758 weakly suppresses
- L15:930 weakly suppresses
- L21:2818 weakly excites
- L5:1838 weakly excites
- L24:1551 weakly excites
- L14:655 weakly suppresses
- L23:2421 weakly excites
- L19:1723 weakly excites
- L10:2797 weakly suppresses
- L21:1364 weakly excites
- L20:357 weakly suppresses
- L9:2970 weakly suppresses
- L14:2228 weakly suppresses
- L23:2064 weakly suppresses
- L12:965 weakly excites
- L20:2624 weakly excites
- L16:85 weakly excites
- L6:798 weakly suppresses
- L17:1416 weakly excites
Feeds later layers
- L27:612 weakly excites
- L27:1527 weakly excites
- L27:1560 weakly suppresses
- L26:1865 weakly excites
- L27:1282 weakly excites
- L27:1018 weakly excites unclear: mixed word fragments, no shared theme
- L26:2466 weakly excites
- L27:1242 weakly excites
- L26:2779 weakly excites
- L27:1298 weakly excites
- L27:1584 weakly suppresses
- L26:1158 weakly excites
- L26:2384 weakly excites
- L27:3067 weakly excites
- L26:831 weakly excites
- L27:983 weakly excites
- L26:2845 weakly excites
- L27:1252 weakly excites
- L27:1870 weakly excites
- L26:960 weakly suppresses
- L26:550 weakly suppresses
- L27:1845 weakly excites
- L27:1591 weakly excites
- L26:2667 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
- illery
- 观察
- hon
- щи
- clist
- wick
- ejs
- chos
- ôn
- obs
Suppresses
- izing
- aggio
- aggi
- aturing
- ization
- izational
- izado
- isation
- displayName
- 링