the 'un' prefix in 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. 26 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 prefix 'un' in the middle of a word, such as unless, ungodly, unselfish and unimportant. Turning it up made the model produce repeated, garbled text, so the steering does not show a clear meaning.
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.
-
(Mt. 12: 41).<|endoftext|>“There shall never be any bond slaverie, villinage or Captivitie amongst us unles it be lawfull Captives taken in just warres, and such strangers
-
will rule in the fear of God.’⏎The preservation of a republican government depends on the faithful discharge of this duty.⏎If the citizens neglect their duty and place unprincipled men in office, the government will soon be corrupted; laws will
-
of the second death in the "lake of fire" which will befall the ungodly.⏎Gehenna, or the Valley of Hinnom, was
-
bad with most characteristics exaggerated.The hero and heroine are usually young.The heroine is usually fair, kind, charitable, and caring.The hero is usually honorable, courageous, unselfish, and caring.Both usually have special abilities or powers.⏎What is
-
entices us to go deeper and reflect upon the meaning of things, to imagine how our ancestors conceived human nature and to think about what they thought was dangerous and unfruitful and why they felt the need to issue an advice and a warning for
-
of the skiers. It has been a peculiar battle. Both sides feel strongly on the matter, but even in the most crucial moments, there has been little uncontained anger. Indeed, the only thing that has been expended at all recklessly
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: Imblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleblebleble
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the time of electrophysiologic testing . Type I atrial flutter was classified as a macroreentrant atrial tachycardia (see "Electrocardiographic and electrophysiologic features of atrial flutter"), while type
-
also include tests for infections (microbiology), changes in body tissue and organs (anatomical histology), and chemicals, eg medication, drugs or poisons (toxicology and pharmacology).⏎In Adelaide, the post-mortem is performed
-
of the cause can be much more difficult.⏎There are some common triggers for hives. While it varies from individuals, these may include:⏎- Allergens (e.g., foods, pollens, animals, insect stings).⏎- Irr
-
As a result, scientists began to devote time to searching for drugs that would kill these disease-causing bacteria. The goal of such research was to find so-called “magic bullets” that would destroy microbes without toxicity to the person taking the drug.⏎
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
- L24:483 excites
- L23:864 suppresses
- L25:239 excites
- L20:605 suppresses
- L20:563 excites
- L24:620 excites
- L23:80 excites blank slot before a year or number
- L23:225 suppresses
- L21:1291 suppresses
- L21:1014 excites
- L21:959 suppresses
- L23:1015 suppresses
- L21:390 suppresses
- L15:59 weakly excites
- L19:305 weakly suppresses
- L21:118 weakly suppresses
- L20:267 weakly suppresses
- L8:328 weakly suppresses
- L22:245 weakly suppresses
- L20:270 weakly suppresses
- L23:196 weakly excites em dash opening an aside
- L17:1821 weakly excites
- L22:54 weakly suppresses
- L16:588 weakly suppresses
Feeds later layers
- L27:1385 weakly excites
- L27:1448 weakly suppresses
- L27:2774 weakly excites
- L27:2073 weakly excites
- L27:941 weakly suppresses
- L27:1273 weakly suppresses
- L27:1160 weakly suppresses word starts of acronyms and names
- L27:2941 weakly excites
- L27:1247 weakly suppresses
- L27:1399 weakly suppresses
- L27:2720 weakly excites
- L27:3051 weakly excites
- L27:1777 weakly excites
- L27:302 weakly suppresses
- L27:2160 weakly excites
- L27:320 weakly excites
- L27:402 weakly suppresses
- L27:538 weakly excites
- L27:1658 weakly excites
- L27:928 weakly excites
- L27:2152 weakly excites
- L27:1066 weakly suppresses
- L27:2474 weakly suppresses
- L27:1403 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
- ][
- mith
- rength
- amental
- ocrin
- tes
- lover
- loy
- matic
- roc
Suppresses
- ble
- LY
- Im
- 让孩子
- BODY
- im
- arin
- argin
- 把这些
- _cast