unclear: mixed multilingual, mid-word splits
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. 24 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a wide mix of words that come right before a word is finished or continued, often in non-English text. Turning it up made the model drift into repeated, rambling sentences about keys and values.
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.
-
8 degrees Celsius, even for a few seconds, is enough to kill most bacteria, viruses, or parasites. Toxins produced by bacteria<|endoftext|>Dit artikel staat op icthealth.nl. Link: https://ictandhealth.com/ucla
-
άφει τετράγωνον σχῆμα τῷ λόφῳ in Einklang zu bringen, ist es unumgänglich nothwendig dass der Ausdruck Roma
-
) equal in magnitude but have opposite direction⏎C) not equal in magnitude but have the same direction⏎D) not equal in magnitude but have opposite direction⏎E) None of the above⏎In a uniform circular motion, which of the following
-
ative Accusative mein (m)meinenmeinemein meine (pl.)meine All ein-words follow the same pattern: ein, kein, possessive adjectives⏎Personal Pronouns in
-
that buys offsets for planting trees (not as good) that might be cut down a few years later<|endoftext|>Hand-coloured lithograph entitled; 'Dampfmaschinen Pferd worauf man in einer stunde von Paris nach Petersburg
-
in 1934, and issued to units in 1936. It introduced an entirely new concept in automatic firepower – the ''Einheitsmaschineng
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: It is now 2023, and there is no longer no longer there is no longer is now again now now again now now now now now now now now now now now now now now now now now now now now now now
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ی منظر، دانشکده معماری و شهرسازی، دانشگاه علم و صنعت ایران، تهران، ایران.⏎عنوان مقاله [English]⏎In
-
replacing or augmenting the electronicssil...⏎Introduceți subiectul și categoria pentru o căutare cât mai relevantă.<|endoftext|>How to be
-
and controlled by an adjustable valve and a scaven<|endoftext|>Deutsch: Faktor / Español: Factor - Factor / Français: Facteur - Facteur / Italiano: Fattori - Fattori⏎A factor, a Latin
-
-, etc.⏎Indefinite Articles in Accusative Case Nominative Accusative mein (m)meinenmeinemein meine (pl.)meine All ein-words follow the same pattern: ein, kein, possessive
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
- L22:525 excites
- L23:253 suppresses
- L21:266 excites
- L18:2749 suppresses
- L21:627 suppresses
- L21:40 excites
- L22:539 suppresses
- L3:1757 weakly excites
- L20:377 weakly suppresses
- L4:2283 weakly excites
- L8:2028 weakly suppresses
- L20:267 weakly suppresses
- L18:560 weakly excites
- L19:1096 weakly suppresses
- L14:1661 weakly suppresses
- L18:2343 weakly suppresses
- L23:2394 weakly excites
- L15:65 weakly excites
- L7:22 weakly suppresses
- L21:795 weakly excites
- L22:589 weakly excites
- L0:1224 weakly suppresses
- L0:1301 weakly suppresses
- L10:2580 weakly suppresses
Feeds later layers
- L25:1942 weakly suppresses
- L27:840 weakly suppresses
- L27:1163 weakly suppresses
- L27:9 weakly excites
- L25:2002 weakly excites
- L25:23 weakly suppresses
- L25:2364 weakly excites
- L26:488 weakly excites
- L27:666 weakly suppresses
- L27:1357 weakly suppresses
- L27:1029 weakly excites
- L26:1591 weakly suppresses
- L25:2985 weakly excites
- L27:531 weakly excites
- L25:2243 weakly suppresses
- L25:1223 weakly excites
- L27:1229 weakly excites
- L25:2135 weakly suppresses
- L25:1597 weakly excites
- L27:1996 weakly suppresses
- L27:1826 weakly excites
- L26:2884 weakly suppresses
- L26:2356 weakly suppresses
- L25:57 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
- (objects
- orian
- Ke
- Ke
- ince
- ippers
- enga
- states
- eneg
- ivant
Suppresses
- no
- no
- nos
- zp
- quires
- No
- zin
- będę
- →no
- No