opening parenthesis after a name
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 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 an opening bracket that comes right after a name or term and introduces another name, a translation, or a brand, such as 'cetirizine (Zyrtec)'. Turning it up made the model's output break into run-together words with no spaces, so the steering tells us little about the pattern itself.
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.
-
-lungring (‘Ol-mo lung-ring) on the eastern edge of Tagzig (sTag-gzig), who brought it to Zhang-zhung (Zhang-zhung) in the remote, distant past
-
urdwara' is from the words 'Gur (ਗੁਰ)' (a reference to the Sikh Gurus) and 'Dwara (ਦੁਆਰਾ)' (gateway in Gurmukhi
-
be holders of valid Super Licences, the highest class of racing licence issued by the FIA. The races are required to be held on tracks graded 1 (formerly A), the highest grade a track can receive by the FIA. Most
-
ing, itching, and pain affecting a wide area of the body (such as an entire arm or leg). Oral antihistamines like cetirizine (Zyrtec), diphenhydramine (Benadryl),
-
animals will not even consider what we eat to be edible (see note below), it’s important for us to share the culinary culture to show that there is another (and better) way.⏎According to Jeffrey Smith, author
-
people are so stupid as to use this stuff .”⏎By contrast, synthetic tetrahydrocannabinol (THC) is available as dronabinol (Marinol) and n
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 ' . ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' '
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ipers with mm divisions⏎- Paper towel⏎- Microscope slide and cover slip⏎AQA: Topic 184.108.40.206 and 220.1<|endoftext|>Re export or Re-export
-
ways. Arranging eight different tiems in a line can be done in 8×7×6×5×4×3×2×1 ways, etc
-
Jake Matijevic."⏎- Curiosity on Mars sits on rocks similar to those found in marshes in MexicoTue, 30 Oct 2012, 11:07:11 EDT⏎
-
Power of a sinusoidal waveform⏎Consider an electrical AC sinusoidal signal which is characterized by its voltage V(t)=Vmax×sin(ωt+ΦV) and current I(t)=Imax×sin(ωt+
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
- L14:140 strongly suppresses
- L18:76 strongly suppresses blank spot before a number
- L18:125 suppresses
- L7:92 excites
- L21:165 excites
- L21:29 excites
- L18:246 excites
- L20:433 suppresses
- L22:65 suppresses
- L19:1174 excites
- L15:184 excites
- L23:10 excites
- L16:396 suppresses
- L11:223 excites
- L19:1207 suppresses
- L18:118 excites
- L13:44 suppresses inside web addresses and citations
- L19:1377 excites
- L22:14 excites
- L16:723 suppresses
- L19:509 excites
- L15:59 suppresses
- L22:349 excites
- L5:7 suppresses
Feeds later layers
- L26:17 strongly excites
- L27:318 strongly excites opening parenthesis after a name or term
- L25:239 strongly excites
- L25:217 suppresses
- L26:247 excites
- L25:222 excites
- L27:337 excites
- L26:240 excites
- L27:11 suppresses unclear: mixed fragments in lists and pronunciations
- L27:344 excites
- L26:199 excites
- L27:68 excites
- L26:263 excites
- L27:220 excites end-of-text token between documents
- L27:323 excites
- L27:732 excites
- L27:431 excites
- L25:304 excites
- L26:459 excites opening bracket before an acronym
- L27:101 excites
- L27:513 excites
- L27:1077 excites
- L27:558 excites
- L27:1194 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
- Several
- regular
- ethical
- against
- onto
- since
- did
- terminated
- composed
- possible
Suppresses
- way
- me
- �
- those
- �
- he
- ver
- Statement
- rescue
- works