unclear: first pieces of split 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 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
It mostly lights up on the first chunk of a word that the model has to finish, such as 'par' in 'parmesan' or 'conce' in 'conceives'. It doesn't clearly respond to one topic, and turning it down or up produced repetitive or off-topic text.
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.
-
ry: Saute broccoli florets with cooked quinoa, garlic, ginger, and soy sauce for a protein-packed and satisfying meal.⏎- Roasted broccoli with parmesan cheese: Toss broccoli flo
-
rets with olive oil, garlic, and parmesan cheese and roast in the oven for a crispy and flavorful side dish.⏎
-
'Avola has spread to nearby cities (Noto, Rosolini, Pachino, Ispica and Victoria) to affect areas of the province of Caltanissetta and Trapani (1968).⏎Prior to
-
, East Java; a green fashion show in Korea; and a project aimed at developing smart grid systems in Vietnam.⏎Young leaders can make a difference⏎Siti Nur Alliah, a British Council Climate Champion from Indonesia, is tackling a burning issue
-
of Mines, Golden, Colorado 1979:43 pages⏎Dust control techniques in coal mines were investigated to perfect methods of generating monodisperse coal dust, develop methods for directly measuring supersaturation, make improvements in
-
jaw that surround the roots of the teeth.⏎The curving part of the jaw into which the teeth are rooted.⏎The socket in the
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: they were half the room half the room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they were room they
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Peace Corps. He created the Foundation Heart Ventures-Grupo de Ap<|endoftext|>is an elementary subatomic particle with infinitesimal mass⏎(less than
-
out night patrols.⏎Ten Nieuport 80s went to Brazil in November 1918, where they were used by the Escola de Aviacao Militar at Campos dos Afonsos. Brazil also received a number
-
8, in the Nihon Shoki 日本書紀 (720) 9 and in the Hizen-no-kuni fudoki 肥前國風土記 (first part of the 8th century)
-
bu Anne mau makan di KFC.⏎Harimau tidak mau berlari ke rumah.⏎Perempuan muntah pohon.⏎Students
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
- L16:454 suppresses
- L24:851 excites
- L23:389 suppresses
- L22:819 suppresses
- L18:541 weakly suppresses
- L11:2040 weakly suppresses
- L21:40 weakly excites
- L20:174 weakly excites
- L6:328 weakly suppresses
- L6:344 weakly suppresses
- L21:160 weakly suppresses
- L23:352 weakly suppresses
- L25:57 weakly excites
- L1:1811 weakly suppresses
- L21:266 weakly excites
- L5:50 weakly excites
- L2:2154 weakly suppresses
- L19:853 weakly excites
- L0:26 weakly excites
- L18:102 weakly suppresses
- L16:463 weakly suppresses 'for' introducing a purpose or target
- L14:95 weakly excites
- L16:134 weakly suppresses
- L25:362 weakly excites
Feeds later layers
- L27:1233 weakly excites
- L27:796 weakly excites
- L27:1504 weakly excites
- L27:2210 weakly suppresses
- L27:779 weakly suppresses
- L27:783 weakly excites
- L27:2604 weakly excites
- L27:1909 weakly excites
- L27:866 weakly excites 'ch' starting a word fragment
- L27:752 weakly suppresses
- L27:568 weakly excites
- L27:2264 weakly excites
- L27:2061 weakly suppresses
- L27:1937 weakly excites
- L27:842 weakly excites
- L27:18 weakly excites
- L27:2190 weakly suppresses
- L27:1881 weakly suppresses
- L27:35 weakly suppresses
- L27:2330 weakly excites
- L27:2380 weakly suppresses
- L27:2 weakly excites
- L27:275 weakly suppresses
- L27:2874 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
- etta
- >window
- met
- cost
- ett
- 师
- 武
- 写字
- -cost
- 谋
Suppresses
- #{@
- utes
- ute
- ươi
- itional
- =#{
- /#{
- ց
- себе
- IDER