h starting a new word piece
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 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 letter h when it begins a new piece of a word, especially after a hyphen or a break, like the "-h" in "re-hydration" or "H" starting "Hypothesis". Turning it up gave odd, repeated words, so the steering does not show a clear theme.
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.
-
can. In-fact, these machine learning tools do out-predict individual traits better than work colleagues with only ten likes; family members, friends, and co-habitants with only 100 to 150 likes; and
-
transform atmospheric CO2 into organic molecules, plants can use the energy from⏎Transpiration decreases as air becomes drier.⏎De-hardening in spring involves gradual re-hydration of the cells, recovery of photosynthetic capacity and a tight control
-
any of these areas and would⏎like to participate in our research programs, please contact us.⏎ARTICHOKE SUPPLEMENTATION FOR DIABETES⏎HYPOTHESIS: Dietary supplementation⏎with Jerusalem artichoke (JA) significantly
-
loses some of its density) when coming in contact with acids. Saliva can dilute and neutralize acid so the tooth can re-mineralize (re-harden). This process isn’t instantaneous, however. While the enamel is soft,
-
OOK OF PHYSIOLOGY.⏎For Students and Practitioners of Medicine. By AUSTIN⏎A COURSE OF ELEMENTARY⏎PHYSIOLOGY AND HISTOLOGY. By Sir M. FOSTER, K. C.
-
the other pronoun, me, which is serves as a indirect object pronoun in me cuesta and as a reflexive pronoun in expresarme.¿Hablas el inglés? —Me cuesta trabajo hablarlo y expresarme,
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: as a child, as a child on a car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car car
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
In the following example, hijos is a masculine, plural noun so the masculine plural direct object pronoun los is used to replace it in the second sentence.Son muy buenos hijos. No los cambiaría.⏎They are good
-
from 1472 to 1554. This work was printed at Madrid in 1601; reprinted by Juan de la Cuesta in 1615; revised and augmented by Andrés González
-
, World Bank says. Wealth remains unequal, and some countries are falling behind.⏎El Banco Mundial analiza la riqueza de los países, tomando en cuenta no solo el PIB, sino también el capital humano y el natural.⏎
-
above the border checkpoint at San Ysidro, California. “En nombre del cielo, les pido posada,” sings one group. “Pues no puede andar mi esposa am
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
- L21:160 suppresses
- L22:245 suppresses
- L23:322 suppresses
- L24:346 excites
- L24:624 suppresses
- L24:375 suppresses
- L24:533 suppresses
- L23:268 weakly excites
- L24:231 weakly suppresses
- L21:266 weakly excites
- L22:322 weakly excites
- L19:388 weakly excites
- L13:1706 weakly excites
- L18:247 weakly suppresses
- L23:2095 weakly suppresses
- L13:1658 weakly suppresses
- L8:576 weakly excites
- L22:700 weakly suppresses
- L20:2338 weakly excites
- L24:482 weakly suppresses inside Italian names and places
- L16:2535 weakly excites
- L12:1269 weakly suppresses
- L16:1173 weakly suppresses
- L17:1759 weakly suppresses
Feeds later layers
- L27:600 weakly excites
- L27:723 weakly suppresses
- L26:878 weakly suppresses
- L27:855 weakly excites
- L26:829 weakly suppresses
- L26:2641 weakly suppresses
- L26:2398 weakly suppresses
- L27:667 weakly excites
- L27:1670 weakly suppresses
- L26:2040 weakly suppresses
- L27:2318 weakly excites
- L27:1483 weakly suppresses
- L27:321 weakly suppresses
- L26:2834 weakly excites
- L27:1085 weakly suppresses
- L26:546 weakly suppresses
- L27:2166 weakly suppresses
- L26:1429 weakly excites
- L26:431 weakly excites
- L27:2331 weakly suppresses
- L26:1843 weakly excites
- L27:636 weakly excites
- L27:1557 weakly excites
- L27:2678 weakly suppresses
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
- ig
- iras
- 睢
- �
- esseract
- isposable
- un
- quis
- elize
- ervised
Suppresses
- 第一百
- palms
- Nd
- orges
- Hearth
- seventeen
- intention
- /cards
- Nhap
- cars