capital L starting a place 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. 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
It fires on a capital L that starts a place or person's name, such as Lusaka, Lorraine, Lhasa or Lourdes. The opposite side is a lowercase l starting an ordinary word like 'lures'. Steering did not produce 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.
-
(Cape Town, South Africa) is a data analytics company using aerial imagery and machine learning algorithms to optimize crop performance for farmers globally.⏎AgriPredict (Lusaka, Zambia) provides farmers with access to information that will help them identify
-
one thousand dead and thirty to forty thousand jailed throughout Spain.⏎The day after the nationwide general strike was called on October 4, 1934, Lluís Companys, now president of the Generalitat, proclaimed the "
-
. In an earthquake, liquefaction can cause the ground to behave like quicksand, as seen in New Zealand and in the 1989 earthquake in Loma Prieta in the mountains of Santa Cruz, Calif.⏎In Japan, most
-
© CRDP de Lorraine, 2012).⏎The most outstanding examples of Renaissance architecture in
-
00,000 are ethnic Tibetans. In a survey carried out in Lhasa in July 1993 on the southern Lingkor, a
-
—on children is exceedingly rare in the Andes (except in cases of child sacrifice), so it is particularly revealing of the unstable sociopolitical conditions in the LIP [Late Intermediate Period] in the Ayacucho Basin.”⏎Previous work
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: there is not this rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule rule
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. They’re now one of the most common types of aquarium fish available, as well as being popular with anglers who enjoy catching them with live bait or artificial lures.⏎You might have heard of Synodontis catfish before, but chances
-
own species. Cally Harper, a doctoral student at Brown University, along with a team of researchers, set out to better understand how bats use their tongues when lapping up nectar. Scientists have known that they have hairy-looking tongues; the
-
9–1892) used both the one- and two-syllable pronunciations:⏎The purple flower droops; the golden bee⏎Is lily-cradled: I alone awake.⏎Both of these poets rhyme flower and
-
for honouring the country’s leaders. At many universities, North Korean students would wake their foreign classmates up and make them bring flowers to the bronze statue of Kim Il Sung in Wonsan to praise him. The foreign students were asked to give
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
- L23:268 excites
- L23:789 excites
- L24:482 suppresses inside Italian names and places
- L18:247 suppresses
- L23:352 suppresses
- L20:2575 excites
- L24:624 suppresses
- L22:461 suppresses
- L23:1760 suppresses
- L21:252 excites
- L18:198 weakly suppresses
- L16:1 weakly suppresses
- L20:1186 weakly suppresses
- L22:2157 weakly excites
- L20:234 weakly excites
- L24:346 weakly excites
- L25:338 weakly suppresses
- L25:744 weakly excites
- L1:1981 weakly suppresses
- L5:315 weakly suppresses
- L3:1599 weakly suppresses
- L18:550 weakly suppresses
Feeds later layers
- L27:515 weakly excites
- L27:2610 weakly excites
- L27:1583 weakly suppresses
- L27:2731 weakly excites
- L27:2593 weakly excites unclear: first letter of a name fragment
- L27:181 weakly excites
- L27:1184 weakly excites
- L27:854 weakly excites
- L27:2503 weakly suppresses
- L27:822 weakly excites
- L27:547 weakly suppresses
- L27:2694 weakly excites
- L27:408 weakly excites
- L27:1250 weakly excites
- L27:1258 weakly excites
- L27:1415 weakly excites
- L27:2742 weakly suppresses 'w' inside split words
- L27:2730 weakly excites
- L27:2866 weakly excites
- L27:1661 weakly excites
- L27:580 weakly excites
- L27:2100 weakly excites
- L27:2492 weakly excites
- L27:1129 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
- ozy
- meld
- ainter
- neys
- ocaust
- iteur
- ушки
- rence
- 身影
- 踏实
Suppresses
- orris
- samt
- занят
- يرا
- %E
- PROCUREMENT
- Specs
- -rule
- postage
- Beet