'in' right after a place or time
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 word 'in' when it comes straight after a place name or a time phrase, like 'located in', 'Jallianwala Bagh in' or 'April in'. Turning it up or down changed the model's text in ways that did not clearly relate to this word.
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.
-
world as it was then and is even now.<|endoftext|>A NASA satellite the size of a small piano shot into space Wednesday atop an air-launched rocket, kicking off a two
-
plagiarism?" And why is<|endoftext|>The last few months have brought national attention to Tulsa, Oklahoma, as plans were unveiled to begin digging there in April in suspected mass graves — locations where corpses may have been dumped almost a century ago,
-
to 200 members of the Royal Army of Sulu, who have holed up in the village of Lahad Datu for the past two weeks in order to press their historic claim to the land. The Philippine and Malaysian governments are
-
ukan⏎Lachit Was born in 1622 on the 24th of November. He was born in the Ahim empire, located in today
-
ius Cicero⏎IN JULY 51 b.c. the Roman statesman and orator Marcus Tullius Cicero arrived in Cilicia, in
-
, the so called Amritsar Massacre of 1919. The massacre took place in a public park called Jallianwala Bagh in the city of Amritsar where British Indian army’s Colonel Reginald D
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: have you had a chance to get to some of the more remote areas that have some natural gas reserves, like peñolesole, peñolesole, peñolesole, peñolesole, peñolesole, pe
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ó, another student of Bortnyik, or László Káldor. One of Nemes’ most famous design is the poster for the
-
age and other places, in Sicily and Corsica, in Spain at Tartessus, and even beyond the pillars of Hercules at Gadeira (Cádiz)".⏎In the time of King
-
inadequate access to water and sanitation facilities. Nigeria has pledged its commitment to meeting Sustainable Development Goal (SDG) 6, which is to “ensure access to water and sanitation for all” by 2030. But the National
-
ounder and then pick the black doll). Watching black children struggle with those questions is very difficult.⏎Showing the video in a racially mixed class is risky, and always provokes interesting and complicated responses from the students. Often the classes can be unpredictable,
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
- L21:728 excites
- L19:2814 suppresses
- L20:2031 excites
- L7:2761 excites
- L19:2643 excites "is" in "what is" questions
- L21:825 suppresses
- L20:2460 excites
- L18:1586 excites
- L22:489 suppresses
- L23:594 excites
- L13:597 weakly excites
- L20:2335 weakly suppresses comma ending an 'if' clause
- L4:4 weakly suppresses
- L21:918 weakly excites
- L10:469 weakly suppresses 'of' after a noun phrase
- L20:219 weakly suppresses "the" after time or place prepositions
- L21:109 weakly excites
- L12:158 weakly excites
- L22:990 weakly suppresses
- L1:1262 weakly suppresses
- L1:1 weakly suppresses
- L15:2370 weakly excites
- L22:543 weakly suppresses
- L22:349 weakly suppresses
Feeds later layers
- L26:1104 strongly excites
- L27:656 strongly excites
- L26:1836 excites
- L26:1046 excites
- L27:1534 weakly excites
- L26:1804 weakly suppresses
- L27:887 weakly suppresses
- L27:891 weakly suppresses
- L26:2153 weakly suppresses
- L27:2175 weakly excites
- L26:787 weakly suppresses
- L26:613 weakly excites
- L27:2410 weakly suppresses
- L26:400 weakly excites
- L27:1314 weakly excites
- L26:937 weakly suppresses
- L26:3013 weakly suppresses
- L26:1128 weakly excites
- L26:1304 weakly suppresses
- L26:2755 weakly excites
- L26:2428 weakly suppresses
- L26:1673 weakly suppresses
- L26:1843 weakly suppresses
- L27:968 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
- what
- what
- What
- What
- whats
- WHAT
- .what
- 什么
- 是什么
- .What
Suppresses
- nerRadius
- viso
- ñ
- ativity
- regation
- ثن
- 口气
- groundColor
- HttpException
- 溟