question marks ending a question
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. 21 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 lights up on the question mark at the end of a question, especially a question asked in the middle of a sentence or paragraph. When it was turned up, the model's text was not clearly about anything. Turned down, the model fell into long strings of question marks.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
. Armstrong said, but by watching larger trends. “We are not pushing the<|endoftext|>Would you be able to live in a house that was just four feet wide? That’s right – four feet wide! Well, Polish architect Jakub Szcz
-
at their heart to do with science. But the journey from evidence to policy is far from simple, always lengthy, frequently divisive and often ineffective. Why is that? To get to the bottom of this question we first need to ask another — what
-
Denial Paradox. When a guy denies he's a crook, what goes through most people's minds is, "why does he have to deny it? There must be some⏎truth to the accusation."⏎But there's something deeper going
-
, would he please hush and go away, and stay away, if he could, for five minutes? By all means, yes, he would. He’d stay away indefinitely if she
-
. But by making sustainability apply to imports, could the vote paradoxically harm farmers in developing countries by imposing a sort of hidden tariff for entry into the Swiss market? Jonathan Hepburn, senior programme manager for agriculture at Geneva-based International Centre for Trade
-
nutrition experts.⏎Healthy eating includes making healthful food choices, which means knowing what and how much you eat. Do you know the difference between serving and portion sizes? While the terms serving and portion often are used interchangeably, they actually mean different
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: how to make?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ask: “If you were forced to return to your home country by the U.S. government, would you bring your children born in the U.S. with you or would you leave them in this
-
Galilee has two famous inventions to his credit. What are they?⏎- Four inventions that were to change the history of the world came from China. What were they?⏎- Which great painter of the Renaissance was also a famous inventor?⏎- This
-
the Sinai Experience at some point during the prayer service so as to make sure that we don’t forget it. Yet we seemingly find no such reminder … or do we?⏎The truth is that the Rabbis did set in place a ritual in
-
to authentic Christianity. As a matter of fact, many who grew up in the church have the same question as those who didn’t: “What is Lent, anyway?”⏎What is Lent, Anyway?⏎Lent’s origin is hidden in the
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
- L17:1069 strongly excites
- L16:1078 strongly excites
- L20:1371 strongly suppresses comma inside a question
- L20:1957 strongly excites
- L18:2812 strongly excites
- L20:63 excites
- L16:2817 suppresses
- L16:1123 excites
- L18:1002 excites
- L15:725 suppresses
- L15:2412 suppresses
- L19:1084 suppresses "and" joining clauses after a condition
- L11:296 suppresses
- L14:1800 suppresses
- L13:2080 suppresses
- L18:310 suppresses
- L5:1800 suppresses
- L20:491 excites
- L20:218 suppresses
- L20:373 weakly excites
- L12:27 weakly excites
- L16:111 weakly excites
- L10:69 weakly suppresses
- L14:247 weakly suppresses "of" before a field or department name
Feeds later layers
- L25:541 strongly suppresses
- L24:402 strongly suppresses
- L27:233 strongly suppresses
- L26:613 strongly suppresses
- L23:164 strongly suppresses
- L27:975 strongly suppresses
- L26:953 strongly suppresses
- L27:838 strongly suppresses
- L27:1080 excites
- L25:1025 suppresses
- L24:2628 suppresses
- L26:11 suppresses
- L23:2365 excites
- L25:683 suppresses
- L27:1576 excites
- L24:125 suppresses
- L27:319 excites
- L27:347 excites
- L23:46 excites
- L27:1801 weakly excites
- L22:328 weakly excites
- L22:282 weakly suppresses period then line break in lists
- L22:321 weakly excites
- L26:1018 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
- ;",
- ;</
- .",
- ;",⏎
- .");
- .";
- .',
- ";
- ."),
- .",⏎
Suppresses
- ?
- ?
- ?⏎⏎
- ?⏎
- ?⏎⏎
- ?⏎
- ? ⏎
- ?"⏎
- ?)⏎
- ?”