unclear: mixed, end-of-text and function 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. 15 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 a grab bag of spots: the end of a document, words like 'from', 'which' and 'could', and ordinary places in sentences. No single theme stands out. When pushed up, the model's writing turned repetitive and wandered toward generic phrases, but that doesn't point to a clear meaning.
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.
-
were based on NASA’s lunar architecture plans.⏎The game itself only takes around 20 minutes to play, and despite its sophisticated 3D visuals—the game<|endoftext|>Disk or partition alignment is something new to me. I have heard of it before
-
for which mankind are warranted, individually or collectively in interfering with the liberty of action of any of their number, is self-protection. That the only purpose for which power can be rightfully exercised over
-
of Publication⏎Blanks, C. M. (1948). San Antonio the Inland Gate. Retrieved from https://digitalcommons.pvamu.edu/pvamu-theses/39
-
insects in their diet. Further down the line, he hopes these mini-cricket factories will catch on as a sustainable local food source for North American cities. "<|endoftext|>Philadelphia is the largest city in the Commonwealth of Pennsylvania, the second largest city on
-
$50 from the mint."<|endoftext|>In the next of our blog posts for Census week, we showcase recent research undertaken
-
than €1 each to produce. 10 confirmed patients have been connected to a breathing machine via a 3D printed valve.⏎Although Fracassi is<|endoftext|>Key: "S:" = Show Synset (semantic) relations, "W
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: (other than the other...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...)(...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
or not and take the default route. Here’s an example of⏎mwhen for Monoids:⏎-- Important numbers!⏎xs = [6, 9, 13,
-
, this function is only declared as:⏎void * memchr ( const void *, int, size_t );⏎instead of the two overloaded versions provided in C++.⏎/* memchr example */⏎int main ()⏎char * pch;⏎char str = "
-
Ones. For this deed, she went down in Assassin history as the legendary Egyptian Assassin who killed Cleopatra with an asp.⏎- 1 Biography⏎- 2 Legacy⏎- 3 Personality and traits⏎- 4 Behind the
-
with <BODY> and end with </BODY>, but in practice this isn't essential. 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
- L14:1609 weakly excites
- L8:2748 weakly suppresses
- L3:1951 weakly excites
- L12:2329 weakly excites
- L3:955 weakly excites
- L10:60 weakly excites
- L7:1805 weakly suppresses
- L2:853 weakly excites
- L8:1455 weakly suppresses
- L8:1513 weakly excites
- L6:477 weakly excites
- L0:37 weakly excites
- L6:2355 weakly suppresses
- L14:781 weakly excites
- L2:1205 weakly excites
- L7:1348 weakly suppresses
- L13:1223 weakly excites
- L7:1806 weakly excites
- L0:3039 weakly suppresses
- L14:777 weakly excites
- L2:590 weakly excites
- L2:1134 weakly suppresses
- L1:617 weakly excites
- L7:2205 weakly excites
Feeds later layers
- L16:1119 weakly excites
- L27:834 weakly suppresses
- L17:1342 weakly suppresses
- L22:1018 weakly excites
- L19:1518 weakly suppresses
- L23:1498 weakly suppresses
- L17:1351 weakly suppresses
- L20:1312 weakly excites
- L18:2483 weakly excites
- L18:2411 weakly excites
- L27:1375 weakly suppresses
- L18:1052 weakly excites
- L18:2925 weakly suppresses
- L16:2849 weakly suppresses
- L25:620 weakly excites
- L23:347 weakly excites
- L21:2699 weakly suppresses
- L16:2196 weakly suppresses
- L18:837 weakly suppresses
- L19:707 weakly excites
- L23:2113 weakly suppresses
- L18:1845 weakly suppresses
- L24:668 weakly suppresses
- L27:841 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
- inger
- ude
- ypes
- aleza
- ube
- ylan
- .enable
- __;
- Alpha
- enable
Suppresses
- ronic
- цикл
- athi
- 衷
- loo
- 更大
- PROFITS
- wider
- broader
- rait