'you' after if, when, and the more
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. 8 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 responds to the word 'you' (and nearby words like 'to' or 'is') when it comes right after a word like 'if', 'when' or 'the more'. It is about the generic 'you' in a sentence. Turning it up did not produce any clear theme in the written text.
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.
-
cold winter and the snow was deep in the woods surrounding the logging camp near the Dungarvon River. But there was plenty of game to be found if you were a good hunter, so the men didn't go hungry. The camp cook
-
almost entirely Muslim. “Also, because of her Malay, Arabic comes easily.” Eunice adds, “And, English helps with French. The more languages you know the easier it is to learn.” According to Angela, her colleague is indeed
-
, according to⏎"If you're an assistant professor proposing to study it, I don't⏎think you'd get tenure," he says. "But one has to exist."⏎laboratory at Stanford University's Hopkins Marine Station⏎in Pacific Grove
-
’s George Leach Centre Wednesday, as the injustices and struggles of Canada’s indigenous people were recounted, one blanket at a time.⏎“It’s very impactful when you relive it, it’s always got that pain,” said Shirley Horn, Alg
-
animals) the the fact that it still needs to be mounted on a "neck" to see much of the field.⏎Acctually, it depends on what you are doing. The greatest advancement of science, wealth and welfare in history has come
-
't quite get to the⏎> issue though.⏎> The problem I'm concerned about is the need to define y in the first⏎> place. If one is chasing through a data structure and finds a need to⏎> change something buried within
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: the current state of the project is the problem. You know, I'm your assistant? Maybe you can ask me questions for me, but I need to make sure I understand what the current situation is. If I think about the problem,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and science for every school year group. We really hope that you can use some, or all, of this content as you see fit, whether that
-
End of Year Math Review is not your typical math printables…this is math review with a twist! A digital version has been added!⏎With this resource, your students will be engaged and having fun all while reviewing what they’ve learned throughout the
-
the argument found in our assigned readings, and that in our in-class and online discussions we have already hashed out what we think the readings mean.⏎Instead of asking you to do that low-level thinking again, this assignment asks you to focus on applying
-
the rapid social changes around us, there exist certain constants like the family, religion, language, national identity, and democracy.⏎But the fact is, as you’ve heard me say many times before, everything changes and evolves, including the family,
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
- L7:169 excites
- L4:499 excites
- L7:194 excites
- L5:2771 weakly suppresses
- L7:1375 weakly suppresses
- L5:222 weakly suppresses
- L1:1063 weakly suppresses
- L7:500 weakly excites
- L3:678 weakly suppresses
- L2:2299 weakly suppresses
- L6:234 weakly excites
- L1:2392 weakly suppresses
- L5:158 weakly suppresses
- L5:52 weakly excites
- L4:2089 weakly suppresses
- L4:515 weakly excites
- L1:294 weakly excites
- L2:2025 weakly excites
- L3:838 weakly suppresses
- L6:1564 weakly suppresses
- L5:1820 weakly suppresses
- L2:1566 weakly suppresses
- L7:959 weakly suppresses
- L3:981 weakly suppresses
Feeds later layers
- L10:1309 excites
- L15:1204 excites
- L10:293 excites
- L9:1279 weakly excites
- L23:2177 weakly suppresses
- L9:576 weakly suppresses
- L17:730 weakly suppresses
- L20:592 weakly suppresses
- L16:3044 weakly suppresses
- L12:976 weakly excites
- L19:2133 weakly excites
- L12:1702 weakly suppresses
- L14:2998 weakly suppresses
- L12:937 weakly excites
- L17:2643 weakly suppresses
- L11:937 weakly suppresses
- L9:1417 weakly suppresses
- L12:1671 weakly suppresses
- L17:1118 weakly excites mid-clause words in technical prose
- L13:722 weakly excites
- L13:2205 weakly suppresses
- L15:86 weakly suppresses
- L15:1113 weakly suppresses
- L9:1236 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
- theor
- anz
- subject
- lick
- etr
- theory
- subject
- nel
- know
- .setView
Suppresses
- 你们
- 成年人
- 你們
- OUS
- 小男孩
- phans
- educators
- exampleModalLabel
- 您的孩子
- .enqueue