words inside dialogue and questions
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 mixed: they peak on different tokens. 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 tends to light up inside spoken lines and questions, especially on words like 'you' in 'Do you' or on the quote mark after a speaker's words. Turning it up made the model drift into unrelated topics, so the link is not clear.
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.
-
garden.⏎‘What are you doing here?’ he cried in a very gruff voice, and the children ran away. ‘
-
ship of the desert - cannot be far behind either. Do you want to learn more? Stay with us at AnimalWised and discover the native
-
immediately. Just as I say, “ Imagine the tip of your left pinky finger… where the nail meets the skin…” Do you think of it? Did you think of your pinky because I mentioned it?
-
drawing a picture of.⏎"God," he said.⏎"Why Johnny," she said, "you can't do thatbecause nobody knows what God looks like."⏎"They will when
-
to sit down with them, explain what you are trying to accomplish, and get them to start preparing for the next day is after
-
beloved local garden guru Carolyn Singer wrote, “’Grandma, why don’t you grow yucky flowers?’ ‘Why would I do that, Marcus?’ I asked, respectful of his three-year-old wisdom. ‘Well,’ he replied, ‘
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: I need to get going and I need to get back to work and I should just give myself a go and do my job and I should have a good go at it and move on and I should have a go at it and just have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
⏎Open Browser and type “”http://Server-IP/drupal”” which will open below screen to configure Drupal. Select ‘Standard‘, click on Save and Continue.⏎Choose language, click on Save and Continue⏎Fill up the details
-
were done to ensure that the dogs would not pull a lever just for fun or for fear of the unfamiliar dog. They also were shown how to give themselves a treat.⏎“This control excludes the possibility that the dogs did not pull on the tray out
-
lot of people, Ginger.⏎I would say I think … yeah. I think you’re talking about more than 80 million people.⏎Wow. Yeah. Okay. Can you define it for us? Can you describe what insulin resistance is in
-
this. All this time, they’ve been moping about in existential despair. What, did you think I didn’t write this blog to amuse myself? Okay, obviously, the paper is about how black flies might be of some use to
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
- L11:2261 suppresses
- L10:375 excites
- L10:1003 weakly excites
- L1:2575 weakly suppresses
- L12:2835 weakly excites
- L14:2753 weakly suppresses
- L10:615 weakly excites
- L7:1893 weakly excites
- L6:1847 weakly suppresses
- L9:2336 weakly excites
- L13:2922 weakly excites
- L11:1030 weakly suppresses
- L12:2767 weakly excites
- L10:706 weakly excites
- L13:1259 weakly excites
- L11:561 weakly excites
- L2:2491 weakly suppresses
- L8:1071 weakly excites
- L13:45 weakly excites
- L2:748 weakly excites
- L12:530 weakly suppresses
- L12:2844 weakly excites
- L1:1648 weakly suppresses
- L9:1747 weakly suppresses
Feeds later layers
- L16:2828 suppresses
- L18:1260 suppresses
- L16:886 excites
- L23:1010 weakly suppresses
- L20:3056 weakly excites
- L19:2603 weakly suppresses
- L17:1528 weakly suppresses
- L22:2470 weakly suppresses
- L19:1973 weakly suppresses
- L16:2197 weakly excites
- L17:780 weakly suppresses
- L16:2377 weakly excites glosses, translations and unit conversions
- L18:2076 weakly suppresses
- L20:1055 weakly suppresses
- L17:2151 weakly suppresses
- L20:2972 weakly suppresses
- L24:2891 weakly excites
- L19:2425 weakly excites
- L16:3043 weakly suppresses
- L21:965 weakly excites
- L18:1224 weakly suppresses
- L20:2466 weakly excites
- L16:1247 weakly suppresses
- L16:2664 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
- oser
- 特殊
- 独特
- السن
- Mana
- Savage
- _male
- 一方
- avery
- 特殊的
Suppresses
- sucesso
- .restart
- Congratulations
- thanked
- xes
- retries
- .Quit
- 啶
- lingen
- 通知