verb right before its object
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. 23 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 fires on a verb or preposition that is about to be followed by its object, such as 'picked', 'chose', 'showed' or 'submerges it in'. Its negative side fires on 'how' after words like 'when and'. Turning it up made the model repeat phrases like 'at least a few'.
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.
-
G) vs. electro-galvanized⏎Hot dip galvanization (HDG) is a process that takes cleaned and treated steel and submerges it in a molten zinc bath. The zinc metal at these temperatures bonds with the steel
-
student figured out who it was — Italian Prime Minister Silvio Berlusconi.⏎Todorov's next example illustrated how facial recognition can have serious consequences. He showed the students a photo of Darrell Edwards, a man who is currently fighting a
-
-D structure, but are as far away from morphine and codeine as possible," Manglik says.⏎The team ran 3 million possibilities through the computer and picked the <|endoftext|>Research on the relationship between microbiome and cancer has made significant progress
-
healthy like replacing full-fat version of foods such as milk, yogurt, and salad dressing with the fat-free or low-fat kind.⏎The best way to give your body the balanced nutrition it needs is by eating a variety of nutrient-packed foods every day
-
hydrogen peroxide reacts with organic material it breaks down into oxygen and water.⏎Whiten Clothes – An Alternative to Bleach⏎Add a cup of Peroxide to white clothes in your laundry to whiten them. Peroxide is great to
-
on in the same way, so I try to work in changes and transitions that make my lectures more interesting and engaging.⏎At the beginning of each lecture, I show and discuss “student questions of the day” that pertain to the previous lecture
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: on a day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day in the day
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, the ancestors of all people trace to Africa. That much is clear. We are all Africans. But there’s been considerable room for debate about exactly when and how many times modern humans made their way out of Africa to
-
sure though – the fascination with space exploration isn’t likely to go away any time soon.⏎“There’s a driving desire within us to find out who we are and how we fit and what the world around us is about,” said Kistler.⏎
-
future? I think my future is better."⏎Walesa spoke of his past – and of his fight. By analyzing our past, we can"understand how to build the future," he said. He then told the crowd that even though
-
about the body from the media.⏎"People don't remember how many factors are related to body images and the role models being pushed upon our children
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
- L22:554 suppresses
- L18:1339 excites
- L17:264 suppresses
- L20:593 excites
- L19:272 suppresses
- L19:2406 weakly excites
- L17:2363 weakly suppresses
- L20:417 weakly excites
- L19:1501 weakly suppresses
- L21:1851 weakly excites
- L19:1980 weakly excites
- L7:403 weakly suppresses
- L16:832 weakly excites
- L22:353 weakly suppresses
- L8:508 weakly suppresses
- L18:2153 weakly suppresses
- L17:1557 weakly suppresses
- L9:1787 weakly suppresses
- L11:810 weakly excites
- L19:1305 weakly suppresses unclear: verbs after a noun or 'that'
- L18:1601 weakly suppresses
- L20:728 weakly suppresses
- L16:1860 weakly suppresses
- L3:2730 weakly suppresses
Feeds later layers
- L25:1581 weakly excites
- L25:1132 weakly excites
- L24:380 weakly excites the word 'at' in place phrases
- L25:2791 weakly suppresses
- L25:664 weakly suppresses
- L25:696 weakly excites
- L26:898 weakly excites
- L24:1785 weakly suppresses
- L26:880 weakly excites
- L27:1484 weakly excites
- L26:148 weakly excites
- L24:655 weakly suppresses
- L27:2184 weakly excites
- L24:1929 weakly suppresses
- L24:1070 weakly suppresses
- L26:2452 weakly excites
- L24:1354 weakly suppresses
- L25:1865 weakly excites 'in' in 'keep in mind'
- L27:168 weakly excites
- L27:213 weakly suppresses
- L24:722 weakly excites 'about' or 'approximately' before a number
- L25:1157 weakly excites
- L25:1836 weakly suppresses
- L26:474 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
- trer
- icens
- atics
- alyze
- osit
- off
- ưỡ
- ateurs
- [out
- oplan
Suppresses
- 为基础
- often
- ithub
- usually
- ushman
- Cp
- often
- 大全
- 为主
- 有些人