'in' as a phrase particle
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. 4 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 'in' when it finishes a phrase, like 'take things in hand', 'went up in the air', or 'close in late June'. The matches are weaker than the top neurons on this page. Turning it up did not produce related writing.
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.
-
for getting regulations published in the Federal Register, Rumsfeld decided to take things in hand. He assembled Jack Grayson, the chairman of the Price Commission, and
-
neighborhood of Seattle and is a University of Washington graduate. During WWII his parents were sent to the Minidoka concentration camp in Idaho.⏎A teenager is pulled back in time to witness her grandmother’s experiences in World War II-era Japanese internment camps
-
source is also<|endoftext|>Since 2002, when HAVA was enacted, computer scientists have released over three dozen studies of the software-based voting systems rushed in by HAVA. Every one of these studies has concluded that software-based optical scanners
-
check Fairness Act would close loopholes, strengthen business incentives to end pay discrimination, prohibit retaliation against workers who share wage information, and bring the Equal Pay Act in line with other civil rights laws. Passing this act does not just affect women,
-
men heard the shells edging closer. “And then bang!—one of the big shells struck the ground right in front of us and we all went up in the air. But we all come down again.”⏎York was showered with honors
-
great migration,” more societies are becoming like the United States, with a number of diaspora communities. But diasporas not only change the societies that took them in, they can also impact the countries they left behind.⏎“There’s a tendency to
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: A problem where a grid has a certain number of rows, and columns, and there's a certain number of rows? Let me first understand the context of the problem. This problem, the user's question, the grid has a certain number
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
In the real world, speakers sometimes also revert to illeism when they want to create
-
In the training room, athletes and athletic trainers often use the application of moist heat to boost
-
In a capitalist society, all human relationships are voluntary. Men are free to cooperate or not
-
?⏎In a common cathode RGB LED, the cathode of the internal LEDs are all connected
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
- L3:493 suppresses
- L2:2307 excites
- L1:1179 suppresses 'in' after 'result' (results in)
- L0:1280 weakly excites
- L0:1161 weakly excites
- L2:2918 weakly suppresses
- L1:290 weakly excites
- L3:1756 weakly suppresses
- L1:1718 weakly excites
- L0:536 weakly excites
- L1:2201 weakly excites
- L3:1116 weakly excites
- L0:2143 weakly excites
- L2:207 weakly suppresses
- L3:2739 weakly suppresses
- L0:1432 weakly excites
- L2:1541 weakly suppresses
- L1:1037 weakly suppresses
- L3:621 weakly suppresses
- L1:2953 weakly excites
- L2:1034 weakly suppresses
- L3:2871 weakly excites
- L2:1899 weakly excites
- L3:159 weakly excites
Feeds later layers
- L7:2878 weakly suppresses
- L10:2604 weakly suppresses
- L20:1356 weakly suppresses
- L18:1372 weakly suppresses
- L6:823 weakly suppresses
- L14:2709 weakly excites
- L12:1808 weakly suppresses
- L11:386 weakly suppresses
- L11:2172 weakly suppresses
- L15:2102 weakly excites
- L6:2072 weakly excites
- L26:1102 weakly suppresses fibromyalgia and hour counts
- L17:272 weakly suppresses
- L8:2722 weakly suppresses
- L6:2844 weakly suppresses
- L5:199 weakly suppresses
- L26:1469 weakly suppresses
- L18:1817 weakly excites
- L16:623 weakly suppresses
- L8:2610 weakly excites
- L17:612 weakly suppresses
- L23:2080 weakly suppresses
- L21:2501 weakly suppresses
- L12:1518 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
- imates
- esub
- 发展机遇
- 覆
- 收购
- trends
- -gen
- (“
- preserve
- ).</
Suppresses
- ulado
- ยาว
- bras
- flater
- _different
- ilitation
- unfinished
- assoc
- measurement
- ?