the word 'of' in 'middle of'
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. 0 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 lights up on the word 'of' when it comes right after 'middle', as in 'in the middle of the road'. Turning it up made the model's text drift into unrelated topics, so the steering results don't add much.
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.
-
. The audience shock-factor came from the garment that Bella Hadid wore, which was a spray-on dress. Hadid walked slowly onto a platform in the middle of the runway. Here she was sprayed by two technicians with a white material. With
-
's where Rob Johnstone chooses to spend his spare time.⏎As the sun begins to set, the astronomer drives out to an abandoned airstrip in the middle of the desert and sets up his complicated telescope as the remnants of the day's light
-
RAND's Robert Lempert studies climate change risk management. In October, his professional and personal lives collided when he was ordered to evacuate his home in the middle of the night to escape a spreading wildfire. Lempert's experience emphasized the challenges
-
sun. Wear sunglasses, a hat and a shirt, especially in the middle of the day (even on overcast days).⏎Eat right. Include at least five
-
deer behavior on the roads, sharing stories of deer not only darting out into the road but jumping right into the side of the car or stopped in the middle of the road, "caught in the headlights."⏎This video (thanks to National Forest
-
difference.⏎When does a child’s apology matter to parents?⏎<|endoftext|>Deep in a North Carolina marsh, a lone swamp sparrow sits on his perch in the middle of the water. He’s singing his usual song. But he’s also aggressively fl
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 think that the problem is to find the number of the number of the number of the number of the number of the number of the number of the number of the number of the number of the number of the number of the number of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Harvard T.H. Chan School of Public Health, Sept. 2012, https://www.hsph.harvard.edu/nutritionsource/vitamins/.⏎- Betta Health, VIC Government
-
0 through Summer 2011. The departments listed here offer sustainability course and programs. View the current complete course listings here.⏎A “sustainability-focused” course concentrates on the concept of sustainability, including its social, economic, and environmental
-
Indo-Aryan language—the Tamils speak the Dravidian language of Tamil. Again, while more than 90 percent of the Sinhalese are Buddhists, both Sri
-
-an-acre program. It is not uncommon for loggers, oil and gas companies, cattle ranchers, and miners to illegally extract resources from so-called “protected” areas.⏎Traditional forest communities are often the best stewards of the land because their
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
None: this is the first layer.
Feeds later layers
- L22:271 strongly suppresses
- L21:625 strongly excites
- L18:853 suppresses
- L22:1055 excites
- L2:761 excites
- L23:50 suppresses
- L23:474 suppresses
- L24:124 suppresses dollar sign before a money amount
- L21:1781 excites
- L27:1687 suppresses
- L21:3030 suppresses
- L21:2021 suppresses the 4 in years 184x, 164x
- L3:459 suppresses
- L20:854 excites
- L24:204 suppresses
- L25:2877 suppresses
- L4:524 excites
- L21:2500 excites
- L25:231 excites
- L26:319 suppresses
- L25:36 suppresses
- L23:1351 suppresses
- L20:2688 excites
- L4:216 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
- gether
- ly
- theless
- ン
- hip
- ful
- plier
- LY
- warts
- ся
Suppresses
- .persistent
- ×⏎⏎
- hình
- 形
- ares
- pressions
- 了
- gráfico
- PRESSION
- unique