percentages and shares in statistics
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. 13 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 lights up on the number in a statistic, like the 20 in '20 percent' or the 60 in '60 percent', especially where a share of a whole is being described. Turning it up made the model write in a more generic, repetitive style that didn't clearly relate to numbers.
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.
-
plaster, clay, or other substance to give it the colour and lustre of bronze. Dutch metal, an alloy of 80 percent copper and 20 percent zinc, is frequently used for bronzing.
-
years; 50.5% boys) and 38 who were unexposed (mean age, 6.8 years; 42.1% boys). The researchers used two 24-hour dietary
-
bigger than one third.⏎You arrange the data set in ascending order. You then find the observation such that a quarter of the observations are smaller than it and three quarters are bigger. That value is the lower quartile. Next find the observation such
-
when thinking about hackers and cyber criminals. While in 2015, 40 percent of attacks stemmed from ‘outsiders’, a surprising 60 percent were actually perpetrated by company insiders.⏎IBM, who produced the figures based on
-
, real estate markets are modeled as a stock/flow market. About 98% of supply consists of the stock of existing houses, while about 2% consists of the flow of
-
can exhibit soft tissue-like behavior. The nanotube arrays have a porosity of around 95 percent, meaning that carbon nanotubes only occupied five percent of the arrays while 95 percent of the array was occupied by air.⏎
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 can't have a similar effect, if I want to for the next time. So I need to make sure that this time, I can have a similar effect, if it's a different one. So a similar effect. So a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2 MHz⏎- Cancer 42 MHz⏎- Negative Thoughts lower our body’s frequency by 12 MHz⏎- Positive Thoughts raise our body’s frequency by 10 MHz⏎-
-
. Using these data, the author estimates the number of rental units and ownership units that would be affordable to prototypical households at specified income levels.⏎Previous Article | Next Article<|endoftext|>Scientists at Stanford University have created a way to produce thin, stretchable
-
45 right before landing: vx = 2cos25 = 1.8126 vy = 2sin25 = -0.845 3. The attempt at a solution Im not sure if my solution
-
Continuous splits are based on a threshold value, such as the split in a regression task where all instances below the threshold value are assigned to one class and all instances above the threshold value are
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
- L12:455 suppresses
- L12:2518 suppresses
- L12:15 suppresses
- L12:1960 suppresses
- L10:466 excites
- L3:2125 weakly excites
- L12:2674 weakly excites
- L12:2503 weakly suppresses
- L3:2036 weakly suppresses
- L2:2812 weakly excites
- L12:2826 weakly suppresses
- L12:606 weakly suppresses
- L12:909 weakly excites leading digit of metric conversions
- L12:1493 weakly suppresses
- L11:1529 weakly suppresses
- L11:1220 weakly suppresses
- L0:278 weakly suppresses
- L12:2591 weakly suppresses
- L9:715 weakly suppresses
- L12:2968 weakly excites
- L12:2945 weakly suppresses
- L1:33 weakly suppresses
- L11:2392 weakly excites
- L12:133 weakly suppresses
Feeds later layers
- L15:1563 strongly excites
- L15:2446 strongly excites
- L19:1343 excites
- L17:1062 excites
- L17:2957 excites
- L14:2418 excites
- L14:2521 excites
- L16:1920 excites
- L18:2084 excites
- L14:2445 excites
- L17:1132 excites
- L15:1921 excites
- L14:1799 excites
- L15:250 suppresses
- L15:2112 excites
- L15:449 excites
- L15:1037 excites
- L14:2375 excites
- L14:1165 excites
- L15:3013 excites
- L19:1989 excites
- L15:2060 excites
- L14:2117 excites
- L15:1032 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
- 另一半
- addCriterion
- ева
- remaining
- complement
- 젯
- .swt
- )__
- vette
- وا
Suppresses
- stag
- ounced
- ANGED
- 進
- tober
- }`).
- izon
- regist
- Skill
- angu