digits inside numbers, not number words
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. 9 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 reacts to a single digit in the middle of a number, such as the 5 in 45 or the zeros in 1,500 or 25,000. It goes quiet, or pushes the other way, on spelled-out large-number words like "hundreds" or "thousands". Turning it up or down did not give a clear new theme in the sample text.
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.
-
ing.⏎location of the weather station site to be surveyed in advance. Sites are listed by name, and give an approximate lat/long to the nearest⏎hundredth of a degree, but
-
Which medical conditions medical marijuana treats⏎- The impact smoking or vaping has on the lungs and other body systems⏎- Four engaging marijuana prevention units, each 45 minutes long:⏎- Marijuana and the brain⏎- Marijuana and the body⏎-
-
walking distance of this marker. Montrose Park (about 400 feet away, measured in a direct line); Watching the Flames (about 700 feet away); Dumbarton Oaks (about 700 feet away);
-
,034.⏎Eureka is the largest coastal city between San Francisco and Portland, and the westernmost city of more than 25,000 residents in the 48 contiguous states. It is the regional center for government
-
reduced by 500 to 750 calories. In general:⏎- Eating plans that contain 1,200–1,500 calories each day will help most women lose weight safely.⏎- Eating plans that contain
-
members of the population consume no more than 2,300 mg of sodium.*** The American Heart Association recommends no more than 1,500 mg
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: The first time, it was a question of a single person, but if you have more than one, or three, or five, or even one, or two, or three, or four, or five, then it's something,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Sun inside them! Stars only appear small in the nighttime sky because they are so far away.⏎In the long ago, astronomers once thought that stars numbered in the hundreds or thousands, instead of the now-estimated “billion trillions.” About
-
, impossibly large slice of silicon that does it all. Apple’s most powerful chip to date has 114 billion transistors packed into over a hundred processing cores dedicated to logic, graphics, and artificial intelligence, all of it connected
-
dots written in layers that are only a few nanometers thick that cover the surface of millions of saucer-sized spinning disks. These hard disks are stacked by the thousands in server farms worldwide.⏎In recent years, the technology for growing uniform magnetic layers
-
hundred twenty-four years after it was first proposed, Kenneth Appel and Wolfgang Haken cleverly reduced the problem to checking the properties of 1,936 maps by computer. The result was hundreds of pages of hand analysis combined with thousands
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
- L8:41 suppresses
- L5:1109 excites
- L3:1129 suppresses
- L7:415 weakly suppresses
- L6:586 weakly suppresses
- L8:1939 weakly suppresses
- L5:1662 weakly suppresses
- L7:1313 weakly excites
- L7:1643 weakly suppresses
- L4:1263 weakly excites
- L4:1845 weakly excites
- L4:326 weakly suppresses
- L4:97 weakly suppresses
- L3:202 weakly suppresses
- L1:2110 weakly suppresses
- L5:1671 weakly excites
- L3:2436 weakly suppresses
- L7:2580 weakly excites
- L7:2343 weakly suppresses
- L8:780 weakly excites
- L8:2079 weakly excites
- L3:1771 weakly excites
- L7:1506 weakly excites
- L5:1432 weakly excites
Feeds later layers
- L10:2300 suppresses
- L12:2284 suppresses
- L13:2603 excites
- L12:2116 weakly suppresses
- L10:1031 weakly suppresses
- L12:217 weakly excites
- L23:104 weakly suppresses
- L21:2833 weakly suppresses
- L12:824 weakly suppresses
- L10:2791 weakly suppresses
- L11:1783 weakly suppresses
- L25:1436 weakly excites
- L13:2563 weakly suppresses
- L10:2499 weakly suppresses
- L13:2679 weakly suppresses
- L15:2708 weakly suppresses
- L12:947 weakly suppresses
- L14:2379 weakly suppresses
- L16:1410 weakly excites
- L11:1016 weakly excites
- L19:2291 weakly suppresses
- L14:2765 weakly suppresses
- L11:301 weakly suppresses
- L25:2573 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
- VA
- hip
- اظ
- AYOUT
- itti
- inea
- преимуществ
- @property
- па
- alance
Suppresses
- mers
- atology
- zens
- umd
- odic
- 团长
- zos
- 重重
- zell
- aptcha