the digit zero inside numbers
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. 3 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 digit 0 when it sits inside a bigger number, like the zeros in 510,000 or 47 0uF. It is quiet on most other text. Turning it up made the model ramble about unrelated topics.
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.
-
510,000 Martini-Henry rifles in reserve.⏎The rifle used
-
Curtiss and purchase a plane for $7,000 ($157,000 in today’s money). While waiting for it to be built, he
-
December 1782 to March 1783, St. Augustine was home to 21,000 to 22,000 people.⏎If You Go⏎Doors to the Flagler College Auditor
-
, Carleton Journalism adjunct professor Barbara Freeman maintained that women working for newspapers in those early days were well-educated, bright, progressive and informed. They meshed well with other professionals, such as doctors, teachers and civil servants who joined them
-
few hundred quadrillion oscillations per second, corresponding to wave-lengths between a few hundred and a few billionths of metres. Known in everyday life as that part of the radiation we receive from the sun that causes our skin to tan⏎Syn
-
enrolments a year.⏎Currently people only complete about 300 hours of the 510 available and only 21% leave with a functional level of English.⏎Australia
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: of people who are interested in the time, and I have a lot of people who are interested in the time. So of people. I have a lot of people who are interested in the time. And I have a lot of people who
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
it. What this implies is that caffeine is doing something to your internal brain chemistry to make this happen – and where there’s neurochemistry, there’s a ton of crazy internal mechanisms that actually make that effect happen.⏎One can easily and efficiently
-
serious ways with a series he calls "Tons." Each piece is made up of one ton of nails, and to date he
-
Bisbee's intuitive working method comes closest to the surrealist process of automatic drawing, allowing the subconscious mind to direct the hand.⏎In addition to the "Tons" series, Bisbee has recently produced more intricately patterned pieces. These
-
ular/lymphoid tissue encircling the back of the throat. Tonsils are visible at the back of the mouth, one on each side. The
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
- L2:2034 weakly excites
- L1:1206 weakly excites
- L1:788 weakly suppresses
- L2:36 weakly excites
- L1:2918 weakly excites
- L1:2244 weakly suppresses
- L2:49 weakly suppresses
- L0:1624 weakly suppresses
- L2:2692 weakly excites
- L0:433 weakly excites
- L2:2685 weakly excites
- L0:2527 weakly excites
- L2:1805 weakly suppresses
- L1:1421 weakly suppresses
- L0:2451 weakly excites
- L2:1669 weakly suppresses
- L0:371 weakly suppresses
- L1:1429 weakly suppresses
- L2:2052 weakly suppresses
- L1:1374 weakly suppresses
- L2:421 weakly excites
- L2:2699 weakly suppresses
- L1:1536 weakly excites
- L2:1851 weakly excites
Feeds later layers
- L4:2754 suppresses
- L26:1621 suppresses
- L4:1537 weakly suppresses
- L7:2354 weakly excites
- L17:774 weakly suppresses
- L4:2401 weakly excites
- L9:1220 weakly suppresses
- L7:378 weakly suppresses
- L4:2938 weakly suppresses
- L9:547 weakly excites
- L24:855 weakly suppresses
- L15:2518 weakly excites
- L7:1483 weakly suppresses
- L17:1270 weakly suppresses
- L7:1678 weakly excites
- L19:2039 weakly excites
- L24:484 weakly excites
- L25:1046 weakly excites
- L7:2932 weakly excites
- L14:2993 weakly suppresses
- L17:342 weakly excites
- L15:2881 weakly suppresses
- L19:819 weakly suppresses
- L11:1555 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
- icemail
- :\\
- lemetry
- 中关
- emento
- ddit
- rim
- pend
- poste
- ILog
Suppresses
- ATURE
- ᵃ
- 有一种
- }`).
- !”⏎⏎
- )⏎⏎
- ’⏎⏎
- match
- ()⏎⏎
- ty