zeros inside long 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. 5 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
The neuron reacts to a zero digit in the middle of a long number, such as '2,500,000' or '0.0000000001'. Turning it up did not produce any clear change in the topic of the 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.
-
0 meters or as 2570000000 millimeters.⏎- Colorado (western U.S.) 233
-
of a nanometer-sized magnetic grain in just a few picoseconds (1 picosecond is 0.000000000001 seconds). As a comparison, today’s hard drives take a few nanoseconds
-
0/15 IP: 10101100.00010000.00000000.00000000 SM: 1111
-
this child will become a doctor by the age of 40 years.⏎What I have just described<|endoftext|>|Area||2,500,000 sq km|⏎|Lowest Point||-5,267 m|⏎
-
drinking water after treatment with chlorinated compounds. Brines from shale gas wells contain dissolved bromide at concentrations up to 2,000,000 ppb and bromide can react with the chlorinated disinfectants and with
-
600,000 year earlier than had previously been documented.⏎- Our ancestors began making multicomponent
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 have a new friend, I have a new friend, and I have a new friend again. I have a new friend again, and I have a new friend again. So, I have four new friends. What is the total number
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
0⏎network => 172.16.0.0/15 IP: 101011
-
aryons, which are hadrons made of three quarks, make up almost all the matter we see in the universe today," said Brookhaven theoretical physicist Swagato Mukherjee, a co-author on a paper describing the new results in
-
diverse career in live performance presentations includes: Entertainment Operations Manager for the Tropicana, Production Coordinator for Lincoln Center Festivals, and Associate Producer/Writer for Westinghouse TV. He currently serves on the Executive Board of the Stage Managers’ Association
-
the plant.⏎Young plants should be watered Every day.⏎7) Ripsali bacciferra-⏎8) string of tears-⏎Pack of 8 hanging succulent⏎About the combo :⏎Succulents are the plants that store
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
- L2:1918 excites
- L4:1233 suppresses
- L1:1276 excites
- L1:21 suppresses
- L4:1180 suppresses
- L4:943 suppresses
- L2:1791 suppresses
- L3:219 weakly suppresses
- L4:1401 weakly suppresses unclear: letter fragments inside names
- L0:2319 weakly excites
- L0:2471 weakly excites
- L1:2211 weakly suppresses
- L1:1087 weakly suppresses
- L1:2288 weakly excites
- L1:1598 weakly excites
- L4:1117 weakly excites
- L0:1811 weakly suppresses
- L2:676 weakly suppresses
- L0:3043 weakly suppresses
- L1:1827 weakly suppresses
- L4:1808 weakly suppresses
- L4:1049 weakly excites
- L2:2442 weakly suppresses
- L0:1130 weakly suppresses
Feeds later layers
- L14:246 strongly suppresses
- L15:115 excites
- L17:166 suppresses
- L19:529 suppresses
- L19:1004 excites
- L17:360 excites
- L18:570 excites
- L20:83 suppresses
- L10:1228 excites
- L20:1040 suppresses
- L22:316 suppresses
- L15:96 suppresses
- L18:533 excites
- L10:445 suppresses
- L18:404 excites 'get', 'go' and 'got' after a subject
- L22:384 excites
- L18:137 suppresses 'the' after 'to' in fixed phrases
- L18:541 suppresses
- L18:288 excites
- L20:53 suppresses
- L17:389 excites
- L21:1112 suppresses
- L19:1865 suppresses
- L22:1141 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
- nd
- keiten
- es
- NING
- 肩
- ister
- loor
- ecute
- nings
- (pr
Suppresses
- ented
- orrow
- 行って
- Ỗ
- leased
- icator
- tees
- amarin
- 书法家
- returned