comma inside large 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. 8 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 fires on the comma in big numbers like 1,526,006 or 1,332, the one that splits thousands from the rest. Turning it up made the model's text drift into ordinary sentences and numbers, which tells us little about the neuron itself.
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.
-
(250.2 km2)|⏎|• Water||0.7 sq mi (1.7 km2)|⏎|Elevation||1,148 ft (350 m)|⏎|Time zone||Eastern
-
As of the 2010 Census, the city had a population of 1,526,006, growing to 1,547
-
silver production were between 1872 and 1883. At the peak of its prosperity, the town had an estimated population of 6,000"⏎"
-
Ireland. The most notable jewel in the crown is a pale yellow brilliant in the centre of the front cross, weighing four carats. The crown contains 1,332 other diamonds, as well as pearls edging the band. Some
-
Mahadev Temple is located in Dhyansar village of Una district of Himachal Pradesh. It is situated at an altitude of about 1,500 meters above sea level and offers stunning views of the surrounding hills and
-
1, in World War II, the German battleship Bismarck sank the British battle cruiser Hood in the North Atlantic, with the loss of 1,300 lives.⏎In 1975 President Gerald Ford approved more
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: not, h. o. d. o. e. s. a. g. and. c. and. c. and. c. and. c. and. c. and. c.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
second voyage to the New World the French arrived off the coast of Florida on Thursday, June 22, 1564 around three or four o’clock in the afternoon. They landed about thirty leagues south of the St. Johns
-
nine. 3. My friend (to ring) me up at eight o’clock tomorrow. 4. Where Kate (to go) when you (to
-
it’s been planted. Watering slowly and deeply is better for your tree’s over-all health then fast and frequent waterings. Watering before 10 a.m. and after 8 p.m. can reduce the amount of water lost
-
few fires but the drought hit us hard.⏎In the hills surrounding the town and further afield the evidence was obvious with many stands of trees brown and dying.⏎Although this
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
- L4:1657 excites
- L6:1299 suppresses
- L6:2016 excites
- L6:2363 suppresses
- L6:809 suppresses
- L7:2370 suppresses
- L6:184 excites
- L4:2827 excites
- L5:1387 excites
- L7:1684 excites
- L7:2139 excites
- L6:616 excites
- L5:2867 excites
- L7:2026 excites
- L6:2940 suppresses
- L6:1587 excites
- L7:2975 suppresses
- L6:785 excites
- L6:2938 suppresses
- L3:1661 suppresses
- L6:2264 excites
- L4:2974 suppresses
- L6:1695 excites
- L7:663 suppresses
Feeds later layers
- L15:646 strongly suppresses
- L16:592 strongly suppresses
- L12:576 suppresses 'of' in fixed phrases like table of contents
- L9:929 suppresses 'each' in 'each other'
- L17:1042 suppresses
- L9:209 suppresses
- L18:1321 suppresses
- L10:1705 suppresses
- L21:278 suppresses
- L16:2117 suppresses
- L9:116 suppresses
- L20:404 suppresses
- L19:107 excites
- L9:735 suppresses
- L16:286 suppresses
- L17:533 suppresses
- L10:678 suppresses
- L15:966 suppresses
- L11:300 suppresses
- L18:828 suppresses
- L25:645 suppresses
- L13:1552 suppresses
- L15:2051 suppresses
- L16:239 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
- obil
- assigned
- returnValue
- Hayes
- perplex
- _w
- NF
- NV
- väl
- Michaels
Suppresses
- 场上
- 份
- ₋
- 霉素
- 荪
- 项
- iture
- ICIAL
- HandlerContext
- stance