the zero in round numbers like 100
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
This neuron fires on the zero inside numbers like 100, 1000 or 300, and on the word "seventy." It is about the size of a round number being written out. Turning it up made the model talk about numbers and counting, which is loosely related.
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.
-
74 Illingworth accompanied George Armstrong Custer’s expedition into the Black Hills, and he captured over seventy images. Illingworth was taken to court for embezzling government property
-
Valley called TransitCamp. Inspired by a similar event in Toronto, the idea was to brainstorm what you might do with transit-agency data. Nearly 100 people came. One guy was looking to build a Web site that combined an online
-
It's likely to get lost. Your character will appear wh<|endoftext|>The 45th annual PDK/Gallup Poll collected data from over 1000 Americans ages 18+ regarding the approach to improving the
-
stability and serenity. Other significant plants on the property are especially notable for their maturity and precise placement. Some rhododendrons are over 100 years old, their gnarled trunks as attractive as their flowers. Tr
-
-25 percent higher than houses with no trees.⏎Trees generate jobs and contribute raw materials for buildings, newspapers, books and more than 15,000 other forest products. Trees are renewable, biodegradable and recyclable.
-
regulating your intake and staying healthy, both physically and mentally.<|endoftext|>Tech moves fast! Stay ahead of the curve with Techopedia!⏎Join nearly 200,000 subscribers who receive actionable tech insights from Techopedia.⏎A
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: 1/1, 2/2, 3/4, 5/5. 5/5? Well, I need 50000 to be 10000? Let me try that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
reformer” led to his⏎appointment as head of Scots Church, Collins Street (1875-83) in Melbourne.⏎Yet, “almost from the hour of his arrival”, controversy was never far away⏎and his essays,
-
, which heightens the intensity of this sound installation. The volume also oscillates—some parts are loud, others whir and buzz at levels that are almost inaudible, then it becomes tinkly and jolting before cycling back to
-
and Ukraine, approximately in the 10th century. The Cossacks were known for their military
-
excess weight, more than it is considered healthy for their height and age.⏎People with this disorder may go to great lengths to diet, almost to the point of starving or exercise even if they are underweight.⏎It’s all
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
- L4:2624 suppresses
- L3:830 weakly excites
- L4:546 weakly suppresses
- L4:1905 weakly suppresses
- L4:2951 weakly excites
- L2:2968 weakly excites
- L1:1810 weakly excites
- L2:1002 weakly excites
- L1:2607 weakly suppresses
- L3:2111 weakly excites
- L3:632 weakly excites
- L1:1047 weakly excites
- L2:1874 weakly suppresses
- L4:3049 weakly suppresses
- L3:1086 weakly suppresses
- L3:2492 weakly suppresses
- L3:453 weakly suppresses
- L1:2797 weakly suppresses
- L0:1067 weakly suppresses
- L2:2789 weakly excites
- L1:2278 weakly excites
- L3:2703 weakly suppresses
- L0:593 weakly excites
- L3:3048 weakly suppresses 'called' after 'so'
Feeds later layers
- L10:1636 excites
- L9:620 excites
- L7:534 excites 'much' or 'many' after 'not' or 'little'
- L7:755 weakly excites
- L15:2484 weakly excites
- L22:1736 weakly excites
- L6:1037 weakly excites
- L16:1273 weakly excites
- L15:2350 weakly excites
- L11:2495 weakly excites
- L9:781 weakly excites
- L8:391 weakly excites
- L7:245 weakly excites 'even' in 'and even' lists
- L11:2797 weakly excites
- L18:1309 weakly excites
- L7:561 weakly suppresses 'all' or 'every' before a group
- L13:1083 weakly suppresses
- L8:2933 weakly suppresses
- L10:1063 weakly excites
- L26:1200 weakly excites
- L22:1056 weakly suppresses
- L6:1877 weakly excites
- L25:1731 weakly excites
- L20:4 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
- EPROM
- ander
- inct
- adero
- agem
- raph
- ��
- .dw
- dpi
- !"⏎
Suppresses
- 格
- 不对
- inen
- ais
- 使者
- 推进
- galement
- 第一节
- orney
- hungry