trailing zero in decimal 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
This neuron fires on a zero that pads out a decimal number, like the 0 in 1.000 or 10.00 or 0.750. Steering it produced rambling text that did not carry the number theme forward.
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.
-
if the marked price is 1,000 and has a discount rate of 25% should take the following relationship: 1.000×25 / 100 , or , 1.00
-
Human Development Index (HDI) that is regularly estimated by the United Nations Development Programme (UNDP) measures the progress. It was 0.750 in 2013 – a notable improvement from 0.55
-
a packed agenda to read aloud Boylan's request. The council members, struck by Boylen's youthful enthusiasm, passed the hat and collected $10.00.⏎Flush with funds, Boylan and Cook experimented with various rocket nozzle designs
-
1%), high (45.8%), and extra-high (67.1%) intensities, as well as clear-cut harvesting (100.0%), with non-harvesting as the control, based on permanent sample plots
-
tons)|⏎|Length:||205 ft (62.48 m)o/a|⏎|Beam:||33 ft (10.06 m)|⏎|Draught:||11.5 ft (
-
one serving there are 20 mg of Calcium and 1.08 mg of Iron. In addition, there are 610 mg of
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: What a good job do you have? The same thing, but not that much less? Maybe that, but not that much less. Let's see. I think the user might be. So, I need to check. Let me do
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is a quarter and three fourths is three quarters so tehrefore three fourths is bigger.⏎three tenths is more than a quarter⏎1 and a half is bigger⏎About a quarter of all stars are bigger than the sun, some
-
3.5 million African-American slaves, and now the Union Army began to recruit a second all
-
number of calories they get from food and beverages (energy IN) and increase their physical activity (energy OUT).⏎For a weight loss of 1–1 ½ pounds per week, daily intake should be reduced by 500 to
-
2.4 \times 100 \div 3.2$.⏎You could
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
- L4:440 excites
- L4:2290 weakly suppresses
- L4:2366 weakly excites
- L0:1006 weakly suppresses
- L4:2780 weakly excites
- L3:918 weakly excites
- L4:338 weakly excites
- L4:1423 weakly suppresses
- L0:2355 weakly excites
- L2:2834 weakly suppresses
- L3:2282 weakly excites
- L2:2143 weakly excites
- L3:1299 weakly suppresses
- L3:1316 weakly excites
- L1:317 weakly suppresses
- L1:154 weakly excites
- L0:1392 weakly excites
- L3:576 weakly suppresses
- L3:2569 weakly excites
- L3:1813 weakly excites
- L0:379 weakly excites
- L1:2654 weakly excites
- L3:604 weakly suppresses
- L4:1594 weakly excites
Feeds later layers
- L15:1855 suppresses
- L9:1373 weakly excites
- L21:1589 weakly suppresses
- L16:2783 weakly excites
- L8:981 weakly suppresses
- L19:928 weakly excites
- L11:2468 weakly excites
- L12:2262 weakly suppresses
- L18:2419 weakly suppresses
- L15:1103 weakly excites
- L23:1109 weakly suppresses
- L11:1941 weakly excites
- L11:890 weakly suppresses
- L11:1106 weakly suppresses
- L20:3052 weakly suppresses
- L24:2386 weakly excites
- L21:3000 weakly suppresses
- L26:3 weakly suppresses
- L20:1764 weakly suppresses
- L12:2447 weakly suppresses
- L20:2072 weakly suppresses
- L16:2280 weakly excites
- L18:1699 weakly excites
- L10:1848 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
- том
- MULT
- pháp
- 磋
- zespo
- venth
- ่ว
- _module
- alle
- forme
Suppresses
- ificates
- (wait
- Wait
- (inner
- etric
- exels
- rawer
- 相对
- lsx
- tık