zero after a decimal point
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 lights up on a zero that comes right after a decimal point, as in 0.0025 or 0.01. Turning it up made the model ramble on unrelated topics rather than talk about numbers.
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.
-
is freshwater (water that we can use), and we only have access to 1% of Earth’s freshwater. This means that we rely on 0.00025% of Earth’s water to support billions (people?). Imagine
-
of 0.02). Plug your population size and required margin of error into the formula. The
-
breadth and depth remain the most extensive, unexplored habit.⏎Activity 2 – Freshwater Habitats⏎Freshwater Habitats cover less than 0.01 % of the earth’s
-
96 %). The most important part of laser apparatus is laser crystal. Most commonly used laser crystalis manmade ruby consisting of aluminum oxide and 0.05% chromium. Crystal rods are round and end surfaces are made reflective.⏎A
-
A single heat pulse will switch the magnetic orientation of a nanometer-sized magnetic grain in just a few picoseconds (1 picosecond is 0.000000000001 seconds). As a comparison,
-
5.9 g in 125 ml pot of plain whole yogurt and 0.03 g in 25 g of cheddar cheese.⏎Cow’s milk allergy
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 need to know how to find the right side of a person. This is a common issue in the human brain. There are a lot of people and a person is not able to. In some cases, a person can be seen,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
with self-esteem are recognizable by the fact that:⏎- They believe in themselves and their abilities⏎- They do not lose faith in themselves when they make a mistake⏎- They do not hesitate to try new things
-
by the Nazis, tortured, and executed—at the age of 38. He can serve as a companion and guide for those who are struggling to keep faith in times and places of despair.⏎His heroic witness while in Hitler’s prisons should challenge
-
one birth?⏎There are many karmas recognized from ancient time, accepting the first two alternatives gives rise to a highly confusing situation. People will lose all faith in the karmas prescribed by the śāstras. The third alternative has to
-
an ignorant man. When he tells you he has a hunch about how it is going to work, he is uncertain about it. When he is pretty sure of how it is going to work, and he tells you, "This is 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
- L4:440 excites
- L4:987 suppresses
- L2:2708 weakly excites
- L3:622 weakly suppresses
- L2:902 weakly suppresses
- L4:2715 weakly excites
- L0:2306 weakly excites
- L0:781 weakly suppresses
- L0:817 weakly suppresses
- L2:1322 weakly suppresses
- L4:1626 weakly excites
- L0:904 weakly suppresses
- L0:3018 weakly excites
- L0:1224 weakly suppresses
- L0:1839 weakly suppresses
- L3:359 weakly suppresses
- L0:1301 weakly suppresses
- L3:1925 weakly suppresses
- L1:1841 weakly excites
- L3:625 weakly excites
- L2:1211 weakly suppresses
- L0:510 weakly excites
- L0:2261 weakly suppresses
- L1:2240 weakly excites
Feeds later layers
- L21:518 weakly suppresses
- L9:735 weakly excites
- L16:239 weakly suppresses
- L11:1546 weakly suppresses
- L10:1284 weakly suppresses
- L6:224 weakly suppresses
- L6:616 weakly excites
- L17:2782 weakly suppresses 'same' in 'at the same time'
- L6:1565 weakly excites
- L25:1599 weakly excites
- L11:1852 weakly suppresses the digit 1 in numbers
- L10:1602 weakly suppresses the number 1 as a unit value
- L8:2724 weakly suppresses
- L7:2344 weakly suppresses
- L19:395 weakly excites
- L11:2897 weakly suppresses
- L12:2830 weakly suppresses
- L23:123 weakly suppresses
- L26:661 weakly excites
- L8:2197 weakly suppresses
- L7:2262 weakly suppresses
- L21:278 weakly excites
- L13:2812 weakly suppresses
- L25:783 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
- ダー
- CAA
- vals
- izr
- чество
- ADA
- unist
- oria
- ッシ
- 一事
Suppresses
- eway
- ṭ
- ,"
- ,”
- ellung
- 찜
- emption
- uide
- trer
- obscured