empty slot just before a number
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 at a point where a number is about to appear, such as "all the[[ ]]4 volunteers" or "its[[ ]]47th close pass". The number itself isn't in the text yet. When it was turned up, the model's replies began writing numbers and quantities, like "60% of the day" and "45 years".
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
. Even if you don't make it all the way to the top of the mountain or hike all 40 miles of the trail, you'll have to hike just to get back
-
m) wide. Among all of the 37 shipwrecks discovered at Yenikapi, there were a
-
of the giant planet Jupiter and its tiny lava-covered moon Io.⏎When the solar-powered spacecraft completed its 47th close pass (perijove) from Jupiter on December 1
-
plane crash survivor to avoid starvation, and abusing a kitten.⏎All the 4 volunteers put the actions on the 'wrong' side of the scale. But
-
incorporate it into your daily life, the more you will find you use it.⏎A degree of experimentation is necessary to get exact cooking times for<|endoftext|>Here are the 15 types of distorted thinking:⏎1. Filtering: You take the negative details
-
ities, including the arms, legs, hands, and feet.<|endoftext|>From the time of Aristotle (384-322 BC) until the late 1500’s, gravity was
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: that I can get more information about the possible. I am not sure if I am getting it. - - - - - - - - - - - - - - - - - - - - - - -
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
be influenced by his internal characteristics; and the second is in which the possible causes of a person’s actions are related to his situation. This is the theory of attribution in social psychology. However, Allah’s Messenger said,⏎“No child is born except
-
some of the dangers of this type of attitude:⏎* If people feel that they have little control over their future, they will be unwilling to put a great deal of effort into maintaining their physical well-being. They will not see the point in exercising
-
and disorders, addiction is multifactorial, resulting from a combination of genetic, social, psychological, and environmental forces.⏎Genetics plays an important role in conferring vulnerability to addiction. Genetic factors influence how substances are metabolized and experienced. Such
-
idical approaches, which give rise to expectations. Frustrated expectations result in an effective response. Music only has meaning unto itself and the music listener ascribes any extra-musical meaning. This includes any emotional meaning. The unfolding of
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
- L1:664 weakly excites
- L3:1830 weakly suppresses
- L0:2423 weakly excites
- L2:2544 weakly suppresses
- L3:814 weakly suppresses
- L4:53 weakly excites blank slot before a number
- L4:1438 weakly excites
- L4:292 weakly suppresses
- L2:1694 weakly suppresses
- L2:2204 weakly excites
- L1:636 weakly suppresses
- L2:2037 weakly excites the digit 5 in decimals like 1.25
- L4:2314 weakly suppresses
- L4:654 weakly excites
- L1:1237 weakly suppresses
- L4:205 weakly excites the digit 2 inside citations
- L1:846 weakly suppresses
- L1:487 weakly suppresses
- L0:1269 weakly suppresses
- L2:1922 weakly excites
- L4:1456 weakly excites
- L2:2192 weakly suppresses
- L4:2175 weakly suppresses
- L4:299 weakly excites
Feeds later layers
- L19:1419 weakly excites
- L8:2828 weakly excites
- L13:1308 weakly excites
- L18:390 weakly excites
- L19:1947 weakly excites
- L10:868 weakly excites
- L26:423 weakly suppresses
- L26:172 weakly excites
- L8:2749 weakly excites
- L15:2091 weakly excites
- L25:477 weakly excites
- L16:262 weakly suppresses
- L21:175 weakly excites
- L14:549 weakly excites blank spot right before a number
- L27:1036 weakly excites
- L16:1717 weakly suppresses
- L17:222 weakly excites
- L27:628 weakly excites
- L10:1081 weakly excites
- L14:616 weakly excites
- L16:120 weakly suppresses
- L16:92 weakly suppresses
- L16:64 weakly suppresses
- L25:333 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
- ponsored
- ledged
- 九十
- _ENCOD
- lesi
- stamp
- 提出的
- 万名
- цикл
- 预定
Suppresses
- acle
- inalg
- Freund
- nard
- keit
- Constant
- geo
- Constants
- imore
- fug