dash inside a number range
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. 13 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 the dash or hyphen that sits between two numbers, such as a time range (4pm-9pm), a pair of years (1338-1573) or a span of dates. When it was turned up, the model's replies drifted into unrelated, rambling text, so the steering does not show a clear effect.
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.
-
be impacting these fisheries. They will be sharing the results of this project during two upcoming community events.⏎When: Friday, July 27, 4pm-9pm⏎Where: Dolly’s Hall⏎What: Naknek community members
-
former was the native faith of Tibet, while the latter had been introduced by Tibet’s first emperor, Songtsen-gampo (r. 617 – 649). According to traditional Tibetan accounts, there was much rivalry between
-
by Marshal Ferdinand de Marsin (1656-1706) and the Elector of Bavaria Maximilian II (1679-1726). They had at least 60,0<|endoftext|>Per
-
bring subsistence foods to share!⏎When: Tuesday, July 31, 3pm-7pm⏎Where: Bristol Bay Native Association⏎What: Researchers will talk about
-
a defender of marriage (Image: Alexander and his “friend” Hephaestion, from the Getty). The lawmaker writes:⏎Alexander the Great (356-323 BC) married a Bactrian woman – modern day⏎Af
-
and his successors Orkhan (ruled 1326-59) and Murad I (ruled 1359-89) were western Asia Minor and southeastern Europe, primarily the Balkan Peninsula
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 have been writing about writing as a child, as a child, as a child (as a child, as a child, as a child, as a child, as a child, as a child, as a child, as a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
) 2 √2⏎What is the centripetal force acting on a car of mass 2000 Kg turning around a circular bend of radius 50 meters at the constant speed of 36 km/hr?⏎A
-
oltage AC power supply⏎- Two capacitors, 0.1 µF each, non-polarized (Radio Shack catalog # 272-
-
was aimed to investigate the patterns of liana-tree interactions in tropical dry evergreen forest of Point Calimere Wildlife Sanctuary (PCWS), south- east India. This study also analyzes the impact of various tree functional
-
conuts and banana leaves<|endoftext|>const void * memchr ( const void * ptr, int value, size_t num );⏎void * memchr ( void * ptr, int value, size_t num
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
- L11:79 suppresses
- L8:144 suppresses
- L10:224 suppresses
- L10:392 excites
- L12:1008 suppresses
- L10:406 excites
- L11:1128 excites
- L7:248 excites
- L9:981 weakly excites
- L8:288 weakly suppresses
- L12:242 weakly suppresses
- L12:1672 weakly excites
- L11:1261 weakly excites
- L11:525 weakly suppresses parentheses and code fragments
- L9:1084 weakly suppresses
- L12:829 weakly excites
- L10:195 weakly excites
- L2:728 weakly suppresses
- L1:700 weakly excites
- L11:575 weakly excites
- L11:219 weakly suppresses
- L10:44 weakly excites
- L11:51 weakly excites
- L9:677 weakly excites
Feeds later layers
- L15:167 weakly suppresses
- L24:25 weakly suppresses
- L27:590 weakly excites
- L17:109 weakly suppresses
- L17:3052 weakly excites
- L16:213 weakly suppresses
- L17:1255 weakly excites
- L25:2555 weakly excites
- L17:2474 weakly excites
- L15:2819 weakly excites
- L22:1445 weakly excites
- L21:2569 weakly excites "that" before a need or must clause
- L15:208 weakly suppresses
- L14:227 weakly suppresses
- L22:2387 weakly excites
- L22:400 weakly excites
- L15:363 weakly suppresses
- L17:1129 weakly suppresses
- L15:1452 weakly excites
- L14:1360 weakly excites
- L27:2338 weakly excites capital P starting a name
- L27:2701 weakly excites
- L15:2385 weakly excites
- L17:397 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
- 让她
- 但她
- ��
- aña
- uffs
- réuss
- antha
- Inherits
- 对她
- herence
Suppresses
- (
- ly
- ab
- `⏎⏎
- igs
- obs
- abs
- .
- sample
- cn