lowercase m opening a word fragment
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. 27 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 lowercase 'm' that begins the middle or end of a word, as in 'mussels', 'mucus' or 'molar'. Pushing it up or down changed the model's output into rambling or repeated text, which does not reveal much about the letter.
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.
-
population of this fascinating creature and they are currently balancing on a knife-edge.⏎Populations are found in some of the rivers flowing into Loch Ness.⏎Freshwater pearl mussels are very slow-growing and live at the bottom of clean, generally fast-flow
-
Sanskrit.⏎|karsharpan / karsha (coin)||cash|⏎|mukh (face)||Mouth (old english muth)|⏎|kapal (head)||as in per “capita”,
-
. This process is more formally known as tearing apart. Place your book on a flat surface, and with a sharp knife, cut right through the endsheet, mull and the covering material as depicted in figure 180. While doing
-
.95 million years ago rekindle debate over whether our own genus, Homo, split into multiple lineages early on.⏎- Analysis of tartar, molar wear and tooth chemistry in the nearly two-million-year-old hominin known as
-
ia and finally home to Cartage.⏎The flies we buzzing merrily around his bed. As he tiredly met the dawn, he thought
-
consequences. Egan provides a chilling account of how sea lamprey, zebra and quagga mussels and other invaders have made their way into the lakes, decimating native species
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: Whenineededineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineedineed
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
: the German-Fascists invaded the territory of the country, a war began. Young boys and girls-the same age as the current students of KSMU-went to the front as soldiers, nurses, doctors, sacrificing themselves and defending their
-
The COVID-19 pandemic has worsened problems that were already bedeviling Nigeria’s health system, including run-down or inadequate medical equipment, underpaid and unmotivated medical staff, and unequal access to health care (Andolu Agency,
-
1986 because of the Chernobyl nuclear disaster; Centralia in Pennsylvania, cleared out over 20+ years due to an underground coal-mine fire that’s been burning since 1962, which has made
-
OH and South Bend, IN established all-time records for those locations. Until this date, Erie, PA, was one of the few non-mountainous spots in<|endoftext|>The New Student's Reference Work/Alcinous⏎
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
- L22:1138 suppresses
- L24:1190 excites
- L21:1679 excites
- L17:230 excites
- L21:505 suppresses
- L22:2646 excites
- L20:234 suppresses
- L20:417 excites
- L20:1120 excites
- L20:87 excites
- L18:133 excites
- L10:311 weakly suppresses
- L7:922 weakly suppresses
- L7:291 weakly excites
- L19:251 weakly excites
- L1:333 weakly excites
- L16:34 weakly suppresses
- L19:381 weakly excites word-start 'acc' fragments
- L5:1642 weakly suppresses
- L13:2993 weakly excites
- L21:463 weakly suppresses
- L9:1834 weakly suppresses
- L16:255 weakly excites
- L0:73 weakly excites
Feeds later layers
None: this is the last layer.
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
- UsersController
- Partition
- роб
- Partition
- behalf
- writeln
- icularly
- pours
- PLIED
- pping
Suppresses
- anch
- itia
- oras
- ine
- 带
- ounter
- Lİ
- ishly
- ix
- ito