mid-word 'electro' and 'ro' fragments
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. 21 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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
It reacts to pieces of longer words, especially the 'electro' in electrolyte or electrostatic and the 'ro' in styrofoam or wardrobe. The steered outputs did not show a clear theme.
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.
-
area, and/or the genitals. "LEOPARD," an acronym for such characteristic abnormalities, stands for (L)entigines; (E)lectrocard
-
water as a heat source. Movie clips of setup and convection in action are included on this page.⏎- Clear pan (plant dish)⏎- 4 styrofoam cups⏎- Food coloring⏎- Hot water (scalding hot
-
- foil take-home containers – ordered too large of a meal at a restaurant? Usually you get your leftovers in a foil container with a cardboard lid⏎- styrofoam containers – again you might get this from a restaurant for leftovers or as
-
0V Polyester Film Capacitor⏎- 470uF/25V Electrolyte Capacitor⏎- 22uF/25V Electro
-
0 degrees C).⏎Electrostatic spark discharge results from the internal molecular friction that generates high-voltage electric charges
-
rid of? Think responsibly about how you dispose of things – can it be passed on, sold, or donated? And before you go fill the presses or wardrobes again, take time. Is there a way to reuse or repurpose something
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: the study of a lot of this under a lot of this, like the study of the meaning of a lot of this under the study of a lot of this. This is like a lot of this under a lot of this. So,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
treats of the best means of directing school affairs so as to secure efficiency. It includes not only school economy, but also school discipline
-
tiny alterations to the lifestyle over a duration of time. Modifications need to be considerable plenty of to generate a obvious distinction, however not so big that your particular entire body adapts in their mind and backside out from
-
divided into revocable and irrevocable living trusts.⏎How and When Does a Living Trust Terminate?⏎The manner in which a trust terminates depends, to some extent, on the type of trust in question. If the trust is a
-
records?⏎Paper vs. Processor⏎Stealing paper records isn’t too difficult on a small scale, but thieves would probably need a truck to steal
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
- L19:1370 strongly excites
- L20:374 suppresses
- L18:112 suppresses 'over' meaning time or duration
- L16:118 suppresses
- L14:392 suppresses
- L2:1795 suppresses
- L17:213 suppresses
- L20:1964 excites
- L20:1398 suppresses
- L16:177 excites zero digit inside dates and codes
- L17:366 suppresses
- L2:7 excites
- L0:44 suppresses
- L19:1241 suppresses
- L18:569 excites
- L19:53 excites
- L20:1273 excites
- L19:598 suppresses
- L4:221 suppresses
- L18:954 excites
- L20:2652 excites
- L18:1675 suppresses
- L19:1083 suppresses
- L0:2969 weakly suppresses
Feeds later layers
- L27:1585 strongly excites
- L24:631 excites
- L22:2705 excites
- L27:3069 excites
- L23:2981 excites
- L25:3055 excites
- L26:1877 excites
- L23:864 excites
- L27:2777 excites
- L27:2742 excites 'w' inside split words
- L23:1773 excites
- L27:2960 excites
- L22:2122 excites
- L25:1006 excites
- L23:678 suppresses
- L23:2291 excites
- L26:2439 excites
- L27:1002 weakly excites
- L23:2574 weakly excites
- L26:2784 weakly excites
- L24:3020 weakly excites
- L25:826 weakly excites
- L23:2815 weakly excites
- L25:1449 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
- ffset
- ptions
- cks
- cket
- ftware
- pts
- LOGY
- ʊ
- logy
- 써
Suppresses
- mol
- andard
- 婚后
- nbr
- rush
- nak
- uan
- way
- nation
- wang