the '3' in '3D' and '2D'
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. 23 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 the number before a 'D', as in '3D' or '2-D' for three-dimensional things. When turned up, the model's replies filled with a repeated capital D, so it is tied to the letter D after a number.
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.
-
separate GPU is less important than with more complicated or moving images.⏎In video games, a separate graphics card helps to turn a set of instructions into a complex 3D image with surface textures and shading that are constantly changing. The 3D
-
and some ways it can be used in the classroom. Augmented Reality (AR) is the combination of reality and virtual reality, and allows you to put 3D content and images into the real world. AR has many applications in various fields
-
laser scanning.⏎In the overall Scan-to-BIM process, a laser scanner is used to acquire a precise 3D scan of the project’s real-world conditions.⏎The scan data is then loaded
-
IM creates a data-rich, linked, and constructible workflow. Having precise measurements to work with is essential for creating accurate models. The accuracy of measurements and 3D model content in BIM projects has reached new heights thanks to 3D
-
our right when it starts and know its next move, maybe we can also stop it from growing at all.⏎EMBL scientists have created a new, realistic 3D testbed that could help them achieve that goal by studying cancer cells as they
-
US decarbonization target. The Ministry of Energy’s solar energy research shows that the new installed capacity in the next five years<|endoftext|>Holograms help build 3D nanostructures⏎Mar 1, 2002⏎A new
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 use of232222222222222222222222222222222222222222222
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
takes place in the mid-18th century, Rzewuski created a gallery of types from the Polish gentry—bullies, drunkards, squabblers, and ignoramuses from Prince Radziwiłł’s ret
-
extreme makeover. Back in the 1970s, squatters, hawkers and manufacturing industries crowded the banks of the river, leading to
-
about so they can see the animals in their native habitats, rather than simply pictures in a textbook.⏎The workshops, which are funded by the Bristol-Myers Squibb Center, are offered free to the teachers. "The goal of the
-
, and established farmers and artisans known as Grensboere, or Border Farmers. Amongst the Voortrekkers were poor men too belonging to the squatter or bywoner class. Together these groups were later called Voortre
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
- L17:128 excites
- L22:747 excites
- L22:332 suppresses
- L19:723 excites
- L15:12 excites
- L4:205 weakly excites the digit 2 inside citations
- L19:258 weakly suppresses
- L20:225 weakly suppresses
- L5:1101 weakly excites
- L7:229 weakly excites first digit of a split number
- L18:1490 weakly suppresses
- L0:2289 weakly suppresses
- L8:463 weakly excites
- L7:556 weakly excites
- L18:569 weakly excites
- L12:713 weakly suppresses
- L8:20 weakly suppresses
- L13:33 weakly excites
- L2:49 weakly suppresses
- L20:1823 weakly suppresses
- L13:545 weakly suppresses 'of' or 'to' linking two things
- L7:303 weakly suppresses
- L16:1220 weakly excites
- L19:2307 weakly excites
Feeds later layers
- L27:552 strongly excites
- L26:2996 excites
- L27:343 excites
- L26:2168 excites
- L26:901 excites
- L25:3008 excites
- L25:224 excites
- L25:279 excites
- L25:617 excites
- L24:371 excites
- L25:313 excites
- L25:543 weakly suppresses
- L27:294 weakly excites
- L27:1202 weakly excites
- L26:1217 weakly excites
- L27:1680 weakly excites
- L27:1440 weakly excites
- L25:379 weakly suppresses
- L25:2657 weakly excites
- L25:232 weakly excites
- L27:1023 weakly excites
- L25:1244 weakly suppresses
- L25:138 weakly excites
- L26:834 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
- D
- D
- -D
- d
- Ds
- D
- 𝐷
- *D
- [D
- ,D
Suppresses
- lops
- zon
- 2
- lop
- 二十四
- 3
- tps
- neh
- 1
- apons