inside all-caps abbreviations
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. 0 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
This neuron fires on a letter in the middle of a short capitalised code or abbreviation, such as the 'W' in BTW or the 'M' in CDM. Turning it up did not make the writing about any single topic.
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.
-
BTW € 125,00<|endoftext|>Hell Maximillian (172
-
the Space Science Institute in Boulder, Colo.⏎For more information about the Cassini-Huygens mission visit http://saturn.jpl.nasa.gov/home/index.cfm. The Cassini imaging team homepage is at http://ciclops.org.
-
x 638 mm; watermark: Atlas holding the celestial spheres; slightly frayed and darkened edges, else in fine condition.⏎Incl. BTW € 151,25⏎Excl.
-
1,700.<|endoftext|>Risk assessment for construction⏎The Construction (Design and Management) Regulations (CDM Regulations) are intended to ensure that health and safety issues are properly considered during a
-
power point. Figure 3 shows the operating range for PWM controllers.⏎Figure 2: PWM Charge Controller⏎Figure 3: Power Curve for a PPV Panel⏎MPPT Charge Controllers⏎MPPT or the Maximum Power Point Tracking provides
-
CPSC. Their 2011 Fireworks Annual Report noted that children younger than
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 need to make the next time to be in the game of the game of the world. " I need to do something, I need to do the next. " So, I need to do the next. " So, I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. You can export a word list with a word frequency count to import into other visualization tools. http://www.antlab.sci.waseda.ac.jp/software.html⏎Voyant Tools: Voyant Tools is a free web based suite
-
& Co. (04/01/2000), and Others|⏎The novel opens in form of a letter by Wiliam Crimsworth to his school friend, describing his life after he left the school. The first
-
transatlantic networks that wove the exiles of African descent, the gens de couleur, together across the Atlantic. Exploring the careers of Joseph and Sévère Courtois provides a window
-
vère penned the letter urging his brother to support his universal liberation enterprise, Joseph had just come back from fighting in the Napoleonic wars in Europe. Sévère had participated in multiple revolutionary coups and moved from New Orleans to Cartag
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
None: this is the first layer.
Feeds later layers
- L2:50 suppresses
- L1:1031 suppresses
- L4:1595 weakly suppresses
- L6:1854 weakly suppresses
- L12:163 weakly excites
- L1:332 weakly suppresses
- L19:1584 weakly suppresses
- L2:2174 weakly suppresses
- L1:156 weakly suppresses
- L13:1643 weakly suppresses
- L24:2826 weakly suppresses
- L1:1707 weakly suppresses
- L17:419 weakly suppresses
- L3:189 weakly suppresses
- L10:1620 weakly excites
- L1:1753 weakly suppresses
- L1:2450 weakly suppresses
- L5:15 weakly suppresses
- L19:342 weakly excites
- L3:1488 weakly suppresses
- L24:2621 weakly suppresses
- L25:1486 weakly suppresses
- L1:3055 weakly suppresses
- L22:2814 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
- następ
- şek
- ôte
- _note
- 它
- anguage
- it
- dön
- ingu
- arn
Suppresses
- 网讯
- 大的
- ístico
- 米尔
- allas
- リー
- 闭
- 个项目
- isons
- misplaced