commas in keyword lists
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. 3 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 commas that separate items in a list of keywords or tags, especially near the end of the list. It goes negative on the middle of some longer words. Turning it up or down did not produce a clear, consistent change in the text.
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.
-
central china, chinese government, conservation efforts, early 1970, giant panda, giant pandas, hsing hsing, ling ling, mainland china, mei xiang, mother and baby, mountain forests, national zoo, pandas
-
General technology, Google, Media, Services, Augmented reality, eyeglasses, feature, featured, glasses, Google, Google Plus, Google Project Gass, google virtual reality glasses, New York Times, Project Glass, technology, turn by
-
asymmetric radial transects with fixed orientation, but a workable solution for polygonal transects remains elusive.⏎Keywords: Boundary overlap, Inclusion region, Reflection method, Transect configuration⏎- We recommend that you also print this page and attach it
-
, Hau believes future practical applications include improved communications technology, television displays and night-vision devices.⏎Tags: Lene Vestergaard Hau, physicist, spped of light.<|endoftext|>The Real
-
Gaming, Services, Software, 3 million dollar grant, Associate Professor, Augmented reality, Bill & Melinda Gates Foundation, Common Core, Common Core standards, Education, education arcade, education program, Eric Klopfer, Filament Games
-
and reduce inequalities in household decision-making.⏎keywords malaria, treatment-seeking behavior, inequality in household decision-making, Ghana, child mortality, chloroquine, home treatment, pre-packaged antimalarial drugs, prevention⏎Malaria, a mosquito-b
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 am a good person. I have a good home. I am a good friend to the Earth. I am a good planet. I am a good... and I am a good... and I am a good... It's the first
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
elage. It was under the umbrella of American power that today even the smallest European state can enjoy its national identity and independence in unprecsented security. Domestically America in the 20th century gradually implemented a series of progressiveand
-
voltage drop waveforms are in-phase with each other.⏎Given that the power supply frequency is 60 Hz (household power frequency in the United States), calculate impedances for all components and determine all
-
bacteria Yersinia pestis, and can be contracted by humans and household animals if proper precautions are not taken,” officials from Jefferson County Public Health (J
-
, which was published in 1979, the year she died.⏎‘Impossible to understate Rhys’ importance’⏎Jean Rhys is a household name across the Caribbean due in large part to the fact that “Wide Sarg
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
- L2:733 weakly excites
- L2:2558 weakly suppresses
- L2:1036 weakly excites
- L2:2346 weakly excites
- L1:583 weakly suppresses
- L2:1657 weakly excites
- L2:41 weakly suppresses
- L2:2784 weakly suppresses
- L2:1494 weakly suppresses
- L1:649 weakly excites
- L2:1710 weakly suppresses
- L0:1046 weakly suppresses
- L1:2463 weakly excites
- L2:75 weakly suppresses
- L1:2160 weakly suppresses
- L0:895 weakly excites
- L2:929 weakly suppresses
- L1:3038 weakly excites
- L2:122 weakly excites
- L1:997 weakly suppresses
- L2:2149 weakly suppresses
- L0:1512 weakly suppresses
- L1:2605 weakly suppresses
- L1:597 weakly excites
Feeds later layers
- L12:552 weakly excites
- L18:1935 weakly excites
- L7:2713 weakly suppresses
- L16:1722 weakly excites
- L17:2618 weakly suppresses
- L6:938 weakly excites
- L16:589 weakly suppresses
- L10:2863 weakly excites
- L7:511 weakly suppresses
- L19:1315 weakly suppresses
- L10:1150 weakly suppresses
- L4:2137 weakly suppresses
- L12:471 weakly suppresses
- L4:831 weakly suppresses
- L14:1020 weakly excites
- L9:775 weakly excites
- L7:1031 weakly suppresses
- L10:865 weakly suppresses
- L6:2934 weakly suppresses
- L23:1120 weakly excites
- L4:1126 weakly suppresses
- L15:1853 weakly excites
- L11:2383 weakly suppresses
- L11:2416 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
- extent
- ológ
- deadline
- CD
- PTS
- 长大了
- 我爱你
- درك
- grade
- 幅度
Suppresses
- etwork
- pomi
- culator
- atab
- �
- idian
- iphers
- lycer
- pee
- dition