unclear: document ends and number gaps
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. 15 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 is hard to pin down. It fires most on the end-of-text marker between documents, and it goes quiet on the gap right before a second year or number, such as in '1989 and [blank] 1996'. Turning it up or down changed the writing in ways that did not show a clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
' labour in a portfolio with him and rather than let the research fall into German hands, (Poland having been invaded by Germany), he has hidden it.<|endoftext|>BURKE TEACHERS STUDY, PREPARE FOR NEW CURRICUL
-
major avoidable risk factors.<|endoftext|>As the rain begins to fall on Tanzania’s Tarangire National Park, thousands
-
days.<|endoftext|>The bread glycemic index enables health conscious people to choose the right bread for the
-
Project:<|endoftext|>After years of experience and "combing through" catalogs⏎to find the best
-
software applications use 255 file type for internal purposes and/or also by different way than common edit or open 255 file actions.<|endoftext|>The Importance of Detoxification – Part 1⏎Updated: Sep 2
-
code<|endoftext|>Jet engines, or gas turbines, suck in air at the front with a fan
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, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the, the,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
findings noted above, based on research by Dr. Michael J. Murray and his associates, were published in Equine Veterinary Journal in 1989 and 1996. An even more recent study in 2005
-
William R. Pogue orbited the Earth for 84 days during the final manned flight of the Skylab space station in 1973 and 1974.⏎He was born January 23, 19
-
55°W and 35°N and 42°N. The patch extends over an indeterminate area, with estimates
-
对比分析的研究成果对泰国学生学习汉语助动词具有较大的帮助,学生能更容易、更好地把握汉语助动词的语义、语法功能与特点。auxiliary verbs have very important functions in Chinese grammar. They are frequently
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
- L13:1578 suppresses
- L6:101 suppresses
- L7:874 suppresses
- L8:410 excites
- L13:44 suppresses inside web addresses and citations
- L10:155 excites
- L6:1801 excites
- L14:375 suppresses
- L6:2477 weakly suppresses
- L11:1602 weakly excites
- L14:775 weakly suppresses
- L7:111 weakly suppresses
- L14:282 weakly excites
- L8:328 weakly excites
- L11:226 weakly suppresses
- L6:1233 weakly suppresses
- L4:96 weakly suppresses
- L0:68 weakly excites
- L8:81 weakly excites
- L10:289 weakly suppresses
- L14:37 weakly excites
- L7:6 weakly suppresses
- L10:356 weakly suppresses
- L8:93 weakly excites period after a middle initial
Feeds later layers
- L16:1112 excites
- L16:1993 excites
- L16:1059 excites
- L20:2998 excites
- L16:775 excites
- L18:1249 excites
- L20:302 suppresses
- L18:166 excites
- L19:376 suppresses
- L17:839 excites
- L18:1775 excites
- L19:2073 excites
- L17:585 weakly excites
- L24:35 weakly suppresses
- L19:386 weakly suppresses
- L17:1997 weakly excites
- L16:1315 weakly excites word or phrase being named or defined
- L19:2335 weakly excites
- L17:1772 weakly excites
- L18:227 weakly suppresses
- L19:730 weakly suppresses
- L17:219 weakly excites
- L19:1767 weakly excites
- L16:797 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
- 这才
- SpaceItem
- 就是为了
- );\
- overwhel
- avanaugh
- %;">
- ");⏎⏎⏎
- 这就是
- apk
Suppresses
- ستمرار
- Torch
- 台
- ться
- ames
- unts
- ,
- Lib
- minimum
- lice