name fragments split mid-word
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. 4 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 about parts of people's names that get split into pieces, such as the middle of a surname like "Jaip-aul" or "Todor-ov". The same kind of name piece makes it strongly negative in other cases. The pattern is clear in the examples but the direction is mixed, and steering 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.
-
Research, Volume 1⏎Jonathan D. Jaipaul ✉️ Department of Environmental Studies. Faculty of Earth and Environmental Studies. Turkey
-
virus from my computer, and they refused."⏎But despite these and other experiences, Vladimir says he's determined to carry on spreading the message about HIV.⏎Jaroslav Jedlicka, AIDS Programme Manager at Prague's National Institute of Public Health,
-
select mode and select all those extra loose segmenets and delete them. Hmm it looks like a rounded pipe to me . Lets<|endoftext|>Jonathan D. Jaipaul, Devya Hemraj & Samantha Providence-Forrester⏎Published: November
-
The use and acceptance of pesticide is far from uniform. Most cities are planning a combination of removal and treatments. That includes Richfield, which is now in its fourth year of pesticide treatment, and Shoreview,
-
shocked students laughed at the blurry photograph.⏎"The jurors didn't have that reaction. They thought it was actually possible from this distance to identify a person," Todorov said. "The fact of the matter is, from this distance, you can
-
; Nordström et al., 2017; von Arx et al., 2012). Floral humidity is created by a combination
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 question is, or none of the things?The answer is: the question is, or none of the things? Answer: the question is, or none of the things? The answer is: the question is, or none of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
--- have a negative reputation that may not be deserved.⏎Nicholas DiFonzo has been studying the nature of rumors for more than 15 years. Di
-
: Pure and Applied Optics, published by the Institute of Physics in London.⏎Giulio Fanti, a professor of mechanical engineering at the University of Padua,
-
The Value of Everything: Making and Taking in the Global Economy by Mariana Mazzucato. Copyright © 2018. Available from PublicAffairs,
-
ica began under Pope Julius II, in 1506 and was completed in 1615, under Pope Paul V.⏎- Donato Bramante was to the first chief architect of St. Peter's Basilica. Many famous
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
- L3:31 excites
- L2:1408 excites
- L2:2519 excites
- L3:658 weakly suppresses
- L3:2130 weakly suppresses
- L3:275 weakly suppresses
- L1:2420 weakly suppresses
- L2:75 weakly suppresses
- L3:2714 weakly excites
- L1:2683 weakly excites
- L1:1464 weakly excites
- L1:2160 weakly suppresses
- L3:2765 weakly excites
- L1:2244 weakly excites
- L2:2291 weakly excites
- L3:530 weakly excites
- L3:464 weakly excites
- L1:1227 weakly suppresses
- L1:2342 weakly excites
- L3:216 weakly suppresses empty slot before a figure number
- L1:2213 weakly suppresses
- L2:1803 weakly excites
- L0:1224 weakly suppresses
- L2:1 weakly suppresses
Feeds later layers
- L9:1086 suppresses
- L7:1184 suppresses
- L6:1067 suppresses
- L17:2205 suppresses
- L21:711 suppresses
- L21:751 suppresses
- L22:245 weakly suppresses
- L25:920 weakly suppresses
- L5:1165 weakly excites
- L25:97 weakly suppresses
- L27:1352 weakly excites
- L23:352 weakly suppresses
- L26:637 weakly suppresses
- L26:869 weakly excites word before him or he
- L27:2018 weakly suppresses
- L27:929 weakly suppresses
- L9:2098 weakly suppresses
- L5:2384 weakly suppresses
- L25:555 weakly excites
- L22:654 weakly suppresses
- L10:1826 weakly suppresses
- L11:416 weakly suppresses
- L7:598 weakly suppresses
- L20:1251 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
- ortion
- orgh
- uggested
- buried
- ác
- ROWN
- 用地
- {}).
- speech
- ency
Suppresses
- |--
- 微信号
- zza
- 经查
- Rosenstein
- isos
- 彻
- estring
- 已经超过
- смотреть