'as' after seen, regarded, described
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. 1 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 reacts to the word 'as' when it follows a verb like 'seen', 'regarded' or 'described', as in 'seen as a problem'. Turning it up made the model write in a more general, wandering way, with no clear link to that word.
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.
-
Juan Jesus Campos and six other people were killed at the Guadalajara, Mexico, airport in what was described as a shootout involving drug gangs.<|endoftext|>Osteogenesis imperfecta (OI), also
-
stability of Australia’s electricity grid are unlikely to be made before then, large-scale electrical storage could very well have a significant role to play.⏎In what he hailed as “a
-
We have to learn how to ask questions. Make sure we leave with the understanding of what we're being told."⏎FR=EE is tackling what's been seen as a longstanding concern.⏎In 1999, the Office of Minority Health
-
he had another calling. He writes the following:⏎<|endoftext|>American music history is considered as one which has undergone tremendous transformation. In other words, what is being seen as the American music today is totally different and not close to what used to be American
-
left Germantown for Virginia because of their dissatisfaction with Lutheranis, whic was non-existent in the Colony of Pennsylvania. Matthais's name is sometimes seen as Castler. He was granted a patent of 300 acres dated September
-
. He has said:⏎“Allah does not love those who sleep too much, and do not work”.<|endoftext|>ANN ARBOR, Michigan -- Cancer is often seen as a disease of genetic changes. But one type of childhood brain tumor has stym
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 have been working on with my family, and I have been working on with my family. I have been working on with my family. I have been working on with my family. I have been working on with my family. I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cacao biomass – from living cacao seeds, tree, or pod tissue – and grows from them, under carefully controlled conditions, an intermediate substance (called callus) that forms the basis of the production of the cacao analogue. At present
-
can be programmed.⏎Our model no. is 884 with a serial no. of 00139⏎Kindly Donated by Callum Wilson.⏎With thanks to http://www.dvq.com/oldcomp/m
-
acid) in 1934. The use of IAA led to some success with tomato root culture (1934), and with tobacco callus development (19
-
cent.; British Mus.; see Holy Island). Carolingian minuscule script and its splendid and complex derivative, known as Gothic, were the principal calligraphic styles from the 9th to the 14th cent.⏎The
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
- L0:1224 suppresses
- L0:1301 suppresses
- L0:178 suppresses
- L0:1001 weakly suppresses
- L0:1868 weakly suppresses
- L0:1049 weakly suppresses
- L0:450 weakly suppresses
- L0:1414 weakly excites
- L0:829 weakly excites
- L0:1503 weakly suppresses
- L0:558 weakly suppresses inside all-caps abbreviations
- L0:1475 weakly excites
- L0:33 weakly excites
- L0:1266 weakly excites
- L0:1241 weakly suppresses
- L0:636 weakly suppresses
- L0:1663 weakly excites
- L0:1374 weakly excites
- L0:2109 weakly excites
- L0:2367 weakly suppresses
- L0:2798 weakly suppresses
- L0:1109 weakly suppresses
- L0:69 weakly suppresses
- L0:953 weakly excites
Feeds later layers
- L21:2790 weakly excites
- L7:2532 weakly excites
- L6:1417 weakly suppresses
- L10:2825 weakly suppresses
- L20:1807 weakly suppresses
- L14:2650 weakly excites
- L9:1065 weakly excites
- L20:1718 weakly suppresses
- L10:417 weakly excites
- L10:2582 weakly suppresses
- L3:2808 weakly suppresses
- L11:2842 weakly suppresses
- L13:2392 weakly excites
- L15:1485 weakly suppresses
- L12:2134 weakly excites
- L14:1172 weakly excites
- L11:2623 weakly suppresses
- L20:2684 weakly suppresses
- L27:2357 weakly suppresses
- L12:720 weakly excites
- L10:1241 weakly excites
- L20:2303 weakly suppresses
- L6:2714 weakly suppresses
- L11:3058 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
- lier
- Primitive
- aña
- pit
- -post
- conquer
- chu
- -pro
- azz
- "["
Suppresses
- 设计方案
- ManagerInterface
- ץ
- 我爸
- .objects
- タイミング
- backgroundImage
- amodel
- ationale
- DialogContent