unclear: syllables inside Sanskrit religious words
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. 22 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 lights up on a fragment in the middle of a Hindi or Sanskrit-derived word, such as the 'ak' in Mahakavya or the 'ad' in Ekadasi. It is not clear what ties these together, and turning it up made the model's text wander and repeat.
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.
-
cooperation.<|endoftext|>Kamayani (Hindi:कामायनी) (1936) is a Hindi epic poem (Mahakavya) by Jaishankar Prasad (1889–
-
and form a halo around her. In front of the goddess are three Mahakalas (enlightened beings who protect those on the path to Buddhah
-
of the Buddhist deity Chakrasamvara⏎- The Buddhist deity Mahakala as a Brahman⏎- The Buddhist deity Simhavaktra Dak
-
n “Souramana” during the auspicious month of Dhanurmasa (Margasira Poushya), out of 24 Ekadasi Tidhis in a year Suklapaksha Ekadasi is termed
-
.⏎Metsäpohjainen piha paransi päiväkotilasten immuunijärjestelmää jo kuukaudessa<|endoftext|>Research Looks at Alternative Power for Military⏎Source Newsroom: South
-
on this day it is further called “Mukkoti Ekadasi”.⏎It is even named as Vaikuntha Ekadasi as
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: 50000000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, which means to combine. Our ultimate goal is union with our eternal, transcendent self.⏎There are many schools of thought on yoga. Some of them contradict one another. There are several schools of thought about yoga. please click the up coming
-
the trees and sang so sweetly that the children used to stop their games in order to listen to them. ‘How happy we are here!’ they cried to each other.⏎One day the Giant came back. He had been to visit his friend
-
objective nature of reality, but primarily predicated on the the fact that humans subjectively interpret the world in terms of objects (vertices) and their respective relationships to one another (edges) (an argument against this idea). The popular data model used
-
studied classical Indian texts such as the Dharmaśāstra and inscriptions make reference to Sanskrit literature. Eventually, while the Cham and Sanskrit languages influenced one another, Cham culture assimilated Hinduism, and Chams were eventually able adequately
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
- L20:2919 suppresses
- L20:952 suppresses
- L19:2162 excites
- L21:1816 excites
- L21:1049 suppresses
- L21:2134 excites
- L21:2296 excites
- L21:1226 suppresses
- L19:1754 excites
- L21:784 suppresses
- L20:2070 excites
- L19:2606 excites
- L19:2104 excites
- L20:1798 suppresses
- L21:2565 excites
- L18:2007 excites
- L21:1903 suppresses
- L21:2419 suppresses
- L20:1160 excites
- L18:2772 excites
- L21:1790 excites
- L20:1794 suppresses
- L19:1933 excites
- L21:1925 suppresses
Feeds later layers
- L27:425 suppresses
- L25:210 suppresses
- L25:120 suppresses
- L25:420 suppresses
- L25:477 excites
- L24:48 excites
- L25:531 excites
- L26:511 suppresses
- L27:249 suppresses
- L26:265 suppresses
- L23:197 suppresses
- L25:334 excites
- L24:221 excites
- L25:298 suppresses
- L25:98 suppresses
- L26:248 suppresses
- L23:390 suppresses
- L25:1114 suppresses
- L23:12 suppresses
- L25:176 suppresses
- L23:122 weakly excites
- L27:348 weakly suppresses
- L27:702 weakly suppresses
- L25:1135 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
- arin
- わけではない
- <<⏎
- 就够了
- 以上的
- >" ⏎
- 的不同
- 的基础上
- 的有效
- Willis
Suppresses
- num
- ou
- tered
- num
- -choice
- OLER
- ECB
- ersonic
- ič
- ournament