the letter r starting a word piece
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. 26 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 a single 'r' or 'A' that starts a fragment of a longer word, such as the 'r' in 'ratiocinate' or 'rancor'. When it was turned up, the model fell into repeating short phrases, with no clear topic.
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.
-
proportion.⏎5. (n.) To exercise the rational faculty; to deduce inferences from premises; to perform the process of deduction or of induction; to ratiocinate; to reach conclusions
-
death, but he only thought that this relationship between Romeo and Juliet be the cure to their families grudge for each other.⏎Their love is forbidden because of a rancor between their families. As they arrange a marriage behind everyone’s back,
-
by a group of nomads called the Aryans living in what is now Southern Russia. It is of course critical to mention the improper usage of the word "Aryan" in N
-
the prey, my son, you have gone up. He stooped down; he crouched as a lion and as a lioness; who dares rouse him? The<|endoftext|>Juvenile striped bass spawned in the Chesapeake Bay were
-
in the province of Taurida. He purchased sheep from the tsar's own flocks and also imported some rams from Saxony. At that time the average Mennonite farm had between
-
prolific; they breed at about six months old, producing four or five litters, of from four to ten young ones in each, in the year. The raptorial birds keep down their numbers in
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: like, like, like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like like
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
OR REARRANGEMENT REACTION. THIS EXPERIMENT WILL DETERMINE THE FACTORS OF ELECTROPHILIC SUBSTITUTION ON DIFFERENT AROMATIC RINGS TH
-
ЛЬНАЯ РАБОТА ПО АНГЛИЙСКОМУ ЯЗЫКУ⏎КОН
-
How Humidity Affects Flu Outbreaks⏎03-05-2009
-
the most part, you will have positive outcomes. Though the process can be long, it is still worth the steps required to get greater hearing abilities.⏎Programming Hearing Aids
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
- L23:2833 excites
- L21:1072 suppresses
- L17:278 excites 'as' in 'as well as' and 'as much as'
- L18:102 suppresses
- L6:328 suppresses
- L2:1699 excites
- L16:84 excites
- L20:174 excites
- L22:961 suppresses
- L16:628 weakly suppresses
- L19:171 weakly suppresses
- L16:8 weakly excites
- L0:49 weakly excites
- L0:2837 weakly excites
- L10:103 weakly excites
- L16:63 weakly suppresses
- L10:237 weakly suppresses
- L5:779 weakly suppresses
- L19:569 weakly excites
- L0:481 weakly suppresses
- L4:430 weakly suppresses
- L21:1233 weakly suppresses
- L20:601 weakly excites
- L20:950 weakly suppresses
Feeds later layers
- L27:2769 weakly excites capital 'K' starting a name piece
- L27:1546 weakly excites
- L27:1086 weakly excites
- L27:1082 weakly excites
- L27:1560 weakly excites
- L27:564 weakly suppresses
- L27:1090 weakly excites
- L27:793 weakly excites
- L27:602 weakly suppresses
- L27:1920 weakly excites
- L27:1488 weakly suppresses
- L27:102 weakly excites
- L27:1911 weakly excites
- L27:2490 weakly excites
- L27:1833 weakly suppresses
- L27:582 weakly suppresses
- L27:834 weakly excites
- L27:2585 weakly excites
- L27:408 weakly suppresses
- L27:2516 weakly excites
- L27:1006 weakly excites
- L27:1282 weakly suppresses
- L27:2502 weakly suppresses
- L27:1349 weakly excites
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
- лер
- iel
- réponse
- цию
- 律
- ту
- óc
- uitive
- uild
- раз
Suppresses
- 迢
- est
- curs
- ñe
- いか
- EST
- 书记
- borne
- зрения
- 컷