comma after a speaker tag
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. 18 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 responds to a comma that comes right after a person's name or a "said" phrase, such as "said Markram," or "Halloran said,", where a quote or a clause is about to continue. It is pushed off by commas in long, formal old-style sentences. Turning it up gave wordy, rambling text without a clear theme.
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.
-
create models to explore how neural circuits are organized, and how they give rise to behaviour and cognition — among the deepest mysteries in neuroscience. Ultimately, said Markram, the HBP would even help researchers to grapple with disorders such as Alzheimer’s
-
.⏎The Cherokee presented an interesting dilemma. It was declared, with a sigh, that most Indians would be removed due merely to the<|endoftext|>Editorial: Team effort
-
ranging from yeast to fruit flies to mice to humans. The idea, said study researcher Donard Dwyer, is that in people with a broken starvation response, a few initial rounds of diet
-
be compared to a wise man who built his house on the rock……⏎Revelation 3:20 'Behold, I stand at the door and knock; if anyone hears My voice and opens the
-
.”⏎Not surprisingly, notes Lo, as the dose of THC increased, so did the impact to menstrual cycle length and
-
protect against enterovirus infection, which can cause high fever, diarrhea and sometimes death.⏎Similarly, they say, there is a strong correlation between rheumatoid arthritis and tuberculosis. Populations with
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: to have, as, or to be, as, or, as, or, as, or, as, or, as, or, as, or, as, or, as, or, as, or, as, or
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
III.) a terrible pestilence devastated London. The dirt and crowding of the old mediæval cities made them at all times nurseries of infectious disease, and when a great
-
among savages than among civilised men. Notwithstanding the differences which necessarily exist between savage and civilised ethics, the two systems must evidently have much in common, for from the days of Columbus to the present, travellers have given testimony of customs
-
effected by Themistocles who had meanwhile created the Athenian fleet which at Salamis proved its superiority over the Perso-Phoenician armada, anc thus precluded beforehand the success of the land-forces.<|endoftext|>The Treasure
-
uary will receive from guests more attention than will be accorded to these ancient weapons.⏎No doubt if one could procure a shell strung with gold and silver cords, after the fashion of the old-time instrument with which the gods made music, 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
- L10:315 excites
- L4:90 weakly suppresses
- L3:1758 weakly excites
- L10:2217 weakly suppresses
- L15:995 weakly suppresses
- L12:35 weakly suppresses
- L1:1007 weakly excites name pieces inside a multi-part word
- L14:1832 weakly excites
- L7:2925 weakly suppresses
- L13:44 weakly excites inside web addresses and citations
- L15:2372 weakly suppresses
- L5:2342 weakly suppresses
- L10:1042 weakly suppresses
- L12:1968 weakly suppresses
- L15:930 weakly suppresses
- L16:245 weakly suppresses
- L2:1207 weakly suppresses
- L16:775 weakly excites
- L16:1529 weakly excites
- L1:2910 weakly suppresses
- L8:1400 weakly excites
- L4:726 weakly excites
- L8:2822 weakly excites
- L4:76 weakly suppresses
Feeds later layers
- L21:344 excites
- L19:625 weakly suppresses
- L19:1034 weakly suppresses
- L22:368 weakly suppresses
- L19:1172 weakly excites
- L19:1414 weakly suppresses
- L21:632 weakly suppresses
- L19:1310 weakly excites
- L22:158 weakly excites
- L26:462 weakly suppresses
- L23:1145 weakly excites
- L25:308 weakly suppresses
- L20:116 weakly excites
- L19:1803 weakly excites
- L21:906 weakly excites
- L21:2161 weakly excites
- L20:596 weakly excites
- L19:482 weakly excites
- L20:1391 weakly suppresses
- L22:786 weakly suppresses
- L20:2133 weakly suppresses letters inside spelled-out words
- L22:1024 weakly excites
- L21:2019 weakly suppresses
- L19:1558 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
- REAT
- yy
- 的一款
- igor
- -word
- 心
- 你就
- ị
- نفس
- odon
Suppresses
- 这批
- 起来了
- ठ
- брос
- gemäß
- udence
- etti
- dings
- .social
- etty