comma after a clause ending
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. 17 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 fires on a comma that comes right after a complete clause, where a list, an explanation or a contrast is about to follow, like "satiated, but" or "offer kid-friendly features, such as". Turned up, the model fell into repeating filler words like "and and and".
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.
-
to send people to space.⏎So if NASA’s Planetary Protection Officer—there is one—is going to live up to the coolest job title in the solar system, the space agency is going to need some different rules for keeping Earth’s bugs on
-
1939 the country stood alone against the might of the German war machine, a plucky but woefully under-equipped little island. It initially suffered setbacks, and finally won the war in 1945 thanks to the intervention of
-
miracles of productivity. Perhaps “still” is the wrong word given the improvements in technology and sheer productive capacity over the last half century. If we often seem paralyzed, unable to accomplish basic tasks, it is because of a failure of nerve and guiding
-
this amount with your child beforehand), it’s time to open an official saving account and deposit it into the bank. Some banks and credit unions offer kid-friendly features, such as online games, rewards for getting good grades, and other perks.⏎4
-
ness, and better memory and recall.⏎One simple reason for this is that ensuring you consume protein, carbs, vitamins, and fats not only keeps you satiated, but also provides you the energy you need to fuel your day.⏎An easy way
-
ier blood sugar levels and sustained energy. Also, complex carbohydrates tend to provide more vitamins and minerals than sources of simple sugars.⏎Sugar has its place in your diet, but many health and diet experts warn that excess sugar consumption can lead to several serious
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:,,,,,,,,,,,,,,,,,,,,,,,,,|||||||||||.||||.||.||.||.|.|
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Today, many keyboard and mice use the USB interface.<|endoftext|>fEMPOWER: workshopping
-
00,000) is higher than that found in most African countries and much higher than in China as a whole (53/100,000
-
2,019 waterbodies had 5 or more years of data. When the University of Florida data are added, the number jumps to 4,800 waterbodies. Of these, 17% had significant
-
salaries for electricians, at $72,030, followed by New York at $70,680, Illinois at $70,430, Hawaii at $68,430 and New Jersey
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
- L16:240 strongly excites comma after 'sometimes' or 'the truth'
- L14:35 strongly suppresses
- L9:77 excites commas in citations and name lists
- L11:154 suppresses commas after citation numbers
- L4:3 excites commas after citation fields
- L10:335 excites commas in catalog, date, and list metadata
- L5:16 excites
- L15:588 excites comma after an introductory phrase
- L3:71 suppresses
- L16:241 excites
- L0:68 suppresses
- L0:146 suppresses
- L16:846 excites comma opening an aside or explanation
- L16:710 excites
- L1:86 excites
- L15:123 weakly suppresses
- L16:1082 weakly suppresses comma linking two clauses after 'and'
- L2:66 weakly excites
- L16:1939 weakly excites line break mid-sentence in wrapped text
- L12:1557 weakly suppresses comma after an introductory phrase
- L5:91 weakly excites
- L16:762 weakly suppresses
- L16:432 weakly excites period after month or initial
- L0:346 weakly excites comma closing an introductory phrase
Feeds later layers
- L19:726 strongly excites comma inside large numbers
- L18:366 strongly excites comma inside numbers and dates
- L23:184 strongly excites
- L20:476 excites commas that lead into a continuation
- L19:383 excites commas after a clause or list item
- L22:144 excites
- L24:902 excites
- L20:1298 excites
- L19:1826 excites
- L19:2645 excites commas in the middle of a sentence
- L18:1394 excites
- L18:2043 excites
- L27:77 excites
- L18:2863 excites
- L20:2082 suppresses
- L19:1506 excites
- L18:1270 excites
- L21:50 excites comma after 'So' or intro phrase
- L21:700 excites comma before a new clause
- L19:1598 excites comma before a qualifying clause
- L26:94 excites
- L25:535 excites commas inside number lists
- L18:227 suppresses
- L27:139 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
- especially
- particularly
- albeit
- which
- particularly
- 特别是在
- 特别是
- namely
- specifically
- 以及
Suppresses
- ,
- ,M
- -on
- -date
- date
- -help
- рат
- MORE
- -stack
- -p