comma inside numbers and dates
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 fires on the comma in written numbers and dates, such as the comma in 100,000 or in 'January 29, 2013'. Turning it up made the model's writing drift into strings of counting and repeated number-like text.
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.
-
.S. trials before the Nuremberg Military Tribunals (NMT) took place from December 9, 1946 to April 13, 1949
-
of the mother continent between 100,000 and 60,000<|endoftext|>Know Your Limits,
-
anoma skin cancer) to estimate the number of people in the U.S. who had ever been diagnosed with cancer and who were alive on Jan. 1, 2007.⏎As of that day, 65% of
-
potential investors and vie for prizes as large as $200,000 in value.⏎Helion Energy Fusion Engine has received about $7
-
our commitment to continued energy efficiency gains, sustainability, and lower operating costs for consumers and enterprises.⏎The study further shows that an enterprise customer using 100,000 PCs running 6th Generation AMD A-Series APU could save
-
required a great deal of structural analysis to ensure it can withstand seismic forces—and curious visitors.⏎January 29, 2013—It’s plain to see that turning the remains of
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: (1) (The) (1) (The) (1) (The) (1) (The) (1) (The) (1) (The) (1) (The) (1) (The)
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
CO2 soil concentrations are due to the volcanic rumblings of the magma still active below Mammoth Mountain. At more subtle levels this research has been able to map, using hyperspectral air borne imagery, the tree plant stress over all of the Mam
-
a month later was life-changing for not just the Soviet Union and the US but for everyone on this planet,” Zelibor said. “It showed us all, we do not need to be bound by Earth’s gravity because we have an entire
-
war:⏎“Think, too, of the great part that is played by the unpredictable in war,” Thucydides wrote, “and think of it now, before you are actually committed. The longer a war lasts, the more things tend
-
the sacred isle of Columba! Here landing, he ascended the loftiest of the hills upon the isle, and "gazing into the distance, found no longer any trace
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
- L14:35 strongly suppresses
- L16:240 strongly excites comma after 'sometimes' or 'the truth'
- L17:118 strongly excites comma after a clause ending
- L9:77 strongly excites commas in citations and name lists
- L4:3 strongly excites commas after citation fields
- L5:16 strongly excites
- L11:154 suppresses commas after citation numbers
- L10:335 excites commas in catalog, date, and list metadata
- L17:585 excites
- L2:66 excites
- L3:71 suppresses
- L15:630 suppresses
- L17:1405 suppresses
- L17:1597 excites
- L15:1213 suppresses
- L16:241 excites
- L16:846 excites comma opening an aside or explanation
- L15:588 excites comma after an introductory phrase
- L12:1557 suppresses comma after an introductory phrase
- L17:116 excites
- L1:86 excites
- L16:1939 excites line break mid-sentence in wrapped text
- L17:2221 weakly suppresses
- L17:2001 weakly excites
Feeds later layers
- L19:726 strongly excites comma inside large numbers
- L19:682 strongly suppresses
- L21:700 excites comma before a new clause
- L20:476 excites commas that lead into a continuation
- L25:880 excites comma after a cause or effect clause
- L19:2645 excites commas in the middle of a sentence
- L26:236 suppresses
- L20:2082 suppresses
- L23:184 excites
- L19:383 excites commas after a clause or list item
- L22:144 excites
- L27:86 excites
- L21:632 suppresses
- L22:368 suppresses
- L19:1036 suppresses
- L20:513 excites
- L21:2616 excites
- L19:2403 excites comma after a full clause
- L27:358 suppresses
- L19:2175 excites
- L25:328 suppresses
- L20:811 suppresses
- L20:173 suppresses
- L21:530 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
- especially
- which
- StringComparison
- Inc
- YTE
- noting
- Cor
- którzy
- but
- including
Suppresses
- ,
- ,⏎⏎⏎⏎
- ,⏎⏎⏎
- 不下
- ,e
- ustin
- ,{⏎
- igraph
- ,...⏎
- ,request