commas inside numbers and asides
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. 15 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 commas that sit in the middle of a bit of extra detail, such as a birth place after a date, or a second name in brackets, and also to the thousands comma in big numbers. It is weakest and most negative on numbers like 350,000. Turning it up or down gave text that does not clearly relate to commas.
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.
-
maintaining change that’s hard—creating new and sustained ways of thinking and behaving. As Mark Twain quipped, “Quitting [smoking] is easy, I’ve done it dozens of times.” Many can begin a positive health behavior change
-
and publish his work.⏎What makes Robert Frost unique?⏎Robert Frost, in full Robert Lee Frost, (born March 26, 1874, San Francisco, California, U.S.—died January 29,
-
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
-
always used to help one another<|endoftext|>Ramses III, Ramses also spelled Ramesses or Rameses (died 1156 bce, Thebes, Egypt), king of ancient Egypt
-
is a growth hormone. So that does not help if you’re overweight, and it just becomes a vicious cycle.⏎Right. And as a former powerlifter, I often heard people say how insulin is the most powerful hormone in the human body
-
patient. There is no universal description of counterfeit drugs, but basically, it is a global phenomenon. Counterfeited drugs are dangerous not because of their counterfeit nature, but because of how they are created.⏎What are counterfeit drugs?⏎A counterfeit drug
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: If I had, I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I... I...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in Africa.⏎In a second report, the international team reveals previously unknown traces of a major eruption of Hecates Tholus less than 350,000 million years ago. In a depression on the volcano, researchers found
-
much of the islands and the average rainfall is in upwards of 150″ per year. Temperature varies little, with the average temperature near 82, and the humidity is always high.⏎Figure 2. Yum! After a
-
vil cells, at pressures between 938,000 and 997,000 times atmospheric pressure (
-
already signed up, you need to do it soon. Failure to do so is a felony punishable by up to five years in prison or a $250,000 fine.⏎And the system has not been updated since women were allowed
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 suppresses
- L9:7 excites
- L9:77 excites commas in citations and name lists
- L14:51 weakly excites colon that introduces a statement
- L10:2069 weakly excites
- L12:1557 weakly suppresses comma after an introductory phrase
- L14:348 weakly suppresses
- L9:16 weakly suppresses unclear: periods inside numbers and names
- L4:673 weakly excites unclear: pieces of split words
- L4:1266 weakly excites
- L14:435 weakly excites
- L13:78 weakly suppresses
- L5:5 weakly suppresses digit after a label like paper or type
- L13:345 weakly excites
- L1:226 weakly excites
- L4:3 weakly excites commas after citation fields
- L12:5 weakly excites
- L11:154 weakly suppresses commas after citation numbers
- L13:278 weakly suppresses
- L2:2846 weakly suppresses
- L2:629 weakly suppresses
- L10:33 weakly suppresses
- L3:756 weakly suppresses
- L10:144 weakly suppresses
Feeds later layers
- L16:1599 weakly excites
- L17:2001 weakly excites
- L16:2289 weakly excites 'or' linking two alternatives
- L19:1738 weakly excites
- L20:1134 weakly suppresses
- L25:147 weakly suppresses
- L16:753 weakly suppresses
- L16:2580 weakly suppresses 'and' linking food or health nouns
- L22:906 weakly excites
- L21:274 weakly suppresses
- L18:1819 weakly excites
- L19:806 weakly suppresses
- L26:1248 weakly excites
- L25:429 weakly suppresses
- L23:414 weakly suppresses
- L23:162 weakly excites
- L16:1095 weakly suppresses
- L24:1055 weakly excites
- L17:1614 weakly excites
- L17:2775 weakly suppresses
- L17:2986 weakly excites
- L25:1511 weakly excites
- L16:2614 weakly excites unclear: weak end-of-text signal
- L22:1669 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
- pter
- ktor
- !';⏎
- VG
- =";⏎
- kv
- vt
- KV
- VK
- vn
Suppresses
- 我们一起
- 我们
- Dec
- AB
- Medicine
- ️
- زال
- ound
- Ặ
- izar