ampersand before last author in citations
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. 13 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 "&" or "and" that joins the last two author names in a academic reference, like "Pyszczynski, Solomon, & Greenberg". It is mostly quiet elsewhere. Turning it up gave rambling text that did not clearly connect to citations.
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.
-
helps improve writing competence.⏎Dyson, A. (2008). Playing with textual toys. In Flood, J., Heath, S.B., & Lapp, D. (Eds.), Handbook of research on teaching literacy through
-
L., Allison, A. & James, R. 2016. Dorcopsis atrata.
-
addus VC, Light RW. Pleural effusion. In: Mason RJ, Murray JF, Broaddus VC, et al., eds. Murray and Nadel's Textbook of Respiratory Medicine. 5th ed.
-
ifying fears, and creating dependency. Reminding people of their mortality tends to make them hold more closely to traditional culture (Pyszczynski, Solomon, & Greenberg, 2003); this has implications for mental health,
-
man & Horowitz, 2003). Studies find that people do not treat computers as people outside the mediation context (Goldstein, Alsio, & Werdenhoff, 2002) – just as people do not
-
, 2018, the Trump administration reversed the policy, just eight days after the photo of Yanela was taken (Shear, Goodnough & Haberman
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: I think I have to I have I have I have I have I have I have I have I have I have I have i have 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.
-
“We Were Soldiers Once . . . And Young,” by Lt. Gen. Harold G. Moore and war correspondent Joseph G
-
to read. Reading is fascinating. It's full of surprises. And it will take you places you can't get to in any other way.⏎-
-
:<|endoftext|>The commerce of the ancient world was gathered into the warehouses of Tyre. "Tyrian merchants were the first who ventured to navigate the Mediterranean waters; and they founded their colonies on the coasts and neighbouring islands of the Aegean
-
see the Glossary.<|endoftext|>Madison is preparing for battle against the emerald ash borer, a beetle that has laid waste to millions of ash trees. And city officials are asking for residents' help.⏎A native of eastern Asia, 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
- L4:57 suppresses
- L9:92 suppresses "and" or "or" joining two items
- L5:548 excites
- L10:74 excites
- L12:51 excites
- L1:1250 weakly suppresses
- L0:313 weakly excites "and" joining paired nouns
- L9:1408 weakly suppresses
- L6:622 weakly excites
- L11:1457 weakly suppresses
- L6:527 weakly suppresses
- L1:1710 weakly suppresses
- L0:2266 weakly excites
- L10:1389 weakly excites
- L0:1656 weakly suppresses
- L6:1955 weakly suppresses
- L6:792 weakly suppresses
- L9:619 weakly excites
- L7:2778 weakly suppresses
- L11:1917 weakly excites
- L9:699 weakly excites
- L6:1313 weakly excites
- L3:2973 weakly suppresses
- L10:2432 weakly excites
Feeds later layers
- L16:2442 excites
- L17:2766 weakly suppresses
- L14:2054 weakly suppresses
- L14:1116 weakly excites 'and' inside lists
- L17:727 weakly suppresses
- L16:2049 weakly suppresses
- L17:3021 weakly suppresses
- L23:96 weakly excites
- L16:980 weakly suppresses
- L21:639 weakly suppresses
- L27:2512 weakly excites
- L16:376 weakly suppresses 'or' joining alternatives
- L22:1157 weakly excites
- L20:964 weakly suppresses
- L17:1450 weakly excites
- L18:1755 weakly excites
- L19:408 weakly excites
- L20:2360 weakly suppresses
- L17:293 weakly suppresses
- L15:566 weakly suppresses
- L17:1085 weakly excites
- L18:1625 weakly suppresses
- L19:955 weakly suppresses
- L15:3011 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
- gable
- oz
- oir
- o
- entieth
- icky
- pter
- ]
- respects
- respect
Suppresses
- 不管
- Ethics
- 笼
- NewLabel
- ادات
- 星座
- alot
- роб
- traî
- lems