connective before the last list item
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. 12 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 word like 'and' or 'then' that comes right before the final item in a list or the last step of a sequence. When it was turned up, the written text did not show a clear change in 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.
-
forward 2 squares. What have you found?⏎2. 1/4 turn clockwise, forward 2, 1/4 turn anti-clockwise and then forward 3 _______.⏎3. Forward 2, 1/4
-
the challenges the Tennessee schools faced.⏎West Virginia, New York, Nebraska, and Delaware also strengthened proficiency standards, while New Mexico, Washington, Hawaii, Montana, and Georgia lowered the bar.⏎Uneven at the Start, a new Education Trust report
-
During the last few days we have surveyed butter production costs in 1947 and 1948, compared with 1945 and 194⏎Labor costs in
-
these nine orders there exists three ranks: Seraphim, Cherubim, Thrones—Dominions, Authorities, Powers—Principalities, Archangels, and Angels.
-
1 chapters. Chapter 1 sets the context for RE and climate change; Chapters 2 through 7 provide information on six RE technologies, and Chapters 8 through 11 address integrative issues. References to chapters and sections are indicated with
-
юсь(⏎Y. Раскройте скобки, употребляя глаголы в Present, Past, Future Indefinite, Present, Past, Future Continuous. 1.When I (to open) the window, it
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: the things I keep going with, but maybe I could do a little without. What about… if you speak. Maybe not, but let with the length I keep. You can not be. But I am. Anyway. Because I could
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
isease induced changes in gene flow patterns among Tasmanian devil populations⏎Bruniche-Olsen, A and Burridge, CP and Austin, JJ and Jones, M, Disease induced changes in gene flow patterns among Tasmanian devil populations,
-
; 1945; OFR; 45-37; Fox, I. William; Thigpen, Claude H.; Ginter, Roy L.; Alden, George P.<|endoftext|>Posted September 1st,
-
leg ulcer (4) (6) (7).⏎(1) Posnett, J., Gottrup, F., Lundgren, H., Saal, G. The resource impact of wounds on health-care providers in Europe. J W
-
if you notice any errors which make this publication unusable.⏎XML: View XML⏎Affleck, David L. R.; Gregoire, Timothy G.; Valentine, Harry T
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
- L9:695 excites
- L7:94 suppresses
- L11:1053 excites
- L10:453 suppresses
- L11:47 suppresses
- L9:770 suppresses
- L7:222 weakly suppresses
- L7:48 weakly excites
- L11:1312 weakly suppresses
- L8:1198 weakly excites
- L7:75 weakly excites
- L9:422 weakly suppresses
- L6:354 weakly suppresses
- L8:986 weakly excites
- L3:674 weakly excites
- L10:601 weakly excites
- L3:426 weakly excites
- L6:1115 weakly suppresses
- L10:2952 weakly suppresses
- L8:388 weakly suppresses
- L7:768 weakly suppresses
- L11:1460 weakly suppresses
- L9:259 weakly excites
- L7:63 weakly excites
Feeds later layers
- L13:592 excites
- L14:1258 suppresses
- L13:2501 excites
- L14:683 excites
- L13:2789 excites
- L25:152 suppresses
- L13:1038 weakly excites
- L15:1459 weakly excites
- L15:2845 weakly excites
- L13:2022 weakly excites
- L14:858 weakly suppresses
- L15:751 weakly suppresses
- L19:120 weakly suppresses
- L16:1127 weakly excites opening bracket after an imperial measurement
- L21:958 weakly excites
- L14:776 weakly excites
- L14:1382 weakly excites
- L13:598 weakly suppresses
- L20:2477 weakly suppresses
- L18:571 weakly suppresses
- L14:2266 weakly excites
- L19:228 weakly suppresses
- L15:816 weakly excites
- L14:720 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
- ocene
- bsub
- ươ
- inct
- -has
- fixation
- pisa
- .hasNext
- oplast
- ntax
Suppresses
- mine
- rait
- gesch
- adders
- VERR
- 出差
- 隔壁
- 姐姐
- scri
- arter