'one' before 'of' in 'is one of the'
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. 16 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 lights up on the word 'one' when it comes right before 'of the most...' or a similar phrase, as in 'it is one of the best'. When it was turned up, the model's writing drifted into odd repetition and loose, rambling sentences rather than anything about this idea.
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.
-
reef fish abundance and distribution trends between locations. Armed with the knowledge of where<|endoftext|>Edited by William Byron Forbush This is a book that will never die — one of the great English classics. . .
-
200m photos are uploaded to Facebook each day. These three “v”s neatly encapsulate the reasons why we are experiencing a data explosion, one that is resulting
-
great with yogurt or cereal.⏎2. Eggplant. These beautiful purple veggies are one of the best dietary sources of antioxidants and also provide lots of soluble fiber. Try
-
layout of our collection is much more light, airy and spacious. Following on from Education, I’ve been taking a fresh look at the History section.⏎It’s one of the most extensive in terms of the number of publications. All kinds of stories
-
or a weight belt. These muscles, one on each side, are your true core muscles and strengthening them will give you power
-
how Buddha nature produces the conditions for Buddha nature.”⏎There is much said and written about “Buddha nature”. Maybe to much. In my experience it is one of the most misunderstood terms that has arisen from Buddhism moving to the West. I
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 have been, for example, the first, for the first, but not only for the first, but not only for the first, for the first, for the first. I have no one, no, no, no, no
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Report on Macdonald (Undersea Features) (Wunderman, R., ed.). Bulletin of the Global Volcanism Network, 32:1. Smithsonian Institution. https://doi.org/10.5479
-
urchasing Power Parity Exchange Rates for the Global Poor,"⏎American Economic Journal: Applied Economics,⏎American Economic Association, vol. 3(2), pages 137-66, April.⏎- Angus Deaton & Olivier Duprie
-
ak Dun v. Yukon, 2015 YKCA 18 (the “appeal decision”), the Yukon Court of Appeal unanimously allowed
-
; Magnuson, Terry (2013). "A novel selection marker for efficient DNA cloning and recombineering in E. coli". PLoS ONE 8 (2): e57075. . . http://
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
- L3:244 strongly suppresses
- L4:961 strongly excites
- L10:848 suppresses
- L3:1810 excites
- L0:871 excites
- L11:1852 excites the digit 1 in numbers
- L4:247 suppresses
- L10:1639 suppresses
- L2:1829 suppresses
- L5:559 suppresses
- L10:1284 excites
- L10:2268 excites
- L9:1373 weakly suppresses
- L8:972 weakly excites
- L0:1500 weakly suppresses
- L11:2056 weakly excites the digit zero inside numbers and times
- L0:2393 weakly excites
- L0:1332 weakly suppresses
- L10:800 weakly excites
- L2:2029 weakly suppresses
- L0:2471 weakly suppresses
- L10:1602 weakly excites the number 1 as a unit value
- L10:722 weakly suppresses
- L6:1502 weakly suppresses
Feeds later layers
- L21:518 excites
- L19:720 excites
- L18:2608 excites
- L22:260 excites
- L23:2466 excites
- L23:794 excites
- L19:1251 excites
- L19:2118 excites
- L22:1081 excites
- L17:1056 suppresses
- L19:2221 suppresses
- L21:442 suppresses
- L25:1220 excites
- L19:288 suppresses
- L26:2856 excites
- L19:1331 excites
- L17:3022 excites
- L24:489 weakly excites
- L18:1686 weakly excites
- L17:963 weakly suppresses
- L25:2583 weakly excites
- L17:2705 weakly excites 'of' in 'one of the' phrases
- L19:1752 weakly excites
- L23:1563 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
- ploy
- ển
- Thousand
- 趸
- 被打
- onn
- culator
- //**⏎
- liter
- -shot
Suppresses
- strings
- scr
- 率
- sar
- rats
- walls
- 率为
- labels
- ulous
- istas