zeros inside large money or count numbers
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. 8 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
The neuron fires on a zero in the middle of a big number written with commas, like the second zero in '$350,000'. It is a narrow pattern about digits, and turning it up did not visibly change the sample 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.
-
of effectiveness) each year produces over $350,000 more income for his/her students compared to an average teacher. On the the
-
our organization in Chicago was called upon to assemble some detailed cost figures from butter manufacturers whose 77 plants produced almost 150,000,000 pounds of butter annually. These figures reflected accurately the big increases in the
-
institute said its efforts had provided research opportunities to 15,000 undergraduates, attracted more women and minorities and assisted nearly 50,000 science teachers and students at local schools across the country engaged in efforts to improve
-
5,000 in today’s dollars. Today, that home—even without major remodels or updates—sells for about $350,000.⏎Compare this with nearby Lakeview, NY, a predominantly African American town
-
found itself subject to a migrating swarm of nearly 20,000 honey bees. The species is close to extinction, so the crew on base
-
910.1200(g)), revised in 2012, requires that the chemical manufacturer, distributor, or importer provide Safety Data Sheets (SDSs) (formerly MSDSs or Material Safety Data Sheets) for
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: [A] This suggests that I have been thinking about lately, and [B] could be something. This might be something. I am not sure about, and [X] is a particular case of [B], and [X]
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Regulations, 1992: Main aim of this legislation is establishing tour operators' responsibility towards customers and providing remedies to them (Gibson, 2016).⏎1.2 Surface, sea and air transport law in
-
: Gwartney, J. D., Stroup, R. L., Sobel, R. S., & Macpherson, D. A. (2018). Economics: Private⏎and public choice (16
-
, word-of-mouth marketing, consumption patterns, information traits, credibility.⏎ANTHEUNIS, M.L., & REIJMERSDAL, E.A. (2012). Social connections and the persuasiveness of viral campaigns in
-
<|endoftext|>FROM ISSUE NUMBER 15 ~ SPRING 2013 GO TO TABLE OF CONTENTS⏎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
- L4:0 suppresses
- L7:323 excites
- L4:1272 suppresses
- L2:882 excites
- L7:1889 excites
- L1:122 suppresses 'World War II' and the 'II' after 'War'
- L3:21 weakly excites
- L3:1057 weakly excites
- L4:179 weakly suppresses
- L2:1169 weakly suppresses
- L3:48 weakly excites
- L6:284 weakly excites
- L5:875 weakly suppresses
- L4:2187 weakly excites
- L4:2561 weakly suppresses
- L6:4 weakly suppresses
- L0:735 weakly excites
- L7:2588 weakly excites
- L4:400 weakly excites
- L7:1376 weakly excites
- L3:188 weakly excites
- L2:2469 weakly suppresses
- L1:580 weakly suppresses
- L4:674 weakly excites
Feeds later layers
- L9:898 weakly excites
- L12:478 weakly excites
- L9:307 weakly excites
- L13:1121 weakly suppresses
- L9:826 weakly suppresses
- L19:247 weakly excites
- L9:1185 weakly suppresses
- L10:699 weakly excites
- L26:2543 weakly suppresses
- L11:222 weakly excites
- L14:820 weakly suppresses
- L12:3061 weakly suppresses
- L10:1397 weakly excites
- L14:2397 weakly suppresses
- L17:1017 weakly suppresses
- L23:93 weakly excites
- L10:208 weakly suppresses
- L9:314 weakly suppresses
- L16:1121 weakly excites
- L12:2965 weakly suppresses
- L12:2442 weakly suppresses
- L12:761 weakly suppresses
- L15:1718 weakly suppresses
- L12:2584 weakly 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
- metry
- ละ
- łów
- arkin
- ourcem
- overe
- appers
- ót
- ictionaries
- trackers
Suppresses
- 枨
- uda
- domest
- 优化
- room
- 发育
- 位
- 膜
- Moder
- membr