'as' in 'not as ... as' comparisons
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. 7 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 'as' in comparisons such as 'not as severe as' or 'not as rich as'. Turning it up did not produce writing about comparisons, so the link to the output is unclear.
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.
-
is more easily absorbed and digested by your body, so it can be taken with or without food. It doesn’t contain as much elemental calcium that your body absorbs as calcium carbonate, though, and does tend
-
. They will experience high energy and some impulsive behavior, but it is typically not as severe as mania. This low level “high” is referred to as hypomania
-
Galvanized steel plate. The galvanized steel sheet manufactured by electroplating has good workability. However, the coating is thinner and its corrosion resistance is not as good
-
ender and rosemary are more easily grown from transplants. Plant in full sun. Soil should be somewhat improved with organic matter, but most herbs do not require as rich a soil as vegetables do. However
-
.…It's less powerful than even a small…watch battery.…For processing and mixing,…the weak mic level
-
leaders talked almost half as much, only 19 percent.⏎Later studies revealed that these negative<|endoftext|>Swine flu, these days, may not be as rampant as last year when over 787 cases were detected in the State and nearly
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 a 200g coin, and the coin is either genuine or one of the coins, which is either genuine or the one of the coins, and I need to find the coin of the 200g in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
second bottle to be the fake.⏎In practice, there are some limitations to a manual application of this geometric analysis
-
with a few other lighter and heavier outliers.⏎The main use for the 45-70 Government is as
-
game I played as a child and I’ve always enjoyed its simplicity. If you’ve never played it before, it’s a two-player game of asking yes/no questions to identify a randomly selected character from a panel of twenty-four possible candidates before your
-
slipped disc,” “pinched nerve,” “bulging disc,” or “ruptured disc
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
- L6:1051 suppresses
- L6:2339 excites "as" completing an "as ... as" comparison
- L6:903 suppresses
- L6:2300 suppresses
- L3:1681 weakly suppresses "which" after a noun of "the X of which"
- L6:16 weakly excites
- L1:277 weakly excites
- L6:1983 weakly suppresses
- L5:2143 weakly excites
- L4:2673 weakly suppresses
- L6:1820 weakly excites
- L6:17 weakly suppresses
- L6:1634 weakly suppresses
- L4:2925 weakly suppresses
- L5:2632 weakly excites
- L0:1301 weakly suppresses
- L6:1713 weakly suppresses
- L3:836 weakly excites
- L1:2491 weakly excites
- L6:7 weakly suppresses
- L5:1598 weakly suppresses
- L6:1833 weakly excites
- L3:968 weakly excites
- L3:2014 weakly suppresses
Feeds later layers
- L9:2031 suppresses
- L9:2108 suppresses Rh blood type and related terms
- L11:1031 suppresses
- L10:3064 suppresses
- L17:835 suppresses
- L8:2025 suppresses
- L12:1795 weakly excites
- L12:1153 weakly excites
- L9:255 weakly excites
- L8:2258 weakly suppresses
- L15:2381 weakly excites
- L15:1558 weakly excites
- L22:634 weakly excites
- L23:628 weakly excites
- L14:1771 weakly excites unclear: mixed, mostly technical prose
- L11:2603 weakly suppresses
- L15:1081 weakly suppresses
- L16:2235 weakly excites
- L8:2090 weakly suppresses
- L16:1371 weakly excites
- L8:1012 weakly excites
- L12:1401 weakly suppresses
- L19:2603 weakly suppresses
- L11:2815 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
- _PUT
- _put
- 창
- uong
- Toastr
- düzenlen
- tır
- .figure
- ậy
- 徨
Suppresses
- 一枚
- 党
- -dot
- сен
- inar
- �
- 取证
- �试
- oppable
- INCT