unclear: mixed mid-sentence word choices
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 mixed: they peak on different tokens. 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 reacts to a grab bag of words in ordinary English sentences, such as 'reasonable', 'clear' and 'skills', with no single theme we could find. It tends to go quiet on code and list-like text. Turning it up did not produce a clear change in what the model wrote.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
182) was introduced by Congress in 2011 with the support of President Obama, who in a statement called the legislation “ a common-sense bill that will help ensure that men and women who do equal work receive the equal
-
the spellings are inside the whiteboard :)⏎Quality not yet verified by the community.⏎This resource does not contain any images, words or ideas that would upset a reasonable person in any culture.⏎This resource is licensed by LoubeeSav⏎under
-
who were tested on the basis of clinical suspicion or risky exposure. A human case of brucellosis was considered confirmed if results of serologic tests, such as ELISA or complement fixation test, were positive, in accordance with the World
-
there?⏎There are several types of anxiety disorders. They are:⏎Phobia is about irrational fear of a situation, animal or object, which cannot be explained rationally to the sufferer.⏎A person with a phobia will be triggered by very
-
their specifics differ significantly, the dealers can be any wholesale and retail companies that act as intermediaries between the producer and the consumer. If everything is more or less clear with the commodity sphere, then in the service sector, for example, dealers may
-
ways in the Balkans. But these unique ecosystems in southeast Europe could soon be gone, along with endangered species such as the Balkan lynx, if plans for over 2,000 dams go ahead, conservationists warn.⏎Western
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 am not ready to learn." In this case, the student is ready to learn. So, "Solving the Problem" is the best way to learn. So, "I am learning in a new environment." So, "
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
!⏎xs = [6, 9, 13, 42]⏎-- Without library⏎f :: Bool -> [Int]⏎f b = if b then xs else ⏎-- With library⏎g :: Bool -> [Int]⏎g
-
while loop:⏎given_arr = [1, 2, 3, 4, 5] i = len(given_arr) - 1 while(i >= 0): print(given_arr[i]) i = i - 1
-
> time." (from http://en.wikipedia.org/wiki/Analysis_of_algorithms)⏎> 1 get a positive integer from input⏎> 2 if n> 10⏎> 3 print "This might take a while..."⏎>
-
treated as high, and if it is less than 2.5 volts, then it will be treated as low. So,
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
- L4:526 excites
- L5:1931 suppresses
- L5:328 suppresses
- L4:361 excites
- L6:377 weakly excites
- L2:537 weakly excites
- L5:1044 weakly excites
- L6:2134 weakly excites
- L5:2309 weakly suppresses
- L5:1792 weakly excites
- L3:2323 weakly suppresses
- L2:1493 weakly suppresses
- L6:1037 weakly suppresses
- L2:461 weakly suppresses
- L6:195 weakly suppresses
- L5:1772 weakly suppresses
- L6:792 weakly excites
- L0:128 weakly excites
- L5:908 weakly suppresses
- L1:2868 weakly excites
- L5:237 weakly excites
- L4:1221 weakly suppresses
- L5:399 weakly suppresses
- L6:678 weakly excites
Feeds later layers
- L12:740 strongly suppresses
- L16:2167 suppresses
- L10:750 suppresses
- L10:659 suppresses
- L10:2639 suppresses
- L11:2487 suppresses
- L11:1725 suppresses
- L8:367 excites
- L13:2671 suppresses
- L13:1286 suppresses
- L8:1381 suppresses
- L12:249 excites
- L11:999 suppresses
- L13:2160 suppresses
- L8:1920 suppresses
- L8:248 suppresses
- L13:905 suppresses
- L8:691 suppresses
- L8:1044 suppresses
- L10:545 excites
- L13:1508 suppresses
- L13:960 excites
- L8:158 excites
- L10:203 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
- 函
- ",-
- ={[
- acs
- atura
- 为代表
- guard
- 媒体
- =[⏎
- guard
Suppresses
- oration
- antry
- icum
- 不认识
- IRTH
- chưa
- syn
- itone
- ṅ
- theros