unclear: mid-sentence word choices, mixed
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 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 fires on a grab bag of words in the middle of sentences, such as a number in a list, the word 'five', or 'true' in a comparison. I could not find one clear theme. Turning it up made the model ramble about unrelated topics.
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.
-
do not contain zeros: 1, 2, 3, 4, 5, 6, 7, 8, 9, 13, 14, 15, 16,
-
plants, and a virus blows in across the ocean from a remote island, your lawn has a better chance of surviving and renewing itself than if you have only five kinds of plants. In the same way, having more species of wild birds will
-
no", "yes"]⏎Note that the arguments to⏎bool are in the opposite order of the⏎if then else-syntax; First the false value, then the true value, and finally the boolean. This argument order is consistent with the data type
-
than two numbers are being multiplied, just count the number of minus signs: any even number of minus signs collapses to a plus sign; any odd number of minus signs collapses to a minus sign.⏎Input your answer as:
-
ātauranga strategy for Ngāti Kahungunu are as follows:⏎Strengthen, support and guide Kahungunu ākonga, whā
-
classified by sex in the form of a histogram, it is called age-sex pyramid. The base of the pyramid shows very low age starting from zero, it increases as we move upwards. The top of the pyramid shows the maximum age above 8
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 set of all sets that are a set of some set of some set of some set of some set of some set of ... the set of some set of some set of ... the set of some set of some set of some set of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a branch of mathematics studying zeros of polynomial equations. Modern algebraic geometry is based on more abstract techniques of abstract algebra, especially commutative algebra, with the language and the problems of geometry.⏎The main objects of study
-
work for arbitrary toposes can be deduced by arguing from universal elements. Some simple examples follow.⏎Consider first the construction of internal conjunction .⏎is equivalent to the nerve of .⏎As a warm-up:⏎The colimit of⏎is isomorphic to
-
“Complex Geometry“:⏎In mathematics, complex geometry is the study of complex manifolds and functions of many complex variables. Application of transcendental methods to algebraic
-
classically Diophantine geometry), a subfield of algebraic number theory.⏎- The study of the real points of an algebraic variety is
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
- L9:65 suppresses
- L3:25 suppresses
- L10:276 excites
- L3:461 excites
- L7:334 excites
- L9:41 excites
- L6:1525 excites
- L5:221 suppresses
- L3:924 suppresses
- L9:693 suppresses
- L10:2217 suppresses
- L5:78 excites
- L9:96 excites
- L9:1781 excites
- L0:281 excites
- L10:544 suppresses
- L1:137 suppresses
- L3:105 excites
- L3:448 suppresses
- L3:63 excites
- L6:1459 excites
- L8:77 excites
- L9:487 excites
- L3:408 weakly excites
Feeds later layers
- L21:246 suppresses
- L20:153 suppresses
- L24:263 suppresses
- L26:330 suppresses
- L26:173 suppresses end-of-document boundary token
- L21:987 suppresses
- L24:657 suppresses
- L21:345 suppresses
- L25:299 suppresses
- L24:246 suppresses
- L21:1762 suppresses
- L21:684 suppresses
- L24:287 suppresses
- L22:1196 suppresses
- L24:310 suppresses
- L23:1330 suppresses
- L19:1027 suppresses
- L20:461 suppresses
- L20:157 suppresses
- L25:2827 suppresses
- L21:947 suppresses
- L21:630 suppresses
- L20:870 suppresses
- L22:1080 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
- 始建
- 通行证
- 首家
- USER
- lemetry
- wayne
- webkit
- civilian
- MainThread
- 新一轮
Suppresses
- leurs
- ators
- -unit
- ören
- uations
- jections
- otomy
- erator
- PermissionsResult
- capit