the digit 3 inside 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. 18 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 digit 3 when it appears inside a number, such as the 3 in a Psalm reference, a percentage, a word count, or a year like 2015. Turning it up made the model's continuations drift into unrelated topics like 3D objects and weather, so the effect on the text is not easy to read.
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.
-
6:1-13; Psalms 1:1-6; Psalms 2:1-12; Psalms 3:1-8; Psalms 4:1-8; Ps
-
of 1009 specimens were reported as positive for influenza viruses, 685 (67.9%) were typed as influenza A, 324
-
be assessed by means of a combination of one 1000 word assessed essay [15%], one 2000 word assessed essay [35%] and one two-hour unseen written examination paper in which the student will
-
83 cases with known vaccination status, 87% were unvaccinated, 8% were vaccinated with one dose of measles-containing vaccine, 3% were vaccinated with two or more doses, and 2% were vaccinated with
-
/her own and does not want to move to Curry Island with his/her...⏎This section contains 570 words|⏎(approx. 2 pages at 300 words per page)<|endoftext|>Of all the technology featured at CES this week
-
status silver objects.⏎Elements of the hoard will be on display for the first time at King’s Museum, University of Aberdeen from 20 January to 31 May 2015
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: what is the first non-1010 to 101010 to the 1010...?...? How do I proceed? How do I write the terms?? A... for the first,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
6 lay persons. Participants were presented with a questionnaire using either the term intellectual disability or mental retardation and asked to report on stereotypes, prejudice, and discrimination in response to a vignette. Moderate levels of stigma were found. An indirect effect was
-
. My specific interests are related to species’ responses to environmental change, and alpine plants provide an incredible system to understand changes in range margins and community composition in response to global change. In the European Alps, plant species are shifting their range to
-
anatomy of a spacesuit⏎A spacesuit is made up of practically a half dozen completely different elements and may have as much as 16 layers, in response to NASA.⏎NASA revealed a floor prototype of the brand new exploration extraveh
-
logical links in the /dev directory. The devfsadm command also maintains the path_to_inst instance database.⏎Updates to the /dev and /devices directories in response to dynamic reconfiguration events or file system accesses are handled by devfsadmd
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
- L4:347 strongly excites
- L16:134 strongly excites
- L17:994 strongly excites
- L0:1108 strongly suppresses
- L16:671 strongly excites
- L2:1774 strongly excites
- L2:1866 suppresses
- L16:112 excites
- L16:260 excites
- L10:887 suppresses the B in vitamin B names
- L16:374 excites
- L17:134 excites
- L10:715 excites
- L7:2095 excites
- L17:253 suppresses
- L10:1473 suppresses
- L17:594 excites
- L16:215 suppresses
- L15:129 suppresses
- L16:56 excites
- L15:82 suppresses
- L16:827 suppresses
- L15:96 suppresses
- L5:1984 excites
Feeds later layers
- L20:1308 strongly excites the digit 3 inside years
- L19:196 strongly suppresses
- L19:647 strongly excites
- L21:1189 excites
- L22:1550 suppresses
- L21:252 suppresses
- L21:815 suppresses
- L21:325 suppresses
- L20:411 suppresses
- L20:133 suppresses
- L20:247 suppresses
- L19:1021 suppresses
- L20:183 suppresses
- L19:322 suppresses
- L19:1615 suppresses
- L20:1310 suppresses
- L21:2715 suppresses
- L20:211 suppresses
- L25:2807 excites
- L22:715 excites
- L19:593 suppresses
- L19:879 suppresses
- L19:2848 suppresses
- L19:2826 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
- rd
- -quarters
- rd
- arters
- →rd
- مائة
- cheers
- RD
- lectric
- rbrace
Suppresses
- Warp
- GetUser
- HttpResponse
- GetString
- ọi
- rag
- 률
- SError
- CAS
- ivant