period or digit inside section 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. 12 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 fires on the dot, digit or dash inside numbered references like section 8.C.8.1, figure 5-18 or version numbers. Turning it up made the model write lists and numbered headings, which fits the theme.
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.
-
IFICATION; LEGAL CLUSTERS; GEOGRAPHIC CLUSTERS; TRADE CLUSTERS; ENVIRONMENT CLUSTERS; OTHER FACTORS. I. Identification. 1. The Issue The …⏎BONN, Germany (
-
4.3 feet of gulf water by the end of the century.⏎That rate could swamp
-
re-exported goods might require some additional information.⏎Having understood so much about reexport trade and what are re-exportation, it is<|endoftext|>10.2.2 Restarting TCP/IP⏎TCP/IP is a peer-to-peer, connection-oriented
-
- 3 Pulse transformer⏎- 4 RF transformer⏎- 5 Audio transformer⏎- 6 Other types⏎- 7 See also⏎- 8 References⏎This is the most common type of transformer, widely used in electric power
-
V. That’s 10% of the intended voltage!⏎So what are the traditional
-
that all highly toxic or offensive vapors are scrubbed or adsorbed before the exit gases are released into the hood exhaust system (see section 8.C.8.1 on fume hood scrubbers).⏎It is necessary to check if
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 can make a space that is a square, or a cube or a tetrahedron or a sphere, but I don't need to have a lot of things to put in or take out from. (This is a statement that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Intravitreal injections are the fastest-growing procedure in ophthalmology and in medicine in general (Video 20-10, Fig 20-17). The most common indications for injections include AMD, diabetic retinopathy
-
. Self-equalizing Anchor. A self-equalizing anchor will maintain an equal load on each individual point as the direction of pull changes (Figure 5-18). This is sometimes used in rappelling when the route must change left
-
not be considered medical advice...This answer should not be considered medical advice and should not take the place of a doctor’s visit. Please see the bottom of the page for more information or visit our Terms and Conditions.⏎Thanks for your feedback.⏎3
-
product-term line . A structure such as Figure 5.11 (c) is called programmable array logic , first introduced by Monolithic Memories
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
- L11:2116 excites dots inside version and section numbers
- L5:1316 excites
- L9:799 excites
- L9:1979 weakly excites
- L11:965 weakly excites
- L11:1532 weakly suppresses
- L8:2035 weakly suppresses
- L9:1587 weakly excites
- L4:1285 weakly excites
- L10:1516 weakly excites
- L10:2284 weakly suppresses
- L9:875 weakly excites
- L10:1479 weakly excites
- L11:2505 weakly excites
- L3:972 weakly suppresses
- L6:595 weakly suppresses
- L3:2466 weakly excites
- L11:869 weakly excites 'body' in immune and pain explanations
- L11:1685 weakly excites
- L7:1707 weakly suppresses
- L5:1363 weakly excites
- L4:1721 weakly excites the gap before a decade or century
- L7:868 weakly excites
- L5:26 weakly excites
Feeds later layers
- L13:1797 excites
- L14:1530 excites
- L17:1135 excites
- L20:925 excites
- L17:2548 excites
- L17:756 weakly excites
- L18:1640 weakly excites
- L17:2775 weakly suppresses
- L21:2710 weakly suppresses
- L13:2123 weakly suppresses
- L14:2446 weakly excites
- L16:1652 weakly excites
- L17:2696 weakly suppresses
- L13:852 weakly suppresses
- L13:2498 weakly excites
- L20:2393 weakly excites
- L21:2299 weakly excites
- L14:2354 weakly excites
- L27:393 weakly excites
- L23:1295 weakly suppresses
- L21:2372 weakly excites
- L20:1798 weakly excites
- L16:2306 weakly suppresses
- L23:1769 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
- below
- separate
- below
- -step
- step
- Expl
- 步骤
- Overview
- nutshell
- overview
Suppresses
- anka
- ány
- 索取
- unker
- taskId
- imeType
- Neville
- TES
- ipur
- /books