the digit 6 inside years
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. 20 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 fires on the digit 6 when it sits inside a year or a number, such as the 6 in 1960 or 1860. It is one of the strongest neurons in the set, but it is mostly about one character in a number. Turning it up or down gave text that wandered off into doctors or lists of numbers, so the steering does not clearly show the pattern.
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.
-
as an addon to the ip address. The number of port in computer networking is expressed in 16 bits, which means that there is a maximum of 65,536 ports on a computer. Assign the call coverage path for
-
Maxwell.⏎Cgs systems were gradually superseded by meter-kilogram-second-ampere (MKSA) systems, and finally rendered obsolete in 1960 by the CGPM's adoption of SI, which is a MKSA system
-
⏎Two international road traffic Conventions, namely the Geneva Convention on Road Traffic of 1949 and the Vienna Convention on Road Traffic of 1968, are the foundation of many national traffic laws. Countries signing either or both
-
is.⏎In the east, the Roman Empire later relied heavily on the now only semi-autonomous poleis, which in late antiquity (284–641 AD) experienced a slow decline in
-
i'ite and Sunni—what's the difference?⏎The short answer is that the split occurred in naming a successor to the Prophet Muhammad on his death in 6
-
the Pad There are two types of research<|endoftext|>The Apollo 11 space flight landed the first humans on Earth’s Moon on July 20 1969. The mission, carried out by the United States, is considered a major
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 which the which is the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which the which
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cho IV of Castile⏎|Father||James I of Aragon|⏎|Mother||Violant of Hungary|⏎|Born||8 June 1236⏎|Died||1301 (aged 64–6
-
However, don’t make the presentation too word<|endoftext|>Sore Throat Complications Not Increased with Fewer Antibiotics⏎Am Fam Physician. 2006 May 1;73(9):1619-16
-
rise in obesity.⏎However, other researchers conducted a longitudinal cohort study between May 2020 and April 2021 among 2026 Canadian children between the ages of two to 18 years old, measuring primary
-
20年7月6日 at 9:45 AMに Megan Smith さんが最後の更新を行いました。⏎2020年7月6日 9:45 AM #2107
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
- L18:924 strongly suppresses
- L0:2775 suppresses
- L3:959 suppresses
- L18:123 excites
- L16:26 suppresses
- L5:1722 suppresses
- L16:673 suppresses
- L2:1724 excites
- L18:1023 suppresses the digit 3 inside numbers
- L16:94 suppresses
- L10:951 excites
- L11:981 suppresses
- L2:2029 excites
- L19:1615 excites
- L10:715 suppresses
- L2:1774 suppresses
- L16:84 excites
- L15:82 weakly suppresses
- L19:996 weakly excites the digit 5 inside numbers
- L15:80 weakly suppresses
- L15:114 weakly excites
- L11:2284 weakly excites
- L14:643 weakly suppresses
- L0:996 weakly suppresses
Feeds later layers
- L26:1602 excites
- L22:3017 excites
- L26:2573 suppresses
- L21:921 suppresses
- L23:1670 weakly suppresses 'the' before river names
- L22:1882 weakly suppresses
- L26:1408 weakly excites the 'n' that starts nectar
- L23:1766 weakly excites
- L22:2547 weakly excites
- L26:1943 weakly excites
- L23:2163 weakly excites
- L22:462 weakly excites
- L26:1877 weakly excites
- L23:3014 weakly excites
- L23:1792 weakly excites
- L27:1415 weakly suppresses
- L21:334 weakly suppresses
- L26:1355 weakly excites
- L21:1233 weakly suppresses
- L26:518 weakly excites
- L23:1140 weakly suppresses opening bracket before an acronym
- L26:636 weakly suppresses
- L26:1432 weakly suppresses
- L26:394 weakly 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
- 능
- ��
- ENCE
- axis
- üzerine
- iper
- לידי
- 思った
- accom
- 鬣
Suppresses
- rés
- düş
- ried
- los
- вещ
- owell
- reassuring
- 増え
- pherd
- rog