year in a written date
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 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 year in a written date, such as the 1928 in 'December 7, 1928'. When it was turned up, the model's text drifted off topic and became repetitive, so the steering did not show a clear effect.
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.
-
|⏎|Born||December 7, 1928⏎|Era||20th / 21
-
not aware, the draft is a system of conscription that enlists citizens into the military; the system is run using an online sign up form that assigns you a draft number, and should the number be drawn you will be called on for military
-
to solve a 2,000 year-old mystery: Who is the Cartographer?More⏎We've sent an email with your order details. Order ID #:⏎
-
full Titus Vespasianus Augustus, original name Titus Flavius Vespasianus, (born Dec. 30, 39 ce—died Sept.
-
⏎robert.kern at gmail.com⏎Fri Sep 10 00:10:22 CEST 2010⏎On 9/9/10 4:39 PM, Baba wrote:⏎
-
The American historical painter Washington Allston was born in on the Allston Plantation in the Georgetown district of South Carolina Nov. 5, 1779. His parents sent him to Harvard, where not only did
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: 12, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 19, 18, 20
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
: ||4.4 MB|⏎|Publisher: ||CRC Press|⏎|Date published: || 2015|⏎|ISBN: ||9781498738552 (DRM-PDF)|⏎
-
structure = function.⏎(Image from: https://commons.wikimedia.org/w/index.php?curid=1267900)⏎Want to suggest an edit? Have
-
’s website lists 16474 pharmaceutical reports that can be purchased at http://www.fastmr.com/catalog/browse.aspx?category=16&page=1. The online catalog offers a description for each report, which
-
12.0 MB|⏎|Publisher: ||Apple Academic Press|⏎|Date published: || 2016|⏎|ISBN: ||9781771882903 (DRM-PDF)|⏎
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
- L10:2425 excites
- L9:926 weakly suppresses
- L1:1058 weakly suppresses closing quote marks after a quoted term
- L11:52 weakly excites
- L7:1037 weakly excites
- L4:2451 weakly suppresses
- L9:2453 weakly suppresses
- L10:2847 weakly suppresses
- L10:987 weakly suppresses
- L0:247 weakly excites
- L2:2554 weakly excites
- L8:2854 weakly suppresses
- L10:2832 weakly suppresses
- L8:585 weakly suppresses
- L6:2018 weakly suppresses
- L11:2054 weakly suppresses
- L2:2040 weakly excites
- L10:354 weakly excites
- L10:2394 weakly excites
- L7:1381 weakly excites
- L9:1246 weakly excites
- L8:1049 weakly excites
- L11:703 weakly excites
- L10:499 weakly excites
Feeds later layers
- L22:1076 suppresses
- L19:1905 suppresses
- L15:2351 suppresses
- L16:1676 suppresses
- L18:1712 suppresses
- L24:654 suppresses
- L19:1378 suppresses
- L18:2718 suppresses
- L13:2602 suppresses
- L16:1720 excites
- L17:2672 suppresses
- L17:932 weakly suppresses
- L20:591 weakly suppresses
- L16:2196 weakly excites
- L19:1033 weakly suppresses
- L18:2978 weakly excites
- L24:1467 weakly suppresses
- L17:3011 weakly excites
- L17:2082 weakly suppresses
- L16:901 weakly suppresses
- L16:2637 weakly suppresses
- L15:211 weakly suppresses
- L16:2486 weakly excites
- L14:1783 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
- طار
- orts
- 步行
- monds
- antes
- Ladies
- ệu
- coholic
- +(\
- ritten
Suppresses
- IDs
- ID
- 번호
- 号码
- ids
- .getID
- 号
- ID
- achi
- 编号