ordinal numbers like 16th or 1st
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. 9 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 ending of ordinal numbers such as '16th century' or '38th president', the part that comes after the digits. Turning it up did not produce a clear topic in the model's writing; the output mostly drifted into unrelated word problems and loops.
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.
-
most of his life to learning and implementing the tsujigahana, a 16th-century technique that produces the most obscure and jaw-dropping masterpieces. He used
-
Yoshitoshi. It depicts a difficult battle from the Taiheiki (Chronicle of Great Peace), a Japanese historical epic written in the late 14th century and covering the period from 1319 to 136
-
is traditionally focused on repentance. Speaking biblically, to repent means to make a change in our attitudes, words, and lifestyles. As 16th century reformer Martin Luther taught, the Christian life in its totality is a
-
um), held between 1545 and 1563 in Trent (or Trento), now in northern Italy, was the 19th ecumenical council of the Catholic Church. Prompted by the Protestant Re
-
oke from the ancient myths in mind: That people actually have been named Loke or Lokke:⏎Among the Norsemen in Northumberland in the 12th century, there was a man called Locchi.⏎In Scandia, Lok
-
seven roles in his job, from being the Commander-In-Chief to the Guardian of the Economy.⏎Brief biography of Gerald R. Ford, the 38th president of the United States of America.⏎Timeline of the life of former U.S
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 power of a plant that is not in a container, but just as in the container. So, what do you think that is the most important thing? I have to keep this in mind for me, I think, it's a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Rivne region, about 200 kilometres from the site of the Chernobyl explosion in 1986.⏎They found levels of radioactive caesium in milk above Ukraine’s safe limit for adults of 100 Becquer
-
more than 12 times that for children.⏎“More than 30 years after the Chernobyl disaster, people are still routinely exposed to radioactive caesium when consuming locally produced staple foods, including milk, in Chernobyl-affected
-
children are the main consumers of that milk,” said Labunska.⏎“Though the level of soil contamination in the studied areas is not extremely high, radioactive caesium continues to accumulate in milk and other foods, such that the residents of these villages
-
ic, and are considered to be a primary contributor of arsenic in the diet. Shellfish may also contain methylmercury and certain radioactive isotopes of cesium due to leakage from the Fukushima power plant, depending on origin. See: Ar
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
- L4:2364 strongly suppresses
- L3:2400 excites
- L1:2787 suppresses
- L6:411 suppresses
- L8:785 suppresses
- L3:2600 suppresses
- L6:2402 weakly excites
- L3:2943 weakly suppresses
- L4:867 weakly suppresses
- L3:2780 weakly excites
- L5:2936 weakly excites
- L7:1810 weakly excites
- L2:1221 weakly excites
- L6:1119 weakly suppresses
- L2:2107 weakly excites
- L7:2723 weakly excites
- L2:911 weakly suppresses
- L3:2720 weakly excites
- L5:2455 weakly suppresses
- L1:420 weakly excites
- L6:621 weakly suppresses
- L0:1111 weakly excites
- L1:572 weakly excites
- L6:2296 weakly suppresses
Feeds later layers
- L18:2109 excites
- L12:637 excites
- L14:1601 excites
- L17:1594 excites
- L22:2368 excites
- L25:2232 excites
- L10:1744 excites
- L23:465 excites
- L21:939 excites
- L13:1006 excites
- L25:2583 excites
- L14:2281 excites
- L11:1383 weakly suppresses
- L11:233 weakly excites
- L16:2776 weakly excites
- L10:1623 weakly excites
- L17:1628 weakly excites
- L17:1858 weakly excites
- L18:1662 weakly excites
- L11:2781 weakly excites
- L25:2100 weakly excites
- L20:2953 weakly excites
- L25:2234 weakly excites
- L25:1220 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
- eenth
- 年第
- -first
- attribution
- 第一位
- 髻
- ㄡ
- .getFirst
- 第三次
- teenth
Suppresses
- )?$
- 耗
- ?"⏎⏎
- "?⏎⏎
- ibel
- 鲀
- lệ
- elsen
- ?"⏎
- reste