unclear: word fragments in foreign text
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. 25 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 pieces of longer words in non-English text, such as the middle of a Swedish or Persian word, and it is hard to say what they have in common. Turning it up made the model repeat itself or drift into Chinese characters, so it does not show a clear job.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
to increase, which is a growing public health concern.⏎|Status||Published - 2014 jan. 1|⏎- Folkhälsovetenskap, global hälsa, socialmedicin och epidemiologi<|endoftext|>Russia can
-
: Madison McDermott⏎The Green New Deal Federal Resolution was introduced in 2019 by Rep. Ocasio-Cortez and Sen. Ed Markey as a plan to tackle a trifecta of crises: the climate
-
.⏎2 پژوهشگر دوره تخصصی دکتری معماری منظر، دانشکده معماری و شهرسازی، دانش
-
a representation of both the classes whereas TEU consists of officers only. Though<|endoftext|>نوع مقاله : مقاله پژوهشی⏎1 دانشیار معماری، دانشکده معماری
-
Pine Ridge, he advocated for peace with the European Americans⏎Date de dernière mise à jour : 05/07/2021⏎Ajouter un commentaire<|endoftext|>By DANIEL SUDDEATH⏎NEW ALBANY —
-
uffassung des Dionys. I 88: περιγράφει τετράγωνον σχῆμα τῷ λόφῳ in Einklang zu bringen
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:这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这这
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and 2014, World Bank says. Wealth remains unequal, and some countries are falling behind.⏎El Banco Mundial analiza la riqueza de los países, tomando en cuenta no solo el PIB, sino también el
-
research on BPA. Glass bottles, however, also leach chemicals, including aluminum, explains a study by Clemens Reimann of the Norges Geologiske Undersøkelse in Norway and colleagues. Aluminum is sometimes added to
-
) право голосовать⏎2. Прочитайте текст Mayor-Council и заполните пропуски текста словами и выражениями, данными в
-
⏎1. Соотнесите значения слов и выражений, данных в левой колонке, с их английскими эквивалентами, данными в правой колонке⏎1) legislative branch a) Член
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
- L21:1248 suppresses
- L20:1835 suppresses
- L23:618 excites
- L21:1376 suppresses
- L19:1424 excites
- L21:1698 excites
- L24:434 suppresses
- L24:2963 excites
- L19:2842 suppresses
- L19:639 excites
- L23:2052 suppresses
- L22:735 suppresses
- L18:2872 excites
- L17:1371 excites
- L23:66 excites
- L17:1462 excites
- L20:3022 excites
- L20:541 excites
- L22:186 excites
- L20:54 excites
- L19:400 excites
- L20:1842 weakly suppresses
- L21:607 weakly excites
- L15:3007 weakly excites
Feeds later layers
- L26:507 weakly excites
- L26:405 weakly suppresses
- L27:1458 weakly excites
- L27:694 weakly excites
- L27:127 weakly suppresses
- L26:1052 weakly excites
- L26:479 weakly suppresses
- L27:1771 weakly excites
- L27:472 weakly suppresses
- L26:763 weakly excites
- L26:968 weakly excites
- L27:399 weakly excites
- L26:1356 weakly suppresses
- L26:1515 weakly excites
- L26:2417 weakly excites
- L27:994 weakly suppresses
- L26:1647 weakly excites
- L27:273 weakly suppresses
- L26:716 weakly suppresses
- L27:1168 weakly excites
- L27:2608 weakly suppresses
- L27:2317 weakly excites
- L27:1280 weakly excites
- L26:1913 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
- 那些
- 这些
- degli
- 彼ら
- les
- 三个
- 两类
- 各
- Les
- 們
Suppresses
- 这
- 這
- 而这
- 这也
- 但这
- �
- isplay
- _deinit
- hers
- .hl