unclear: mid-phrase words across many languages
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. 8 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 a random assortment of words in the middle of a phrase, in many languages, with no clear shared meaning. Turning it up or down changed the text the model wrote in unrelated ways, often switching languages.
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.
-
turning off lights).⏎We appreciate the efforts of the contributors to this special issue who are addressing some of<|endoftext|>L’Encyclopédie de l’histoire du Québec / The Quebec History Encyclopedia⏎Ethics and Morals of the Indians of
-
88: περιγράφει τετράγωνον σχῆμα τῷ λόφῳ in Einklang zu bringen, ist es unumgänglich no
-
en and Plan du Havre de Noel, dans la Terre de Kerguelen, as well as an inset at bottom left showing two
-
de Doopsgezinde Gemeente to Amsterdam, 2 vols. Amsterdam: Uitgegeven en ten geschenke aangeboden door den Kerkeraad dier Gemeente, 1883-
-
Is.⏎|Prepared by:||Gordon Rudloff⏎Ohio Department of Agriculture<|endoftext|>Logic & ProofsOpen Learning Initiative⏎Cosa impari in questo corso?⏎Logic & Pro
-
these prepositions must be in accusative case: durch-throughfuer-for gegen-againstohne-without um-around Er kauft das Geschenk fuer mich.<|endoftext|>In case you are searching for easy Preschool Activities, you can
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: vous pouvez-tu dire que vous ne comprenez pas la philosophie des systèmes de gestion de ressources de données, c'est un concept de thèse, mais c'est un système de notation? Pour les réflexions
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
O snap/ring gauges (for checking shafts) and plug gauges (for
-
der Zijpp|⏎<|endoftext|>The Great Pacific Garbage Patch may qualify as the largest garbage
-
04)⏎(4) « La compression veineuse est le traitement de base des affections ve<|endoftext|>With concern over
-
Antalya-Türkiye’de doğdu. Lisans öğrenimini MSÜ Deniz Harp Okulunda tamamlamıştır.⏎
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
- L3:6 strongly suppresses
- L5:136 strongly suppresses
- L3:1090 strongly suppresses
- L5:150 suppresses
- L2:166 suppresses
- L2:2022 suppresses
- L3:113 excites
- L2:24 excites
- L3:530 excites
- L7:130 excites
- L1:1227 excites
- L6:479 suppresses
- L1:651 suppresses
- L4:1147 suppresses
- L2:1733 suppresses
- L6:331 weakly excites
- L0:1090 weakly excites
- L4:1272 weakly excites
- L3:318 weakly excites
- L4:1789 weakly excites
- L2:1934 weakly excites
- L2:1700 weakly suppresses
- L7:92 weakly excites
- L0:83 weakly excites
Feeds later layers
- L15:65 strongly suppresses
- L16:477 strongly suppresses
- L17:487 strongly suppresses
- L17:649 strongly suppresses
- L10:39 strongly suppresses
- L14:381 strongly suppresses
- L17:1021 strongly suppresses
- L11:137 strongly suppresses
- L13:1204 strongly suppresses
- L9:453 strongly excites
- L11:416 strongly suppresses
- L13:929 strongly suppresses
- L20:377 strongly suppresses
- L21:266 strongly suppresses
- L16:1347 suppresses
- L18:560 excites
- L12:1423 suppresses
- L14:1424 suppresses
- L20:267 suppresses
- L16:979 suppresses
- L18:566 suppresses
- L18:72 excites
- L17:2506 suppresses
- L18:567 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
- hers
- 的独特
- 之势
- rove
- 了一口
- ropes
- roach
- ToBounds
- ировки
- 扒
Suppresses
- ác
- dé
- STRU
- enny
- .opend
- annel
- ح
- oire
- .ReadLine
- "));⏎