"other" in "on the other hand"
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. 22 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 reacts to the word "other" in the phrase "on the other hand", which sets up a contrast. When it was turned up, the model produced long, wandering sentences about unrelated topics, and when turned down it fell into "and and and" loops.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
Convention were drawn up simultaneously, French and English, both being recognized on an equal footing as working languages. The 1929 Conventions, on the other hand, had been drawn up in French only, as French was still the leading
-
IM Lander will integrate data from both the TRN system and LIDAR technology to achieve an ultraprecise lunar landing.⏎Navigation GUI, on the other hand, assists in human-machine interactions, making it possible for ground operators to make
-
because he banged on about the poor: theologians agree that the Gospels show the influence of Greek tragedy in their structure and tone. Theatre historians, on the other hand, will tell you that the breakthrough came in 1837 with
-
word meaning simply “box.” This ark, on the other hand, is represented by a Hebrew word meaning “chest” or “coff
-
object-oriented editing the furthest, as the program allows you to open up a clip and access a huge number of clip-related parameters. Sonar, on the other hand, was not designed with object-oriented editing in mind, and takes a bit
-
of addressing the overall system in a product’s (or building’s) life, ‘green’ describes single, separate aspects. Sustainable, on the other hand, relates to the whole system encompassed by design, manufacture, and purchase
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: and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and and
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
greater dilemma posed by<|endoftext|>How can we help?⏎You can also find more resources in our⏎Select a category⏎Something is confusing⏎Something is broken⏎I have a suggestion⏎What is your email?⏎What is 1 +
-
⏎WE BELIEVE that one of the most important subjects of the knowledge of God, is the MONOTHEISM, i.e. the belief that there is but one God. As a matter of fact Monotheism (TOWH
-
-reliance, and a highly centralized economic system caused Albania to lag far behind its neighbors in terms of economic development.⏎After Hoxha died in 1985, his hand-picked successor, Ramiz Alia, who became
-
, which abruptly brought to an end the global financial system based on the gold standard. The result was that the LMU effectively came to an end in 1914, although it lingered on as a legal entity until its formal dissolution
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
- L12:942 excites
- L19:2676 weakly excites
- L11:1296 weakly suppresses
- L3:2972 weakly suppresses word endings in multi-token nouns
- L21:226 weakly excites
- L12:2878 weakly suppresses
- L14:143 weakly excites 'the' inside journal and institution names
- L16:473 weakly suppresses
- L6:366 weakly suppresses
- L5:2442 weakly excites
- L20:423 weakly excites
- L12:1094 weakly excites
- L9:2357 weakly excites
- L2:1411 weakly suppresses
- L7:469 weakly excites
- L14:2394 weakly suppresses unclear: mid-word and descriptive phrases
- L10:1452 weakly excites
- L1:599 weakly suppresses
- L15:508 weakly suppresses
- L18:2861 weakly excites
- L20:201 weakly suppresses
- L0:1351 weakly excites
- L1:67 weakly excites
- L19:943 weakly excites
Feeds later layers
- L24:205 suppresses
- L25:6 weakly excites
- L25:1598 weakly excites
- L25:1740 weakly excites
- L26:65 weakly excites
- L23:2802 weakly excites
- L27:1127 weakly suppresses
- L23:958 weakly excites
- L27:289 weakly suppresses
- L23:507 weakly excites
- L27:720 weakly excites
- L25:2961 weakly suppresses
- L27:2634 weakly suppresses
- L23:379 weakly suppresses
- L27:556 weakly excites
- L24:930 weakly excites
- L25:1856 weakly suppresses
- L23:1860 weakly excites
- L24:1184 weakly suppresses
- L23:2120 weakly excites
- L27:743 weakly excites
- L24:1547 weakly excites
- L26:614 weakly excites
- L26:1091 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
- oshi
- ness
- sẵn
- avail
- ARY
- ********************************************************************************
- 可想
- disp
- depth
- alysis
Suppresses
- 和服务
- and
- RequestId
- 和
- lation
- _-
- ству
- apsulation
- And
- jud