comma before 'but' in contrasts
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 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 fires on the comma that comes just before a 'but' in phrases like 'not only X, but also Y'. Turning it down made the model repeat the word 'but' over and over, which fits the word it helps set up.
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.
-
's used to treat mood disorders such as:⏎- mania (feeling highly excited, overactive or distracted)⏎- hypo-mania (similar to mania, but less severe)⏎- regular periods of depression, where treatment with other medicines has
-
5 minutes)⏎Today we will learn that prayer is communicating or talking with God. We may not be able to call God on the telephone or send Him a letter, but prayer is just as real as e-mail or newspapers. Why should we pray
-
are not properly maintained, your home will not only feel very uncomfortable to live in, but you and your family may also become susceptible to respiratory disorders or chemical reactions.⏎
-
iveness while living in a Christian world. It will be suggested that Mudejar self-perception and group-perception were founded not only on religious belief per se, but also on⏎the perpetuation of a social world the organization and mores
-
a time of preparation for Easter. From the church’s earliest days, the resurrection of Christ was celebrated not only each week (on Sunday, the Lord’s Day), but also in a special festival of the resurrection. This festival we call Easter Day
-
due to the corrupted values popularized by the American Dream.⏎Ultimately, Tom is quite saddened by Myrtle’s death, not because he loved her as a person, but because he loved having control over her. Now that she is dead, Tom
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: but but but but but but but but but but but but but but but but but but but but but but but but but可可可可可可可可可可可可可可可可可可可可可可可
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
1949, the right to asylum was laid down in the Constitution. Until 1993, Article 16 of the Constitution (Grundgesetz)
-
new era of hope. Martin Luther King Jr. led the civil rights movement to new heights, ultimately leading to the signing of the “Civil Rights Act” of 1964. The whole world watched in awe and amazement as
-
the other molecule. Hence such an arrangement provides denser packing than one possessing mirror planes. Lucretius over 2000 years ago proclaimed in his De rerum natura :⏎“Things whose fabrics show opposites that match,⏎one
-
168 note) who remarks: — "Um dies (i.e. Varro's description of the extent of Roma quadrata) mit dem unten erörterten Pomerium und der Auffassung des Dionys.
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
- L21:478 strongly suppresses
- L20:1764 strongly suppresses
- L20:1194 strongly suppresses
- L19:2294 strongly suppresses
- L17:2269 strongly excites
- L20:1309 strongly excites
- L16:788 strongly excites
- L15:343 strongly excites
- L19:3030 strongly suppresses
- L15:844 strongly suppresses unclear: verbs in health and risk claims
- L24:2282 strongly excites
- L16:2778 strongly suppresses
- L15:2976 suppresses
- L17:762 excites
- L16:2865 suppresses
- L16:2373 suppresses
- L14:2135 excites
- L17:2817 suppresses
- L14:1903 suppresses
- L14:2480 suppresses
- L15:2542 suppresses
- L18:2396 suppresses
- L14:1531 excites
- L14:2809 excites
Feeds later layers
- L27:568 strongly suppresses
- L26:1232 strongly suppresses
- L27:1084 strongly suppresses
- L27:2009 suppresses
- L27:760 suppresses
- L27:1698 weakly suppresses
- L26:190 weakly suppresses
- L27:1088 weakly suppresses
- L26:157 weakly excites
- L26:83 weakly excites
- L26:1723 weakly suppresses
- L26:836 weakly excites
- L27:245 weakly excites
- L27:1678 weakly excites
- L27:2951 weakly excites
- L27:789 weakly suppresses
- L27:1104 weakly excites
- L27:2228 weakly excites
- L27:2651 weakly suppresses
- L27:1491 weakly suppresses
- L26:275 weakly suppresses
- L26:347 weakly suppresses the 're' in self-reliant and similar
- L26:456 weakly excites
- L27:601 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
- TimeUnit
- ends
- oud
- htag
- zept
- ActionResult
- @\
-
- CellValue
- остью
Suppresses
- 但
- but
- but
- 但是
- 然而
- though
- 但我
- 但在
- however
- However