'only' in 'not only ... but also'
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. 18 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 word 'only' in the phrase 'not only', where a sentence is about to add a second point. Turned down, the model started repeating words like 'only' and 'both', which matches the same idea of paired contrasts.
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.
-
it will be.⏎Don't forget to stretch⏎This will be one of the habits that your kids will really appreciate having learned, later on in life. Not only does this help reduce injury during sports, it will promote over all comfort and well
-
, and for wading birds grasslands offer nesting sites. Collisions between birds and aircraft, a.k.a. bird strikes, are a possible hazard. Not only do bird strikes lead to billions of euros in damages to aircraft, they
-
message to the students that you really do want to hear from all of them, not just the ones who raise their hands. Not only that, you’re giving each student a chance to consider the question, weigh arguments
-
if the electricity is imagined as water, the voltage corresponds to the water pressure, in simple terms, it is the power of electricity. Here, we will not only introduce to you where you’ll find volts in your daily life, but also list
-
mobile power sources as an option. As the need for uninterrupted power has grown, so too has the cost of traditional uninterruptible power systems (UPS). Not only are these systems expensive to purchase and maintain, but they also take up significant space
-
and should not be used outdoors in windy places; jet suction nozzles such as cedar or springs are more sensitive to water level changes. When using, not only should pay attention to water level changes, but also design in the pool body There
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: even more than not even even most even most even most most even most even most even most even most most even most even most even most even most even most most most most most even most most most most most most most even most most most most
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in a positive light. History reveals that this was not always the case. For centuries, Protestants engaged with Irish on their own terms, sometimes for academic reasons but often because it was their everyday language and an integral part of their lives. '
-
, which helps explain why they've been so easily and so largely forgotten.⏎On earlier occasions, too, the part played by Greece has been pivotal - sometimes positive but sometimes negative. And history has recently been repeating itself in other ways, for the
-
of young Protestants towards the Irish language, carried out through questionnaires and focus groups. Some of the students had attended a Gael-Linn language enrichment course, but the rest had little or no exposure to Irish. The results of this research are
-
those in turn were sometimes divided yet again, until the seemingly magical figure of 100 was reached in
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
- L16:1374 excites
- L14:2232 excites
- L10:707 excites
- L17:1730 excites
- L10:339 weakly suppresses
- L4:661 weakly excites
- L5:1755 weakly excites
- L9:2802 weakly suppresses
- L5:2182 weakly excites
- L15:699 weakly excites
- L17:698 weakly suppresses
- L11:583 weakly excites 'both' after a paired subject
- L15:1671 weakly excites
- L4:962 weakly excites
- L14:144 weakly excites
- L1:1977 weakly suppresses
- L15:2336 weakly suppresses
- L13:1792 weakly suppresses
- L16:1059 weakly excites
- L13:2062 weakly excites
- L14:1394 weakly excites
- L4:151 weakly excites
- L10:49 weakly excites
- L12:1165 weakly excites
Feeds later layers
- L26:846 suppresses
- L24:291 suppresses
- L27:655 weakly suppresses
- L27:938 weakly excites
- L20:1684 weakly suppresses
- L21:1349 weakly excites
- L20:2030 weakly suppresses
- L20:2494 weakly suppresses
- L22:1410 weakly suppresses
- L25:1997 weakly suppresses
- L20:1519 weakly suppresses
- L22:477 weakly suppresses
- L20:1494 weakly excites
- L24:887 weakly suppresses
- L20:114 weakly suppresses
- L20:2909 weakly suppresses
- L19:2367 weakly suppresses
- L22:2150 weakly suppresses
- L27:948 weakly suppresses
- L21:280 weakly suppresses
- L25:1957 weakly excites
- L22:2219 weakly excites
- L26:1143 weakly excites
- L21:555 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
- redict
- fert
- Fans
- &w
- inue
- ships
- ctions
- 的精神
- pires
- 一站
Suppresses
- both
- both
- only
- Both
- 既有
- mainly
- Both
- 不仅
- BOTH
- 更多的是