"or" in either-or choices
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. 20 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 "or" when the sentence offers two alternatives, like "true or false", "whether it is X or not" or "one way or another". Turned up, the model started repeating "not not not".
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.
-
foods at the Jimmy Dean Foods division of Sara Lee Corporation discovered that breakfast for more than 60% of Americans consists of cold foods. Whether that’s true or not is debatable but what nutritionists do agree on is that it’s a
-
course antique wood and furniture restoration and the application of traditional restoration techniques and materials. In this context, one might ask: Does restoration increase the value of an object or diminish
-
the art appreciation and there have been discussions concerning its future.⏎Today, lots of people dispute whether art still has an essential duty in the social background of a nation or society. Numerous experts agree that art and also the creation of artworks is extremely essential
-
, doctors need to know if you are taking these products so they can accurately judge whether particular symptoms are related to a medical illness or a side effect from an herbal medicine. Those who take herbal products should also keep
-
ers,” he says, “So the next step is to refine these idealisations and relax the initial conditions to make it more general.” [Nature]⏎To confirm or refute this model, astronomers
-
all kinds has been going on to determine if there’s danger to people or wildlife.⏎“Air monitors in the community, throughout the area. We were testing
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: I have a 324,000} * 0.000324 + 0.000324 * 00.000324 + 00.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
?⏎a demon is a long-running background process that responds to requests for services. The term originated with Unix, but most operating systems use daemons in one form or another. On Unix, daemon names conventionally end in "d". Some examples
-
lasted in the UK until 200<|endoftext|>Technology has become a defining force in our lives, and its relevance is spreading to virtually every field in one way or another. While some people may go their entire lives never needing to recall the knowledge
-
are examples of creating markers, setting markers, and moving point to markers:⏎;; Make a new marker that initially does not point anywhere: (setq m1 (make-marker)) ⇒ #<marker in no buffer>⏎m1to point between the
-
undeserved . They will also suspect that<|endoftext|>Constipation can cause difficult and painful bowel movements of dry, hard stool. Most people deal with it at one time or another. The most common causes of the problem are poor diet and a lack of
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
- L17:2638 suppresses
- L17:1480 excites
- L19:2864 excites
- L13:2939 excites
- L18:1302 excites
- L19:2686 suppresses
- L18:1582 weakly suppresses
- L19:2544 weakly suppresses
- L8:3 weakly excites
- L18:5 weakly excites
- L13:1603 weakly suppresses
- L18:1203 weakly excites end of document marker
- L8:7 weakly excites
- L18:1586 weakly suppresses
- L19:1571 weakly excites
- L14:249 weakly excites
- L17:1407 weakly excites "and" joining questions (wh-question context)
- L2:126 weakly excites
- L18:2036 weakly excites
- L8:818 weakly suppresses
- L19:835 weakly suppresses
- L9:1098 weakly suppresses
- L18:584 weakly suppresses
- L9:2055 weakly suppresses
Feeds later layers
- L21:1044 strongly excites
- L24:334 excites
- L26:406 excites
- L22:2227 excites
- L23:434 excites
- L24:2628 excites
- L21:1538 excites
- L27:714 excites
- L21:1839 excites
- L24:776 excites
- L23:773 excites
- L25:541 excites
- L25:2646 excites 'there' after a because or if clause
- L27:838 excites
- L24:402 weakly excites
- L27:1334 weakly suppresses
- L23:223 weakly excites
- L27:233 weakly excites
- L21:749 weakly excites
- L21:744 weakly excites
- L27:138 weakly excites
- L22:539 weakly excites
- L25:1270 weakly excites
- L21:360 weakly excites 'such' introducing examples
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
- not
- 不是
- ли
- not
- 否
- Not
- Not
- 不是一个
- không
- 不在
Suppresses
- }*
- (CC
- 纽
- ⽅
- »
- 趸
- .dds
- ;c
- tré
- ístico