comma after an introductory word like then or when
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. 6 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 a comma that comes right after a short opening word such as "Then", "When", "How" or "If", where a clause is about to be set apart. Turning it up gave text with no obvious link to the comma pattern.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
McIntire started out on his journey westward not knowing just where Fortune would lead him. He spent almost a year in Missouri, driving a stage coach. Then, in company with three companions, he drove a herd of sheep from Missouri to central
-
that take this arithmetic thinking into account or, more recently, developing approaches to elementary school mathematics that build foundations of algebraic reasoning
-
remains poorly understood,” the Editors say. “This has led to large uncertainties in modelled estimates of terrestrial carbon storage and carbon cycle–climate feedbacks.” How, then, can important measures with such large uncertainties provide any confidence about temperature predictions for
-
.⏎We were confined three days and nights⏎Within those iron bars,⏎When, oh! the welcome news then came,⏎To get abound the cars.⏎They marched
-
can use some, or all, of this content as you see fit, whether that’s to compliment, vary or boost work set by your child's school or, of course, just to give some extra structure to your usual EdPlace approach.⏎
-
an epileptic," but rather "a person with epilepsy."⏎When writing a story or advertisement, the writer should use the term "people with disabilities" exclusively or, at a minimum, as the initial reference. Subsequent references can us terms like
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: the problem of a person who is not in a good position to live. But I think that this is not the same as being a good person. So, I am thinking about a person who is not in a good position to live,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
article, Brian Sekita discusses how GIS is vital to all phases of the emergency management process.⏎Natural disasters such as hurricanes, floods, tornadoes, earthquakes, and volcanic eruptions have impacted the globe since the beginning of time. Technological improvements
-
follow.⏎But heatwaves do happen anyway. To find out if the current heatwave was made more likely by climate change, a team at World Weather Attribution led by Friederike Otto of the University of Oxford, UK has conducted a rapid-response
-
ase.⏎The endonuclease enables for the virus to trick someone into producing viral proteins, which join into new viruses that spread the infection.⏎The team led by EMBL head Stephen Cusack analyzed crystals of endonuclease from
-
(ICTs) can mitigate many issues in SIDS and help to diversify their economies.⏎She added that innovations such as artificial intelligence, blockchain, drones, and mobile money were enabling progress towards sustainable development in SIDS, but this is constrained
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
- L5:694 strongly excites
- L5:1821 excites
- L5:395 suppresses
- L3:373 suppresses
- L2:89 suppresses
- L5:351 suppresses
- L2:180 suppresses 'is' after focus or goal nouns
- L5:2747 weakly excites
- L4:12 weakly suppresses
- L2:2124 weakly suppresses
- L5:995 weakly excites
- L5:99 weakly excites
- L3:552 weakly excites
- L5:781 weakly suppresses
- L5:1217 weakly excites
- L3:466 weakly excites
- L5:976 weakly excites
- L2:1394 weakly excites
- L4:1587 weakly excites
- L3:700 weakly suppresses
- L5:2643 weakly suppresses
- L2:24 weakly excites
- L4:139 weakly suppresses
- L2:1185 weakly suppresses
Feeds later layers
- L7:288 excites
- L10:142 excites opening parenthesis mid-sentence
- L14:220 excites
- L17:2787 excites comma after 'that' or a verb before an aside
- L13:1721 suppresses
- L16:2571 excites
- L9:819 excites
- L9:1371 excites
- L17:2597 excites
- L15:1213 excites
- L16:1767 excites
- L14:774 suppresses
- L14:2402 excites
- L19:1864 excites
- L16:1253 excites
- L13:897 suppresses
- L15:770 excites
- L11:1272 excites unclear: sentence ends in factual prose
- L13:685 weakly suppresses
- L17:2084 weakly suppresses
- L21:2772 weakly excites
- L20:1466 weakly excites
- L8:204 weakly suppresses
- L13:2609 weakly 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
- shaw
- __,
- ,**
- 为主体
- rows
- **,
- *,
- awning
- 为主的
- -,
Suppresses
- opping
- cg
- ORM
- ARENT
- slick
- ogonal
- 功课
- cg
- budget
- cj