"well" after "how"
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 the word "well" in phrases like "how well," asking about how effective or good something is. It also dips on the ends of some adverbs like "textually" or "sparsely." Turning it up did not produce any clear theme in the sample text.
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.
-
ority homes during the instructional period. We administered follow-up assessments in the fall after the instructional period ended and in the spring of the following year to determine how well students maintained and consolidated target academic words. Students in the intervention group maintained their relative
-
5 year old graduation<|endoftext|>Look at the egg⏎picture. Write the name of the egg part in the box. Click the 'check'⏎button to see how well you did. When finished⏎print the worksheet to learn more.⏎Thank you for
-
experiencing through the goggles. The researchers were testing a phenomenon called “path integration” to determine how well the body was able to keep track of itself and its movements within the physical plane
-
that this tailor-made Chinese intervention program is very effective for improving word reading fluency of Chinese ordinary students and those with reading difficulties. For factors that may affect how well children learn to read, it appears that school resources, student academic ability, and
-
up to a year on your maintenance dose to notice a true change. If the treatment is successful, you will continue maintenance shots for three to five years.⏎How well do these shots actually work?⏎As with any treatment, effectiveness will vary from person
-
cadmium in children’s products– today the agency backed down. They are suggesting that the industry self-enforce. Hmmmm? How well did that work out the last time when we were talking about lead. We are
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 been using the word "certain" in a way that's problematic for my social media. I have been using the word "certain" in a way that's problematic for my social media. I have been using the word
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
This coursebook can be used to support any pedagogical or research activities in text technologies, the history of the book, the history of information, and textually based work in the digital humanities. Through careful explanations of the field, examinations of
-
most populous of India’s northeastern states, is an expansive green of rice paddies, bamboo, banana, and sal trees. Bamboo huts dot its sparsely populated hamlets, ringed by low hills in the distance. Along the narrow
-
and construction, complement this landscape. The intimate paths show a human scale appropriate for the private world the Abkhazis wanted to create for themselves.⏎The sculpturally strong Garry oaks predominate the site and provide a unifying sense
-
-bedded poorly sorted sand and gravel. The aquifer is composed of hydraulically connected alluvial and terrace deposits that unconformably overlie the Permian
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
- L4:2743 suppresses
- L0:828 suppresses
- L4:521 suppresses
- L1:1049 excites
- L3:1774 suppresses
- L3:530 excites
- L4:1743 excites
- L2:2960 excites
- L5:125 suppresses "of" after all three, both, ten
- L2:2932 excites
- L2:166 suppresses
- L5:2841 excites
- L0:1383 excites
- L1:818 weakly excites
- L3:682 weakly excites
- L3:1717 weakly suppresses
- L1:1239 weakly suppresses
- L4:1763 weakly excites
- L4:2722 weakly excites
- L3:6 weakly suppresses
- L4:450 weakly suppresses
- L3:228 weakly suppresses
- L5:1214 weakly suppresses
- L5:2641 weakly excites unclear: scattered word fragments
Feeds later layers
- L8:252 strongly suppresses 'of' after 'each' or 'one' or 'all'
- L17:985 strongly excites
- L7:173 strongly excites
- L15:557 strongly suppresses
- L11:810 suppresses
- L8:2755 excites
- L10:167 excites
- L21:716 excites
- L18:602 excites
- L19:663 excites
- L18:217 suppresses
- L16:1433 excites
- L14:280 suppresses
- L12:1058 excites
- L20:519 suppresses
- L10:289 suppresses
- L21:274 suppresses
- L17:1694 suppresses
- L17:831 excites
- L22:404 excites
- L21:68 suppresses
- L7:2164 excites period inside country abbreviations
- L18:452 suppresses
- L17:199 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
- ouble
- 时候
- ones
- counterpart
- ほうが
- counterparts
- pieces
- 片子
- obstacle
- 時候
Suppresses
- 的话题
- ric
- olen
- ymm
- eut
- rics
- ERICAN
- CGRect
- entials
- 的人都