'and' inside organisation or paired names
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. 19 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 'and' when it joins parts of a named organisation or subject, such as 'Health and Medical Sciences'. It goes strongly negative on 'and' or 'or' linking a pair of contrasting words, and steering produced repetitive lists of 'and' or random words.
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.
-
, it is important to understand where people are acquiring information about the virus, whom they trust, and how to overcome mistrust and information. The Center for Health and Research Transformation (CHRT) and the Department of Learning Health Sciences (LHS
-
ongoing concern. And the challenges brought on by the unprecedented COVID-19 pandemic created even more roadblocks for many of us.⏎According to a survey from Environment and Climate Change Canada, just
-
Day, Dr Dan Horton and Professors Susan Lanham-New, Christian Heiss and Roberto La Ragione from the Faculty of Health and Medical Sciences explore the factors that would influence a healthier world for all.⏎What would
-
Education Level: N/A⏎Audience: Teachers; Practitioners⏎Sponsor: Pennsylvania State Dept. of Education, Harrisburg. Bureau of Adult Basic and Literacy Education.⏎Authoring Institution: Lutheran Children and Family Service, Philadelphia,
-
. 2005. Edge Effects in Line Intersect Sampling With. Journal Agricultural, Biological, and Environmental Statistics 10:460-477⏎Get the
-
higher education, particularly in respect of emerging models of teaching and learning.⏎Professor Laura Czerniewicz, director of UCT’s Centre for Innovation in Learning and Teaching (CILT), who was a principal investigator on the study along with Professor
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: a person who a person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person person
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
whether it is practising gratitude yourself, with your family, partner, or finding ways to cultivate gratitude at work. The book explores why gratitude is different for men and women and the impacts this has on relationships, romantic or not. It provides practical
-
the amount of random error involved in measuring student performance. The risk is that teachers and schools may be wrongfully rewarded or punished because value-added techniques either over- or underestimated their students’ learning gains. However, no intelligent users of standardized testing would
-
, and blood pressure cuffs for village health workers to identify high-risk pregnancies and life-threatening post partum and neonatal infections.⏎Terma’s training on pre- and post-natal care (birthing practices, lactation and infant health) has
-
a pin), they are transferred to a tank in a building closer to the water, where water from the Peconic Bay (the water body between the North and South Forks of Long Island) is pumped through their tanks and they feed on
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
- L18:2863 strongly suppresses
- L18:1413 strongly excites
- L16:1659 excites
- L17:2886 excites
- L12:51 excites
- L18:2238 suppresses
- L8:70 excites 'and' between stacked superlatives
- L9:92 suppresses "and" or "or" joining two items
- L16:2498 excites
- L18:104 suppresses
- L17:2250 suppresses
- L14:440 suppresses
- L16:520 suppresses
- L13:977 suppresses
- L16:1388 suppresses
- L16:376 suppresses 'or' joining alternatives
- L4:57 suppresses
- L16:492 suppresses
- L14:745 suppresses
- L18:1738 excites
- L17:1575 excites
- L17:1740 suppresses
- L16:2577 excites
- L14:706 excites
Feeds later layers
- L20:955 suppresses
- L20:439 suppresses
- L20:160 suppresses
- L21:167 weakly suppresses
- L22:202 weakly excites
- L20:165 weakly suppresses
- L20:1200 weakly suppresses 'and' in 'more and more' doubling
- L23:107 weakly excites
- L23:319 weakly suppresses
- L20:156 weakly excites
- L20:80 weakly excites
- L21:247 weakly excites colon after a said or explains verb
- L21:647 weakly suppresses
- L27:2150 weakly excites
- L20:331 weakly excites
- L20:969 weakly excites
- L20:337 weakly suppresses
- L21:1074 weakly excites
- L24:216 weakly excites
- L27:2544 weakly excites
- L20:300 weakly excites
- L26:980 weakly excites
- L22:168 weakly suppresses
- L22:255 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
- MaxY
- defensively
- DUCT
- lessly
- Davis
- ,{"
- vided
- 崖
- .plus
- ,Q
Suppresses
- ätze
- inion
- aska
- pired
- paced
- oding
- }`,
- ized
- ehen