'to' in 'according to'
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. 7 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 lights up on the word 'to' when it comes right after 'according', as in 'according to the CDC'. When it was turned up, the model's writing did not visibly change, so the word pattern is the only clear signal.
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.
-
of Americans who have a history of cancer was 11.7 million in 2007, or nearly 4% of the population, according to new figures from the Centers for Disease Control and Prevention (CDC).⏎That's up
-
you why…⏎Public health is awesome (and I’m not just saying that). Thanks to the achievements of public health, we live longer and healthier lives. According to the National Center for Health Statistics, “life expectancy at birth among U.S.
-
. coli before it enters a newly protected stream. Estimated costs for installing disinfection equipment and operating it for the first year are about $95 million, according to department figures.⏎Without major improvements and more costs, though, disinfection could be
-
about cybersecurity and how it affects your SMB:⏎Why do hackers target small businesses?⏎The size of your corporation doesn’t guarantee your safety against cybercriminals. According to Stephen Cobb, a senior security researcher at the antivirus company ESET, small businesses
-
). Those fires resulted in 155 deaths, 1,100 injuries and $176.1 million in direct property damage, according to NFPA.⏎In addition, NFPA reports that twice as many home candle fires
-
to minimize symptoms.⏎Why are so many alcoholics skinny?⏎More likely, it results from interference with the body’s ability to derive energy from other foods. According to Lieber’s report, experiments in laboratory animals and in heavy drinkers found that alcohol
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: 13, 12, 12, 12, 12. I need to find the six 1s. The 12 and the 12 and the 12. I think I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
after contracting influenza was pneumonia. However, those living in the sunnier communities had lower death rates after developing influenza than those living in the less sunny communities. Vitamin D helped in two ways: first by reducing the aggressive immune response that among other things
-
and farm managers with a global perspective on a range of potential environmental effects and their relative intensity,” said Dr. Michael Rubino, director of NOAA’s Fisheries Office of Aquaculture. “It is a tool that can be used when evaluating proposed
-
0 mg per day⏎1 to 3 years old: 700 mg per day⏎Your diet should also include adequate amounts of vitamin D. Vitamin D is essential for the body to absorb calcium. A lack of vitamin D inhibits
-
in this SPM can be found in the glossary of the SRREN (Annex I).Conventions and methodologies for determining costs,
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
- L5:2208 excites
- L6:2182 excites
- L6:2759 excites
- L6:773 suppresses
- L6:1530 suppresses
- L6:2193 weakly suppresses
- L6:1637 weakly excites
- L6:2285 weakly suppresses
- L6:2692 weakly suppresses
- L6:2842 weakly excites
- L6:1720 weakly suppresses
- L6:1425 weakly suppresses
- L6:2977 weakly suppresses
- L6:2349 weakly excites
- L4:2679 weakly excites
- L3:2043 weakly excites
- L6:961 weakly excites
- L6:1975 weakly excites
- L5:1276 weakly suppresses
- L4:2979 weakly excites
- L6:1222 weakly suppresses
- L6:2758 weakly suppresses
- L0:2804 weakly suppresses
- L2:2751 weakly excites spiral and screw words
Feeds later layers
- L8:127 excites
- L8:941 suppresses
- L18:2681 weakly suppresses
- L9:283 weakly suppresses
- L10:1224 weakly suppresses
- L20:759 weakly excites
- L24:1175 weakly suppresses
- L13:2413 weakly excites
- L9:1392 weakly suppresses
- L8:586 weakly suppresses
- L20:638 weakly excites
- L16:1291 weakly suppresses
- L23:2060 weakly suppresses
- L8:455 weakly suppresses
- L13:2633 weakly excites
- L8:1572 weakly suppresses
- L9:2952 weakly suppresses
- L21:542 weakly suppresses
- L15:830 weakly excites
- L19:2322 weakly suppresses
- L19:1160 weakly suppresses
- L10:1178 weakly suppresses
- L9:2280 weakly suppresses
- L11:361 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
- 中共
- Together
- iero
- conjunction
- AY
- azio
- ferr
- ause
- ay
- 殿
Suppresses
- |x
- MENTS
- uds
- antry
- llx
- ancy
- reeNode
- -proof
- %x
- hotmail