"one" after an ordinal or superlative
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. 2 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 "one" when it follows words like "most important", "last", or "next", as in "the last one". Its firing is weak compared with the strongest neurons. Turning it up made the text break down into loose fragments.
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.
-
well as a sculpture of Roman goddess Minerva commanding attention from the pediment.⏎Among the many notable events that occurred during the building’s lifetime, the most memorable one perhaps was the assassination of Lt Col NS Simpson, the infamous Inspector General of Pr
-
of different chemical products, such as carboxylic acids. The rate at which this occurs depends on a number of factors. Temperature is perhaps the most critical one, since the rate of oxidation doubles for every rise of
-
airports. The most important one: the ability to exert influence beyond the fences of the airport. Since the Rob
-
”. This makes the readability of your PowerPoint smoother and not awkward. Headings should be the biggest of whatever point from the scale you start at, then the next one down for your dot points.You should
-
.⏎Pluto is the last one of the major planet. It's definitely different from the others. Discovered in
-
less likely to develop the disease.⏎Researchers have long been aware of an apparent link between blood-vessel health and Alzheimer’s, but this study is “the first one that shows that treating mild cognitive impairment can have results,” says Anton Porsteinsson
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: what is the value of the total. the total is the sum. the total is the sum. the total is the sum. the total is the sum. the total is the sum. the total is the sum. the total is the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
now witnessing the effect these linguistic variations are having on children born into the computer age with such a high level of access in and out of schools,” the Scotsman quoted report author Lucy Jones, a former student at Manchester University, as saying.⏎“They
-
while another group did not get an apology. The participants who received the apology felt better and viewed the offending child as nicer as well as more remorseful.⏎Another study exposed children to a more distressing event: A person knocked over a tower that
-
be to assume that diet soda is a better option. Unfortunately, research has also found a connection between diet soda and brain health.⏎One study published in Stroke found that participants who had consumed one diet soda daily were almost three times more likely to develop
-
.⏎Two of Wenzel’s findings are relevant to punishments handed down by professional sports leagues. In a 2010 study, Wenzel found that people tend to seek restorative justice rather than retributive justice when they perceive
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
- L1:1322 suppresses
- L1:425 excites
- L1:551 weakly suppresses
- L1:405 weakly excites
- L1:2147 weakly suppresses
- L0:2509 weakly excites
- L1:1639 weakly suppresses
- L1:1691 weakly excites
- L0:2359 weakly suppresses
- L1:2425 weakly excites
- L1:1829 weakly suppresses
- L1:439 weakly excites
- L0:1572 weakly excites
- L0:2281 weakly excites
- L1:2170 weakly suppresses
- L0:1907 weakly suppresses
- L0:1591 weakly excites
- L0:1225 weakly excites
- L1:166 weakly excites
- L1:237 weakly excites
- L0:1264 weakly excites
- L1:684 weakly suppresses
- L1:407 weakly suppresses
- L1:1893 weakly excites
Feeds later layers
- L5:93 suppresses
- L7:267 excites
- L3:618 excites
- L3:93 excites
- L5:188 excites
- L13:2679 excites
- L4:484 excites
- L4:29 excites
- L7:2571 excites
- L3:30 weakly excites
- L5:898 weakly excites
- L4:443 weakly excites
- L8:1822 weakly excites
- L16:2375 weakly excites
- L12:30 weakly suppresses
- L7:34 weakly suppresses
- L12:2575 weakly excites
- L5:167 weakly excites
- L9:277 weakly suppresses
- L14:2555 weakly excites
- L14:625 weakly suppresses
- L8:969 weakly excites
- L14:1389 weakly excites
- L12:1702 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
- レーション
- 顺应
- 网讯
- ско
- opposition
- 这样一个
- ategorie
- south
- Boundary
- بع
Suppresses
- nawet
- tered
- rehe
- anytime
- tering
- ABCDEFGHI
- CERT
- SEE
- even
- ocha