"with" after a group noun
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. 18 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
The neuron fires on the word "with" when it follows a group of people or things, as in "countries with" or "people with". It goes strongly negative on "with" in phrases like "with confidence" or "with permission". Turning it up gave mixed, unrelated 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.
-
. From Flint Michigan and California in the United States of America to Chennai, India and Cape Town, South Africa, 12 out of 17 countries with the highest water stress are in the Middle East and North Africa.⏎Flooding,
-
2002 and 11.7 percent since 1997, the analysis shows.⏎Texas tops the list of the 10 states with the biggest one-year increases in CO2 emissions, with Georgia, Arizona, California
-
need to plant 6,631,898 trees in order to offset trips for less than 1 mile per year.⏎(Of course for anyone with mobility issues, these car journeys are necessary; therefore, the analysis accounted for the
-
everything run more efficiently is an amazing adaptation. And for many people in the modern world this is still a lifesaving mechanism. But for the people in the world with easy access to Coca-Cola and McDonald’s its making it harder to stay at a
-
slightly higher in men, because they tend to spend more time in the sun and use less sun protection than women do. African-Americans, Hispanics, Asians and others with darker skin are not as susceptible as Caucasians (Skin Cancer and Skin of Color
-
for the following reasons:⏎- African Americans still reside mainly in separate and unequal communities. In 2010, in the 100 metropolitan areas with the largest African American populations, 62.5 percent of blacks would have
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 purpose of a bit of care, as in the care of as much frequency precision as possible, in the purpose of precision precision, as in the precision of the frequency as precision, as in the precision of a precision, as in the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
’s book appeared, many other anthropologists have declared their acceptance of the theory, so that to-day we may assert with confidence that the black man is black because of the excessive sunlight of his environment,
-
, without safety screening⏎Reproduced with permission from Fluechtling.net⏎- Millions of refugees guaranteed German standard of
-
the biggest enemies of our health unless they are eaten very sparingly.⏎Nutrition-health relationships have been established for decades as highly related so that today we can say with certainty that eating habits and life are critical to promoting health and by contrast, are
-
herdsmen began storing milk and discovered (probably by accident) that fermented milk was safe to consume.⏎With such an extensive backstory, it’s impossible to know with certainty who invented yogurt. One thing is for sure: the history of yogurt and
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
- L17:2363 suppresses
- L15:1162 suppresses
- L2:863 weakly suppresses
- L7:1056 weakly excites
- L9:31 weakly suppresses
- L4:882 weakly suppresses 'with' after provide or familiarize
- L2:2889 weakly excites
- L4:228 weakly suppresses
- L16:1860 weakly suppresses
- L9:1856 weakly excites "with" after a noun, introducing contents
- L11:1108 weakly suppresses
- L12:73 weakly suppresses
- L0:2342 weakly suppresses
- L15:986 weakly excites
- L6:738 weakly suppresses
- L15:1913 weakly suppresses
- L16:364 weakly excites
- L13:1087 weakly suppresses
- L6:2708 weakly suppresses
- L2:45 weakly excites
- L14:2116 weakly suppresses
- L17:964 weakly suppresses
- L1:1188 weakly excites
- L2:1321 weakly suppresses
Feeds later layers
- L22:2587 suppresses
- L23:2881 suppresses
- L19:1869 suppresses
- L19:1818 suppresses
- L26:1318 suppresses
- L19:1739 suppresses
- L19:2407 suppresses
- L20:1733 suppresses
- L20:1501 suppresses
- L22:1823 weakly suppresses
- L25:2637 weakly excites
- L22:809 weakly suppresses
- L24:521 weakly suppresses
- L23:142 weakly excites
- L24:921 weakly excites
- L19:879 weakly excites
- L21:2352 weakly suppresses
- L20:1900 weakly excites
- L19:1871 weakly excites
- L23:1587 weakly suppresses
- L25:799 weakly suppresses
- L19:2723 weakly suppresses 'of' after a quality or importance adjective
- L22:507 weakly suppresses
- L22:2814 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
- 一族
- спект
- 诱人
- !$
- прод
- 三个方面
- 单身
- cased
- 其次
- –⏎⏎
Suppresses
- precision
- ease
- certainty
- precision
- 精度
- difficulty
- efficiency
- accuracy
- outh
- Precision