"on" after "of all living things" phrases
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. 22 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 "on" in phrases like "the hardiest animals on the planet" or "life on Earth", where a place or world comes next. Turned up, the model talked about the Earth and its surface over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
.<|endoftext|>Even the hardy tardigrade might not stand up to climate change⏎Tardigrades are well-known for being one of the hardiest animals on the planet, but even they have their limits. New research has shown that they
-
and let your imagination run WILD!<|endoftext|>Baleen and sperm whales, known collectively as the great whales, include the largest animals in the history of life on Earth. With high metabolic demands and large populations, whales probably had a strong influence
-
, digestive problems, numbness in the extremities, and skin cancer.⏎Water is the essence of life, but not just human life. Almost every living thing on the planet must have water to survive, and that includes a whole host of bacteria
-
Mamba. …⏎- White Rhinoceros.⏎What is Africa’s deadliest animal?⏎Despite being vegetarian, hippos are considered the most dangerous terrestrial animal on the African continent. On average, they are responsible for the deaths of almost
-
and many other things dirty. There are many type of pollutions such as air pollution, water pollution and soil pollution. Air pollution is harmful to all living organisms on this earth. Air pollution can exist in many type of states such as open burning
-
ocean...⏎and hydrologic sciences⏎Hydrology is the study of the movement, distribution, and quality of water on Earth and other planets, including the hydrologic cycle, water resources and environmental watershed
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: in the context of the future, we will have the same as in the past, but with the same content in the same capacity, and the same use, and the same identity, in the same content, in the same context, in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Everything You Need⏎Who was the first president to live in the White House? Learn about this
-
reality of his actions is that he only speaks in order to appear sincere as a politician, to look good rather then actually be good. ” to thine own self be true, And it must follow, as the night the day, Thou can
-
Twitter. The Nintendo Wii. Google Maps. Google Translate. Artificial hearts and livers. Face transplants. Self-cleaning windows. Invisibility cloaking. Graphene. The fuel-cell bike
-
the entire Saturn system. A Titan colony is self sustaining as long as it has power to supply these processes, and invaluable food resources
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
- L20:2975 suppresses
- L19:2280 suppresses
- L21:1210 suppresses
- L0:2559 excites
- L21:3017 suppresses
- L21:797 suppresses
- L12:1693 weakly suppresses
- L13:1213 weakly excites
- L20:1722 weakly suppresses
- L19:2967 weakly excites
- L17:1165 weakly excites
- L2:2795 weakly excites
- L9:2862 weakly suppresses
- L9:1120 weakly excites
- L12:2240 weakly suppresses
- L15:349 weakly excites
- L21:1949 weakly suppresses
- L1:2478 weakly suppresses
- L21:2469 weakly suppresses
- L5:1318 weakly excites
- L21:792 weakly suppresses
- L11:1 weakly excites
- L21:508 weakly excites
- L16:2690 weakly excites
Feeds later layers
- L26:948 excites
- L27:995 excites
- L27:2905 excites
- L25:530 excites
- L24:2125 excites
- L26:2384 weakly excites
- L25:2813 weakly excites
- L24:2651 weakly excites
- L25:1282 weakly excites
- L25:312 weakly excites
- L25:2811 weakly suppresses
- L25:2432 weakly suppresses
- L24:2566 weakly excites
- L25:2895 weakly excites
- L26:1910 weakly excites
- L26:981 weakly excites
- L24:716 weakly excites
- L24:1842 weakly excites
- L24:1054 weakly excites
- L23:152 weakly excites
- L23:1455 weakly excites
- L27:2244 weakly excites
- L25:499 weakly excites
- L24:3011 weakly excites 'of' in national and founding phrases
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
- earth
- Earth
- 地面
- earth
- ground
- 地球
- Earth
- -earth
- 地
- 地上
Suppresses
- UserManager
- _SCOPE
- AGIC
- 前瞻性
- oping
- ryptography
- ileged
- UserService
- ?>><?
- cox