the word 'I' as subject
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. 26 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
It fires on the word 'I' when it starts a clause, as in 'I am working' or 'I think'. Turning it down made the model repeat 'am' over and over, and turning it up pushed the writing toward first-person talk.
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.
-
itself, as in "I am doing my genealogy" or "I am working on my genealogy." What does that statement communicate to a non-g
-
out of the kindness of my heart, because they have some historical significance (as in the case of my review of the LM741) or most likely I think the technology is important and interesting. All opinions are my own and I would
-
proclaiming, “I am an American . . . I have constitutional rights.”⏎- A group called Health Freedom Idaho protested the Boise mask mandate
-
This scientist article is a stub. You can help Wikiquote by expanding it.|⏎- It is because I know all that science can bring to the world that I shall continue my efforts to ensure that it contributes to the happiness of all men,
-
the Lord shone around them, and they were filled with great fear. 10 And the angel said to them, “Fear not, for behold, I bring you good news of
-
, 2014. Last Updated: August 18, 2014.<|endoftext|>The more I read about this, the more confused I become. Illuminance, as I understand it, measured in lux, is the
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: am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am am
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
industry, the main sectors include: food, mining, light, oil, engineering.⏎THE ECONOMY OF VIETNAM TODAY⏎Now Vietnam’s economy is developing rapidly, and for a couple of
-
⏎culture is their culture; that same Aramaic language is their language.⏎stele of King Neram-Seen⏎THE PEOPLES OF MESOPOTAMIA⏎THE SUMERIANS— The history of⏎Ancient Iraq
-
and Decanters⏎E for Equipment⏎F for Fermentation⏎G for Glass and Grape⏎H for Harvest⏎I for Infusion⏎J for Jerobo
-
,000,000,000 Styrofoam cups each year. Styrofoam is NOT recyclable; it cannot
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
- L25:2560 strongly suppresses
- L25:2082 excites first-person 'I' starting a clause
- L25:2173 excites
- L22:1302 excites
- L25:1061 suppresses
- L25:2985 excites
- L23:2342 suppresses
- L18:569 excites
- L23:1871 excites
- L25:2152 weakly excites
- L25:1936 weakly suppresses
- L24:3009 weakly excites
- L22:1996 weakly suppresses
- L23:1712 weakly excites
- L25:983 weakly suppresses
- L23:1521 weakly suppresses
- L23:2864 weakly excites
- L25:2572 weakly excites
- L23:319 weakly suppresses
- L24:1084 weakly suppresses
- L24:2531 weakly suppresses
- L2:1629 weakly excites
- L25:1554 weakly excites
- L8:16 weakly suppresses
Feeds later layers
- L27:464 strongly suppresses
- L27:1292 suppresses
- L27:871 weakly suppresses
- L27:236 weakly suppresses
- L27:3014 weakly suppresses
- L27:422 weakly suppresses
- L27:1130 weakly suppresses
- L27:618 weakly suppresses
- L27:868 weakly excites
- L27:2088 weakly suppresses
- L27:2829 weakly suppresses
- L27:908 weakly excites
- L27:713 weakly excites
- L27:1599 weakly excites
- L27:2917 weakly suppresses
- L27:2775 weakly excites
- L27:2243 weakly suppresses
- L27:1593 weakly excites
- L27:2532 weakly excites
- L27:2200 weakly suppresses
- L27:2626 weakly excites
- L27:437 weakly suppresses
- L27:1256 weakly suppresses
- L27:854 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
- úa
- lessness
- ocrates
- ucas
- скачать
- sını
- keletal
- отов
- itous
- iev
Suppresses
- am
- am
- (am
- Am
- AM
- Am
- -am
- AM
- _am
- _AM