"of" inside titles and names
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
This neuron lights up on the word "of" when it sits inside a title or a fixed name, like "The Diary of Anne Frank" or "Music of the Spheres". When it was turned up, the model's writing drifted into rambling, repeated phrases.
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.
-
the report. "The estimate is transparent and conservative," says particle physicist Georg Weiglein at the University of Durham, UK. "I ...<|endoftext|>The Diary of Anne Frank Essay written by l
-
in American History: Selected Documents Celebrating the Manuscript Division’s First 100 Years features drafts of Langston Hughes’ poem, “The Ballad of Booker T.” In this 1941 poem,<|endoftext|>Whenever the st
-
in her lifetime, a towering legacy left behind. Rest in peace, brave comrade,” she penned.⏎Former Chair of the<|endoftext|>The Kalima or the Declaration of Faith is two short sentences and is the essence of Islam. It is given below
-
5:13-16 (The Prayer of Faith)⏎Have the kids find James 5:13-16 in
-
equal to the sum of the squares of the sides containing the right angle." But Pythagoras had an all-inclusive vision, which also included the "music of the spheres" (the sound that no doubt had to be produced by the planets
-
the 13th USENIX Security Symposium, 2004.⏎Garfinkel, S. and Shelat, A., "Remembrance of Data Passed: A Study of Disk Sanitization Practices," IEEE Security and Privacy
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: for example, the famous mathematician, 1942, 1975, 1985, 1995, 1995, 2001, 201
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the speeds of two satellites placed in stable circular orbits of radius R and 2R respectively?⏎C) 1 / 2⏎D
-
's own Apple II.⏎Jobs' vision of "good design" meant not only building aesthetically pleasing machines, but devices that had a utility of function like user-friendly GUIs akin to those seen on early Macs.⏎Near
-
Like⏎their giant counterparts, Humboldt squid are enigmatic. No one has⏎seen them mate or lay eggs. No one has watched them develop from⏎egg to adult. No one knows how many exist.⏎These tentacled titans
-
of the Journal of the American Academy of Religion eschews this empirical approach. Twelve scholars of Hinduism three of them practicing Hindus attempt to answer this question by methodical analysis. None can be said to successfully respond to the question, but their
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
- L15:1888 suppresses
- L16:1594 suppresses
- L11:2859 excites
- L11:1168 suppresses
- L14:935 suppresses
- L16:652 excites
- L17:30 excites
- L10:1644 suppresses
- L17:317 excites 'of' after a noun phrase
- L10:365 excites
- L5:493 suppresses
- L1:788 weakly excites
- L12:2554 weakly excites
- L10:155 weakly suppresses
- L17:147 weakly excites
- L17:1721 weakly excites 'by' introducing a method or means
- L4:23 weakly excites
- L16:1088 weakly excites 'of' after a noun phrase, as in 'of course'
- L17:188 weakly suppresses 'a' before a noun phrase
- L9:227 weakly suppresses
- L16:1286 weakly suppresses
- L6:16 weakly excites
- L6:352 weakly suppresses
- L16:413 weakly excites
Feeds later layers
- L19:1180 strongly excites
- L19:1023 suppresses
- L19:1427 suppresses
- L25:284 excites
- L21:1400 excites
- L25:2973 excites
- L19:2522 excites
- L19:2115 excites
- L19:626 excites
- L19:2593 excites
- L24:359 excites
- L27:250 excites
- L20:1047 weakly excites
- L24:443 weakly excites
- L24:360 weakly suppresses
- L23:79 weakly excites
- L23:258 weakly excites
- L25:2331 weakly excites
- L25:403 weakly excites
- L19:535 weakly excites
- L22:533 weakly excites
- L19:663 weakly suppresses
- L27:2096 weakly excites
- L23:1394 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
- dives
- itled
- erties
- ponents
- states
- .getError
- rates
- stances
- [val
- очных
Suppresses
- 芒
- 戶
- GORITHM
- 军
- OCUS
- 冈
- 솜
- 逊
- 户
- Axe