'of' inside title-like noun 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. 8 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 'of' when it links two nouns in a set phrase or title, such as 'Sunday of Sundays', 'Harvest of Memories', or 'Copyright of the Federal'. The steering samples do not show a clear effect, so the pattern is known mainly from the snippets.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
on Sunday, the Lord’s Day), but also in a special festival of the resurrection. This festival we call Easter Day, and it is celebrated as the Sunday of Sundays!⏎Lent, as a season of preparation, is traditionally focused on repent
-
are twice as likely as whites to have had 10 or more spells of unemployment over their prime working years.⏎-<|endoftext|>In an ideal world, a material that is inherently resistant to its service environment complies with the mechanical, formability and
-
4th of July should be a time when we come together to honor our country by celebrating our
-
Ewan Boyce and later the property of her husband Captain William Boyce<|endoftext|>A Point of View: Making friends the shared currency way⏎- 2 March 2012⏎- From the section Magazine⏎Greece is falling out
-
lected After Runoff Victory|⏎Federal Elections (94-2004): Election Results for the US Senate⏎and the US House of Representatives. Copyright of the Federal⏎Elections
-
- Get a towel (10 seconds)⏎- Answer the phone (2 minutes)⏎- Answer the front door (4-6 minutes)⏎ABC & Ds of Drowning Prevention⏎Read the following ABC and Ds of drowning prevention to learn
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: how do you know what you think? And what do you think about? What do you think about? What do you think about? What do you think about? What do you think about? What do you think about? What do you
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
lectronics has not been given much consideration previously, though much effort has been devoted to the development of general-purpose nanoelectronic systems, i.e
-
ptic. They are used traditionally to treat headaches[⏎If you have any useful information about this plant, please leave a comment. Comments have to be approved before they are shown here.<|endoftext|>Save water with wind power⏎21 March 20
-
as the behaviour of those in our local community.⏎Cultural – Listening to and incorporating elements of<|endoftext|>- A book listing the names, addresses, and telephone numbers of the people in a particular area⏎- means a Telephone Directory published by BT or
-
journalistic institution, Adafruit strives to be a fair, informative, and positive voice within the community – check it out here: adafruit.com/editor
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
- L5:5 suppresses digit after a label like paper or type
- L4:290 excites
- L4:4 suppresses
- L6:58 excites
- L7:195 suppresses 'of' in set noun phrases
- L6:43 excites
- L6:136 suppresses
- L4:250 excites
- L4:1300 weakly suppresses
- L1:978 weakly excites
- L3:941 weakly suppresses
- L4:315 weakly suppresses
- L6:10 weakly excites
- L7:1759 weakly excites
- L3:315 weakly excites 'of' after a noun phrase
- L3:786 weakly excites
- L3:72 weakly excites
- L6:60 weakly suppresses
- L7:56 weakly excites
- L7:1760 weakly suppresses
- L7:386 weakly suppresses
- L4:493 weakly excites
- L5:509 weakly excites the word 'even' after a clause
- L0:2535 weakly suppresses
Feeds later layers
- L10:959 excites
- L10:319 suppresses
- L10:121 suppresses
- L10:502 excites
- L9:976 suppresses
- L10:152 suppresses
- L10:67 suppresses
- L18:67 weakly excites unclear: mixed mid-sentence words
- L9:290 weakly suppresses
- L9:708 weakly excites
- L15:1101 weakly excites
- L9:61 weakly suppresses
- L14:251 weakly suppresses
- L12:99 weakly excites
- L9:742 weakly suppresses "of" after an agency or department name
- L15:795 weakly excites
- L15:638 weakly suppresses "the" inside institution names
- L11:453 weakly suppresses
- L16:689 weakly suppresses 'the' in fixed phrases like 'ahead of the curve'
- L9:172 weakly excites
- L20:146 weakly excites
- L10:29 weakly suppresses
- L15:1580 weakly suppresses
- L14:2349 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
- ately
- adamente
- weise
- /span
- eward
- /style
- atement
- emachine
- 'em
- �回
Suppresses
- t
- d
- фор
- 领域
- ,“
- ĩnh
- pé
- ích
- powerful
- 的能力