'of' after a 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. 17 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 'of' when it follows a noun and is about to name what that noun belongs to, as in 'the size of', 'the length of' or 'the native language of'. When turned up or down, the model drifted into repeated phrases and percentages, which does not clearly match the trigger.
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.
-
transistor radio, because it’s starting to look a lot like the 1950s.⏎Unfortunately, this won’t be the nostalgic ’50s of big cars and pop music.⏎The 1950s that could be
-
consider the following aspects.⏎- Spelling – Convention⏎The spelling of Chinese words using the Latin alphabet (romanisation) seems to be more art than
-
the size of a similar-size hoofed mammal and eyes only a third as large.⏎The
-
technical documentation. However, it is often not the native language of the readers (or even of the authors) of such documentation. Many readers have
-
monopoly on the newly discovered region. He never waiver in his belief that he had found an alternate route to Asia. Columbus had found a paradise just not the one of his hopes and aspirations. This paradise was right before his eyes, yet he lost
-
of the most technically demanding ever written for the piano, he is known more for his nuance, his expressive depth and his ability to conjure up the melody of the human voice from the instrument’s keys. For this sublime blend of skill and
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: what is what, which is itself, to what it do. In this, I am both of what I am and in what I am. In what I am, are there be be be be be be be be be be be be
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
reports that traffic engineers could use to adjust traffic lights or improve roads.⏎Bill Gates and Paul Allen were high school students at Lakeside School in Seattle. The Lakeside Programmers Group got free
-
From Tree Roots⏎- Locate sewer lines – You can call your local public works office or the national 811 “Call Before You Dig” number to locate underground utilities on your property. Before
-
perfect companion for study and revision for nursing and healthcare students from the publishers of the market-leading at a Glance series. This easy-to-read, accessible guide brings together all the key principles of caring for patients with medical and
-
and offers meditative benefits.”⏎This demand fits with the overall interest he’s seeing among clients for less intensive fitness routines that are accessible and mesh with an on-the-go lifestyle.⏎Why gardening is exercise⏎Gardening engages all the major muscle groups
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
- L15:888 strongly excites
- L16:1594 strongly excites
- L16:652 strongly suppresses
- L14:935 excites
- L9:2661 excites
- L11:2859 suppresses
- L10:321 suppresses
- L10:365 suppresses
- L5:493 excites
- L15:1888 excites
- L10:1644 excites
- L10:556 suppresses
- L9:227 excites
- L15:2253 excites
- L4:392 suppresses
- L1:740 suppresses
- L9:662 suppresses
- L9:74 suppresses
- L10:120 suppresses
- L16:513 suppresses
- L2:1439 excites
- L10:1289 excites
- L10:1278 excites
- L10:1377 suppresses
Feeds later layers
- L18:1782 strongly suppresses
- L19:1023 suppresses
- L20:438 suppresses
- L19:644 suppresses
- L19:1180 excites
- L19:1539 excites
- L18:1041 excites
- L21:713 excites
- L20:596 suppresses
- L21:614 suppresses
- L25:284 excites
- L20:1883 suppresses 'the' inside a mid-sentence phrase
- L23:79 excites
- L24:153 excites
- L25:544 excites
- L18:41 excites
- L20:219 suppresses "the" after time or place prepositions
- L18:376 excites
- L19:2115 excites
- L19:2815 suppresses
- L21:675 excites
- L20:47 suppresses
- L20:738 excites
- L20:809 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
- cles
- clus
- hung
- {text
- C
- Pascal
- ikers
- 的任务
- the
- 的看法
Suppresses
- 外围
- propio
- ligt
- เสม
- educação
- liest
- ดาร
- 绪
- -largest
- (primary