'a' before a noun phrase
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 "a" when it comes between a verb or preposition and the noun that follows, like "has a long history" or "a good idea". Turned up, the model produced vague, repetitive phrases with "a lot" and "a kind".
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.
-
which you will read and write about for Exercise 1), Havelock launches his main argument about the shift from oral culture to literate culture, and is a good way for us to start wrapping up the oral unit.⏎Read the full article
-
do this by using a moisture meter. If your succulent is indoors, it’s a good idea to get a small spout watering can. It’s also a good idea to avoid using high pressure streams of water as this can damage the roots
-
beautiful discourse of the Holy Quran and Sunnah, Muslims come to learn about the various rights that Allah (SWT) has enjoined upon His creation – in a practical and detailed manner. Without the fulfillment of these basic rights, the whole institution
-
, and has evolved over time to reflect changing artistic styles and techniques. In this blog post, we'll explore the fascinating world of Japanese art.⏎Japanese art has a long and rich history, dating back to prehistoric times. Early examples of Japanese
-
& Conditions here⏎This set is designed to get students thinking and talking about math. With the 120 color-coded task cards, you will always have a problem at your fingertips. Use the cards on the carpet or in small groups—even
-
's a "citation?"⏎When you write a research paper, at the end of the paper you include a list of the resources you used to write the paper on a page titled "Works Cited" or "References." Each entry in that list
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: the same of a one of the the other, but the one of the one of the one of the one... one of the. The one of the, but the one of the, but the one of the... one of the,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
A growing body of evidence suggests that increasing poor people’s access to better financial tools can help
-
one place, and every day was exactly the same, and nothing that you did mattered?" Phil asks two local guys at the<|endoftext|>Courtesy of EarthSky⏎A Clear Voice for Science⏎Here is the view northward on April evenings. Now the
-
need time. In fact, Rahner thought of an entire century! Therefore, it is our task to deal with the globalization of society and the Church.⏎A Globalized World and a Universal Church⏎In 2013, the Roman
-
a 2013 study, Peter Strelan of the University of Adelaide looked at<|endoftext|>November 19, 1998⏎A weekly feature provided by scientists at the Hawaiian Volcano Observatory.⏎Hawaiian honeycre
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
- L16:880 strongly suppresses
- L16:2248 strongly excites
- L4:137 strongly excites
- L9:59 strongly suppresses "a" meaning per (a day, a year)
- L11:331 strongly suppresses
- L16:1309 suppresses
- L4:49 excites
- L11:2859 suppresses
- L11:875 excites
- L16:2062 suppresses "other" and contrast words
- L0:173 excites
- L1:1062 suppresses
- L10:365 suppresses
- L16:2611 excites
- L3:2559 suppresses
- L3:1810 excites
- L15:1848 excites
- L3:597 suppresses
- L13:1234 excites
- L5:225 suppresses
- L10:1162 suppresses
- L6:640 excites "the" introducing a noun phrase after a name or preposition
- L16:1949 suppresses
- L16:342 excites
Feeds later layers
- L20:901 strongly excites
- L22:1881 strongly excites
- L18:635 strongly excites
- L21:592 strongly excites
- L19:2088 excites
- L18:1084 excites
- L23:218 excites
- L20:835 excites
- L18:216 suppresses
- L19:2115 suppresses
- L22:260 excites
- L27:387 excites 'a' right after 'played' or 'play'
- L22:211 excites
- L22:2381 excites "a" before "bit" or "little"
- L18:314 suppresses
- L27:2632 excites
- L25:1586 suppresses
- L25:390 excites
- L18:2389 excites
- L25:320 excites
- L18:2608 excites
- L19:751 excites the article 'a' before a noun
- L18:602 excites
- L24:489 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
- lot
- few
- very
- certain
- solution
- variety
- combination
- ikip
- number
- wide
Suppresses
- 的一种
- slashes
- 这三个
- 的一个
- 的一款
- languages
- lard
- anches
- 有一天
- plings