comma after 'in this article'
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. 16 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 comma that follows a phrase like 'In this article' or 'In this essay', where the writer is announcing what the text will cover. When it was turned up, the model's replies drifted into rambling, unrelated topics, so the steering results don't add much.
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.
-
know, there are no modules that can identify offshore areas that do not touch other neighbor areas in the same layer.⏎In this article, I’ll just discuss spatial dissolving to keep it simple and focused. Figure
-
number of academic journals is because it helps them choose a journal to submit their articles.⏎In this article, I will try to answer and discuss this question as comprehensively as possible. We
-
effort and time, while smart work is about finding the most effective and efficient way to get things done. In this essay, we will discuss the differences between hard work and smart work and analyze the benefits of
-
sequence classification, and web content classification.⏎Reinforcement machine learning⏎A reinforcement learning algorithm, or agent, learns by interacting with its environment. In reinforcement learning, there is no answer in the data but the reinforcement agent decides what to do to
-
exploited⏎What is a narrative therapy?⏎Narrative therapy is a style of therapy that helps people become—and embrace being—an expert in their own lives In narrative therapy, there is an emphasis on the stories we develop and carry with us through our lives
-
look at some of the specific foods that are considered clean and maybe some simple recipes using foods with which we’re already familiar.⏎Tomorrow<|endoftext|>In this Artificial intelligence Essay, we will learn about Artificial intelligence, Types of Artificial Intelligence & Applications of Artificial Intelligence
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 if: "Let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—let—
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
tray' and covering it with paper.⏎My cousin was 24 years old on Friday April 5th in 1974. On what day⏎of the week was she born?⏎A tetromino is made up of
-
. The historic school house, built in 1872, still stands.⏎A historical marker at Cherry Creek provides additional information:⏎"Here, at one time, was the largest town in White Pine County. Part of the Cherry Creek Mining
-
framework you use to set goals, taking the time to plan what you want to achieve can make the difference to obtain what you want. Sort of if you know where you are going, you’ll get there. Otherwise, you may end up in
-
are, how you move, it can measure sound, light intensity and so on. Combine this with other data such as your email and social media activity and over time we can build a picture to help identify potential triggers in
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
- L15:853 strongly suppresses
- L10:931 excites
- L11:949 excites
- L4:1750 suppresses
- L9:361 excites
- L10:289 excites
- L6:502 suppresses
- L3:681 suppresses
- L9:2442 suppresses
- L6:2233 excites
- L4:1179 excites
- L12:962 suppresses
- L6:2843 suppresses
- L1:2751 suppresses
- L5:976 suppresses
- L4:1834 suppresses unclear: time spans and 'once you've'
- L5:224 excites
- L6:1829 suppresses
- L15:3010 excites
- L0:2135 suppresses
- L6:327 weakly suppresses
- L15:2521 weakly excites
- L11:1098 weakly excites
- L10:371 weakly suppresses
Feeds later layers
- L17:2162 excites
- L17:1845 excites
- L17:3069 excites
- L17:1201 excites
- L17:2238 excites
- L17:2213 excites
- L18:509 excites
- L17:472 suppresses
- L22:689 excites comma closing an intro clause
- L17:2524 suppresses
- L22:2294 excites
- L19:1982 excites
- L18:867 excites
- L19:2397 excites
- L19:2103 excites
- L17:2659 excites
- L19:1953 excites
- L17:2766 excites
- L17:2816 excites
- L18:1811 suppresses 'and' linking two descriptive claims
- L22:1124 excites
- L18:1246 weakly suppresses
- L20:242 weakly suppresses
- L17:1662 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
- 一切都
- %"><
- família
- heads
- 你不
- 一切都是
- ausal
- >Main
- matter
- 我们将
Suppresses
- 者
- 提醒
- eming
- 在他
- тор
- tics
- 欢迎
- り
- 让她
- 支持