unclear: document ends and instruction verbs
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. 14 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 is hard to read. It fires most on the end-of-document marker between web pages, but also on some other spots. It is strongly negative on action verbs in how-to instructions, like 'take' or 'plant'. Steering did not give a clear picture of what it does.
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.
-
’ or ‘raw’ and ‘war’. Use the update function to automatically scan for overlaps, the numbers should update every time you click outside the word list.<|endoftext|>ambience: dingoes howling⏎Australian Aborigines first named these wild
-
. FormsCan be used for anything<|endoftext|>We found 1,708 reviewed resources for state of the union address⏎How can
-
ates and the data subjected to analysis of variance and the means were compared by Tukey test.⏎News published in Agência FAPESP Newsletter about the scholarship:<|endoftext|>Image: Blazing volcano⏎"This is an erupting volcano, showing the process
-
instructions and pre-fetching. Relevant Computer Science topics such as models of computation and formal grammars are addressed, and their practical value explained.⏎Share Low-Level Programming<|endoftext|>New Delhi: For possibly the first
-
utf8mb4) bytes per character can be used. So if one specifies VARCHAR(50) CHARSET utf8mb4, …[Read more]<|endoftext|>To actually reduce pounds in a healthier manner, it is perfect to burn fat gradually
-
also become unstable. In most cases, the treated tooth is then crowned. How exactly to proceed further, the doctor and patient discuss in the root canal aftercare<|endoftext|>Black locust is a member of the legume
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: A question that is a problem, a question that can be solved, the problem can be solved, a problem can be solved... a problem. The problem, as the problem... a problem. The problem can be solved. The problem,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
considerations for when and how to take the medication include:⏎- Monopril comes in tablet form. It is taken once a day.⏎- It is best to take it on an empty stomach at least one hour before a meal. However, it
-
provide the community with great swimming, boating and fishing opportunities.⏎The Garvey-Glenn drains 17.5 square kilometres of farmland and empties directly into Lake Huron,
-
succulents also need nutrients for their growth and development.⏎How to propagate succulents:⏎Cut a single leaf or stem from the succulent very gently and plant this cutting separately.⏎Where to place succulents?⏎Succul
-
needs a space in Linux you can use a backward slash followed by a space, to escape the space in a name. If this isn't done, Linux will treat anything after the space as an option or additional command.⏎Finally, escape characters are
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
- L12:2347 suppresses
- L13:2636 suppresses
- L12:1022 excites
- L9:1444 excites
- L13:1456 excites
- L6:487 suppresses
- L6:2326 suppresses
- L6:2692 excites
- L12:1653 suppresses
- L10:1395 excites
- L10:132 suppresses
- L12:525 suppresses
- L13:1928 suppresses
- L13:1906 suppresses
- L6:1829 suppresses
- L13:2299 suppresses
- L13:792 excites
- L12:1066 suppresses
- L12:2680 suppresses
- L11:487 excites
- L6:1001 suppresses
- L4:118 weakly suppresses
- L13:645 weakly suppresses
- L5:1535 weakly excites
Feeds later layers
- L16:1934 strongly suppresses
- L15:1028 suppresses
- L15:1702 suppresses
- L18:2830 suppresses
- L16:1077 suppresses
- L15:1047 suppresses
- L17:429 excites
- L18:2243 suppresses
- L16:1551 suppresses
- L15:2633 suppresses
- L16:2180 suppresses
- L18:2842 suppresses
- L16:2157 suppresses
- L19:2228 suppresses
- L18:2971 suppresses
- L18:1299 suppresses
- L19:398 suppresses
- L16:2612 suppresses
- L18:867 suppresses
- L18:358 excites
- L16:1313 suppresses
- L15:27 suppresses
- L17:663 suppresses
- L15:37 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
- -state
- conut
- -Identifier
- ilitary
- .getLog
- national
- ulatory
- uel
- .lin
- abetic
Suppresses
- disposed
- 大片
- 它们
- 这两
- 这支
- гран
- 这只
- 这种方法
- esto
- 這些