prepositions opening a phrase ('after', 'on')
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. 21 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 short connecting words like 'on', 'after' and 'of' that lead into a phrase, such as 'Based on' or 'and each day after that'. When it was turned up, the model kept producing the word 'this' over and over.
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.
-
or does not have an aneuploidy. Recent recommendations from the American College of Obstetricians and Gynecologists (ACOG) are quite clear on that issue.⏎In those same recommendations, ACOG also states
-
.⏎- Have them develop an explicit formula for each sequence.⏎2. A company is offering to pay you one penny on your first day at work and each day after that your salary would triple. Assume you work five days a week.⏎a.
-
movements and synthesize the speed, so that the movement required for tooth surface cutting is transformed from the rotary movement to the linear movement as far as possible. Based on the analysis of the tooth surface cutting movement, the revolution and rotation of the blade
-
you would need to⏎pick out 15 sweets because the first 9 could be mice and the next 5⏎you pick out could be frogs and after that you're definitely going⏎to pick out a worm. As well as that
-
But, a bill in Congress would do just that. And if it passes, not just farmers and ranchers would be affected
-
operate kindergarten-through–12th grade (K–12) schools across the United States is creating safe and secure environments that support effective teaching and learning. To help them meet this obligation, the U.S
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 way in a **pre-sim** (pre-**sim************) **** **** **** **** **** **** **** **** **** **** **** **** **** **
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and our research aims to use statistical techniques and other methods such as artificial intelligence to interpret large data sets.⏎“While we have this mathematical and statistical knowledge, what’s missing is the framework that can capture all of this complex
-
This pressure causes them to forget even the things they know. Students who face math anxiety can use find the limit calculator for help.⏎- Poor Foundation⏎Foundation skills are very important when it comes to learning and solving hardcore mathematical problems. If a student
-
attention to the connectivity of the components in a circuit diagram. Arrows, lines and other symbols should be used to show how different components are connected. This information is crucial for figuring out the operation of the circuit.⏎By following the above circuit diagram
-
. It is possible to send and receive bitcoins using a Bitcoin wallet.⏎Even when the value of an asset drops, it’s possible to make a profit, which is counterintuitive
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
- L18:2923 excites
- L19:1450 excites
- L19:1204 weakly suppresses
- L18:769 weakly excites
- L19:2138 weakly excites "of" in "matter of" phrases
- L17:1374 weakly excites
- L9:1995 weakly suppresses
- L15:3019 weakly suppresses
- L19:2052 weakly excites
- L20:593 weakly excites
- L18:2153 weakly suppresses
- L19:1759 weakly excites
- L3:2858 weakly suppresses
- L19:2558 weakly excites
- L19:1181 weakly suppresses
- L19:858 weakly suppresses
- L13:2159 weakly suppresses
- L11:163 weakly suppresses
- L16:1059 weakly excites
- L19:786 weakly excites
- L16:89 weakly excites
- L3:2629 weakly excites
- L6:828 weakly excites
- L19:1914 weakly suppresses
Feeds later layers
- L26:405 strongly excites
- L27:528 strongly excites
- L22:1469 strongly excites
- L24:803 excites
- L26:586 excites
- L26:1191 suppresses
- L26:64 weakly suppresses
- L23:1068 weakly suppresses
- L23:2734 weakly excites
- L24:306 weakly excites
- L24:859 weakly suppresses
- L26:1052 weakly suppresses
- L26:2261 weakly excites
- L27:1055 weakly excites
- L27:616 weakly excites
- L22:735 weakly excites
- L24:343 weakly excites
- L25:185 weakly excites comma after a city in an address
- L24:1975 weakly excites
- L25:2653 weakly excites
- L22:2794 weakly suppresses
- L23:334 weakly excites
- L22:1031 weakly excites
- L24:2045 weakly suppresses acronym introductions
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
- this
- 此
- this
- This
- .this
- [this
- 这个
- =this
- This
- (this
Suppresses
- assin
- Serviço
- üş
- 说自己
- 一所
- 的画面
- reeNode
- artz
- 的日
- ******/⏎⏎