words that introduce a clause
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. 10 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 lights up on the small word right before a new statement begins, most often "that" after verbs like "tell", "show" or "advise", as in "shows that" or "Tell students that". Turning it up didn't produce a clear theme in the sample outputs, so its job is only partly understood.
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.
-
came up with an innovative robot design. They added a tail to their robot "Tailbot," which is essentially a toy car with a tail attached. It turns out that robot design could become much more than just a research tool.⏎Team leader Robert
-
your employees⏎Do a small search online, and you’ll soon find out that there are many different technical skills out there. But which ones are the most important
-
watered with rain and tap water, because they have more minerals for growth.⏎First, I had to decide what plant to test. After researching this, I decided to do bean plants, due to the quick speeds that they grow at. My
-
provides students with the opportunity to work with complex shapes to find surface area in a given context.⏎Display the image for all to see throughout the activity. Tell students that they calculated the volume of this heart-shaped box in a previous lesson and today they
-
describe graphs.⏎Can you find examples on the cards that match each definition? Can you draw some more examples of your own?<|endoftext|>Since 2012, the global economy has been experiencing average growth rates of less than 3 percent.
-
enough to allow entry for a fully-equipped firefighter.⏎- For determining egress window sizes and placement, the International Building Code advises that every bedroom must contain at least one egress window. It must be at least
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 role of a leader in the company. The company and the organization that is being led by the person in question. What about? This is a problem. The company is led by the person in question. The problem of the company.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
through all of it.⏎So what’s the key? Goals have to be actionable. That is the only way to make sure they will know what steps to take, and don’t waste valuable
-
without the benefit of God’s grace and mercy. Because of his own sin, Saul found himself in a hopeless predicament, and killed himself. Only in the sense that God’s word was the immovable object against which Saul rebelled could it
-
Most of us recognize this from our own lives. We cannot become who we are without caring for the relations that care for us.⏎David Brooks called this the meaning of covenant, the bond between singular persons that protects not private interests (as contracts do
-
substance, so that one is consumed in the process. Anyone who wants to comprehend Christ’s message of light … must comprehend this one thing: the mission, the duty to shine, to draw others, to see, to heal, to do good
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
- L8:2452 excites
- L8:894 suppresses
- L5:2901 weakly suppresses
- L7:1834 weakly excites
- L6:2089 weakly excites
- L8:172 weakly excites
- L8:1260 weakly excites
- L8:1608 weakly suppresses
- L6:686 weakly excites
- L9:265 weakly excites
- L6:2301 weakly suppresses
- L4:2929 weakly excites
- L2:1593 weakly suppresses
- L1:731 weakly suppresses
- L6:759 weakly excites
- L7:2568 weakly excites
- L9:1874 weakly suppresses
- L7:2205 weakly suppresses
- L6:1700 weakly excites
- L7:3009 weakly suppresses
- L9:1353 weakly excites
- L6:666 weakly excites
- L6:1123 weakly suppresses
- L3:972 weakly excites
Feeds later layers
- L18:2701 excites
- L11:1546 weakly suppresses
- L12:1420 weakly excites digits starting a percentage or rate
- L18:1960 weakly suppresses
- L21:656 weakly excites
- L12:1141 weakly suppresses
- L20:220 weakly suppresses
- L23:2027 weakly excites
- L24:89 weakly excites
- L11:2 weakly suppresses unclear: blank before a year or number
- L13:2766 weakly suppresses
- L14:1594 weakly excites
- L17:295 weakly suppresses
- L19:2564 weakly excites "be" after a modal verb
- L12:835 weakly excites
- L12:1324 weakly excites
- L15:2523 weakly excites
- L11:1072 weakly excites
- L17:2152 weakly excites
- L21:2091 weakly excites
- L16:2728 weakly excites
- L14:1738 weakly suppresses
- L15:1652 weakly suppresses
- L13:1912 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
- 的气息
- gyr
- 容忍
- Jaune
- $ar
- にする
- AREST
- 蟠
- /rs
- 很开心
Suppresses
- ension
- ensity
- W
- ie
- w
- �
- 引
- lli
- mega
- uality