unclear: word pieces around turbines and wind shear
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. 3 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 a mix of unrelated word pieces, most often 'turbines' in talk about engines and 'shear' in weather reports. It also dips on 'was' or 'is' in 'where was' questions. Turning it up made the model's writing wander without a clear theme.
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.
-
cables, as in vertical-lift drawbridges, power shovels, and cranes. Before the extensive use of electric motors, steam engines or water turbines often supplied the power for factory machinery. One engine or turbine might run a whole
-
September 13 local wind shear was having its effect on the storm, and Hurricane Humberto’s winds had begun
-
4 hours. They predicted that very strong wind shear, dry air and a track over marginally warm sea surface temperatures could result in
-
One might even say that Americans are a hero-worshiping people.⏎How can the American people love both equality and inequality simultaneously? How
-
mud wallows or taking a nap in a big, fluffy hay bed. No, those aren’t warthogs you are looking at—these are one of the most critically endangered pigs in
-
a lot of clay. I always dug the hole at least 25 percent deeper than the root ball and refilled that portion with watered-down good soil before placing the tree in the hole. The rest of the hole would
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 big company, the way to get a good life, a life for a resident. Let me try to the weather. Okay, that's a life for a resident. Let me try to the weather. So, this is the way
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
was this wreck, where was it located, why was it important? When did it happen?⏎2. What
-
expelled the only red state in the national organizaton" -- but<|endoftext|>Raising Our Children to Be Gatekeepers⏎And the LORD said unto Cain, Where is Abel thy brother? And he said, I know not: Am I my brother
-
2. Every day the boss enters the office⏎at half past nine⏎3. My friend will ring me up at eight o’clock⏎4. Where was Kate going when you met her yesterday?⏎does your brother work?⏎6. What
-
or deep? Explain the type of plate boundary it was near. Why are earthquakes common to this area?⏎Where was the deepest earthquake during the week? On what type of boundary was this?⏎Where
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
- L1:1159 weakly excites
- L1:2536 weakly suppresses
- L2:1657 weakly excites
- L1:2654 weakly excites
- L2:229 weakly excites
- L0:710 weakly excites
- L2:115 weakly excites
- L2:75 weakly suppresses
- L0:433 weakly excites
- L0:260 weakly suppresses
- L2:2480 weakly suppresses
- L0:1036 weakly excites
- L1:535 weakly suppresses
- L1:643 weakly suppresses
- L2:1584 weakly excites
- L1:2683 weakly excites
- L1:2723 weakly excites mid-word syllables in capitalised names
- L0:1398 weakly suppresses
- L1:2340 weakly suppresses
- L0:1617 weakly suppresses
- L0:1568 weakly suppresses
- L2:2471 weakly suppresses
- L0:2019 weakly suppresses
- L0:1989 weakly excites
Feeds later layers
- L9:2421 weakly excites
- L8:505 weakly excites
- L15:2175 weakly suppresses
- L19:1418 weakly suppresses
- L4:1594 weakly suppresses
- L18:1510 weakly excites
- L26:168 weakly suppresses
- L12:1491 weakly excites
- L13:2407 weakly excites
- L22:1641 weakly suppresses
- L13:1603 weakly excites
- L8:2436 weakly suppresses
- L26:1050 weakly suppresses
- L7:939 weakly suppresses
- L25:1079 weakly suppresses
- L14:367 weakly excites
- L6:2911 weakly suppresses
- L17:1759 weakly excites
- L10:104 weakly excites
- L8:2285 weakly suppresses
- L4:844 weakly excites
- L18:1734 weakly excites
- L8:2041 weakly suppresses
- L23:3008 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
- per
- hard
- pers
- еб
- 短线
- conde
- ере
- SSION
- use
- итель
Suppresses
- anthology
- onnement
- scratched
- 百花
- _Collections
- deductible
- TAIL
- .hu
- ozilla
- *T