unclear: noun heads in expository text
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. 12 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 the main noun of a phrase in informative writing, such as 'holes' in 'black holes' or 'chains' in 'food chains'. Its meaning is hard to pin down, and turning it up made the model's text drift into random lists and counting rather than a clear topic.
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.
-
supermassive black hole in the new galaxy develops an insatiable appetite. However, this ferocious appetite is unsustainable.⏎The growth of supermassive black holes, which are found in the centers of all normal galaxies (including our Milky Way
-
er will harpoon itself into place to keep it from ricocheting back off into space.⏎It may hold secrets on the origins of life⏎Comets are balls of ice and dirt left over after the dawn of the solar system, and some
-
number using prime factorization.⏎Algebra.com [offsite link]⏎Algebra.com has a variety of sample problems plus online tutors and more.⏎Converting Decimals -
-
slide. There is also a child-sized theater with a stage curtain for dramatic presentations. Science content will include life cycles, invasive species that threaten our lakes, food chains, water quality, and identification of fish species.⏎Mounts of game birds and
-
educate our patients on the signs and symptoms of gum disease and how to treat, and ultimately prevent it.⏎Did you know that most people do not even know that they have periodontal disease? This is because a lot of the time, pain
-
- Adding Mixed Numbers⏎Learn how to Add mixed numbers with like and unlike denominators.⏎Fractions - Adding Unlike Denominators - LCM⏎Least common multiples are very useful in finding the least common denominator when adding and subtracting fractions.
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: I want to add to the problem. I want to add to the problem. So that's what I want to add to the the problem. But I want to add to the the problem. So that's the problem. But I want
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
if you fall into one of the high risk categories listed below. This will keep your doctor on the
-
gone from technology being a shiny object–a TV–to being an indispensable part of one’s everyday life. Technology is changing as rapidly as you change the channels on the TV set. The challenge and opportunity in education is to understand what’s needed
-
of a tile.⏎So what can I do now?⏎Once the glaze on tiles has worn, it can't be rejuvenated in any way.⏎Likewise,
-
3 seconds I had to have a tripod and a cable release, also with an exposure of 33 seconds I would need to use the bulb mode as my cameras set exposures stops at 30 seconds.⏎Second get the exposure right; this
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
- L11:2210 suppresses
- L9:374 excites
- L7:651 excites
- L11:937 excites
- L7:34 suppresses
- L9:55 excites
- L11:119 weakly suppresses
- L5:705 weakly suppresses
- L11:860 weakly suppresses
- L11:172 weakly excites
- L11:1294 weakly suppresses
- L10:661 weakly suppresses
- L10:624 weakly excites
- L8:275 weakly suppresses
- L11:114 weakly excites
- L11:1247 weakly suppresses
- L10:2421 weakly excites
- L11:223 weakly suppresses
- L7:327 weakly excites
- L10:967 weakly excites
- L9:1726 weakly suppresses
- L10:1644 weakly suppresses
- L7:103 weakly suppresses
- L9:1553 weakly suppresses
Feeds later layers
- L17:1706 strongly excites
- L17:1974 suppresses
- L17:1848 excites
- L18:1238 excites
- L16:2507 excites
- L20:807 excites
- L13:1647 excites
- L13:2063 suppresses
- L19:654 suppresses
- L21:913 excites
- L19:626 excites
- L19:592 suppresses
- L25:990 excites
- L24:488 excites
- L16:1317 excites
- L25:374 excites
- L15:1920 excites
- L20:1857 excites
- L21:392 excites
- L17:340 excites
- L27:1052 excites
- L13:1912 excites
- L20:240 suppresses
- L25:423 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
- adoras
- ших
- URN
- tures
- _dicts
- semblies
- arsers
- iefs
- ящих
- шие
Suppresses
- a
- 一份
- 一个
- 活动
- 一颗
- a
- 恭喜
- 把我
- 因为你
- 因为我