unclear: scattered mid-phrase words
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 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 grab bag of words that sit in the middle of a phrase, with no clear shared theme. When it was turned up, the model's text wandered into unrelated topics, so there is no clear story about 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.
-
growing love handles despite diet and exercise. Also deteriorating athletic performance, a general sense of fatigue or falling asleep right after dinner can all be signs of low testosterone, but they are often written
-
as being a List of Users.⏎Sometimes, it’s also handy to define a list of multiple Entities and Structures.⏎Until 184.108.40.2
-
68k-palmos-coff-gcc -v should report its version number to be (at time of writing) 22.214.171.124-kgpd-07109
-
and would like to contribute something to the GNU project, please consider volunteering for one of these projects. If you are seriously contemplating work, please write to `firstname.lastname@example.org' to coordinate with other
-
unkel⏎starring Jim Witter,⏎featuring Ian Tanner⏎“Time it was, and what a time it was”...⏎These opening lyrics to Simon
-
1931, John and Jack, two of the caretakers of the Imaginarium Geographica, discover a plea for help on a medieval manuscript - which seems to have been written by their friend, Hugo Dyson! When
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 person who is not in the right, the right, and the right is not in the right, and the right is not, and the right is not, and the right is not, and the right is not, and the right
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
even after advice⏎- Be obsessed with body image and shape. They often fail to recognize the ill-effects of excess weight loss.⏎The most telling signs about anorexia can be picked around the dining table as they try
-
.⏎Edwin Lacey was a relative of the Revolutionary War general and the son of Isaac and Ruth (nee Twining). Lacey was a tall, gaunt, red-headed bachelor, a Quaker, and a farmer. He was an
-
baby warm.⏎An old custom of washing a newborn soon after birth is being discounted with the knowledge that infants were becoming sick. Instead, a baby should be gently dried with a towel and placed next to the mother's skin. Washing should wait until
-
for a description). So you'll have to call this macro 5 times to get the 5 characters.⏎Remember that when you print a string it should be null
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
- L5:1812 weakly excites
- L9:1766 weakly excites
- L2:444 weakly excites
- L7:152 weakly excites
- L9:2509 weakly suppresses
- L13:1124 weakly excites
- L5:2501 weakly suppresses
- L11:2290 weakly excites
- L12:2254 weakly excites
- L13:1549 weakly suppresses
- L4:273 weakly suppresses
- L10:2144 weakly excites
- L12:3013 weakly excites
- L4:1066 weakly excites
- L5:2700 weakly suppresses
- L8:2501 weakly suppresses 'iv' inside herbivores, energy fuel words
- L9:932 weakly excites
- L10:1837 weakly excites
- L1:2672 weakly suppresses
- L6:799 weakly excites
- L3:2791 weakly suppresses
- L11:2306 weakly excites
- L2:1950 weakly excites
- L12:2208 weakly excites
Feeds later layers
- L18:1509 weakly excites
- L22:2685 weakly excites
- L22:2116 weakly excites
- L21:2047 weakly excites
- L21:2013 weakly excites
- L18:2617 weakly excites
- L27:1770 weakly excites
- L20:2829 weakly excites
- L25:303 weakly excites
- L21:1545 weakly excites
- L22:2344 weakly excites number after 'account for' or 'over'
- L22:2247 weakly excites
- L21:2946 weakly excites
- L20:307 weakly suppresses
- L22:1625 weakly excites
- L23:563 weakly excites
- L18:1530 weakly suppresses
- L16:900 weakly excites
- L24:2891 weakly excites
- L20:1299 weakly excites
- L22:2740 weakly suppresses
- L19:1902 weakly excites
- L15:2423 weakly excites
- L19:790 weakly 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
- 掌门
- Switch
- umer
- clusion
- recess
- 帝王
- universities
- millennium
- discoveries
- aro
Suppresses
- uated
- iffin
- ava
- rated
- Pvt
- 醪
- .Views
- veled
- laughter
- haft