the digit 8 in lists and numbers
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. 18 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 fires on the digit 8, especially when it is a list number, a page or verse number, or part of a sequence. Turning it up pushed the text toward strange number-heavy and repetitive output.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens an email to the maker with the neuron already filled in).
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.
-
their powers of 2 do not contain zeros: 1, 2, 3, 4, 5, 6, 7, 8, 9, 13, 14, 15,
-
to fourfold greater mortality, and are up to eight times more likely to die within the first 24 hours following injury [2-4,7,8].⏎Injury to brain tissue may predispose to acute traumatic coagulopathy
-
pneumonia (nosocomial pneumonia)⏎06 - Measures to prevent surgical site infections⏎07 - Color coding of cleaning materials and equipment to prevent infections⏎08 - Measures to control performance of correct procedure at the correct body site⏎09
-
the uses of it to business?⏎problem 6: Describe the components of MIS.⏎problem 7: Differentiate between desktop PC and portable PC.⏎problem 8: Define the term enterprise application software. describe common types of it.⏎problem
-
x depth x height);⏎3. Temperature range⏎4. Temperature display accuracy⏎5. Temperature stability⏎6. Temperature uniformity⏎7. Heating time⏎8. Internal material⏎SUS304 stainless steel plate, thickness 1
-
.⏎What Causes Cold Sore?⏎The most common cold sore causes is a virus known as Herpes simplex virus type 1 or HSV-1. About 80 percent of people will get the virus before they reach the age of 1
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 was a child in a small town, and I was a boy. I was born with a color that was like a spark, that I could see, and that I could see it with my heart. I could see the truth,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
what’s available in print. This makes finding early examples of colloquial language very difficult, as most non-standard English doesn’t make it into publication. Aight was probably spoken aloud before the 90s, but if nobody recorded it
-
3 and 1994, respectively. Upon first glance, this seems far too late to me. Surely there were rappers rhyming with aight in the 1980s. Why doesn’t the OED cite
-
Italians say: Gluttony kills more than the sword. NB).⏎The company also found
-
”. The song wasn’t what you’d call a smash hit, but I mention it today because the editors of the OED have just put its namesake aight into
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
- L4:847 suppresses
- L16:55 suppresses
- L17:253 excites
- L16:113 suppresses
- L16:142 excites
- L0:2272 excites
- L17:350 suppresses
- L10:233 suppresses
- L5:204 excites
- L16:374 excites
- L17:241 excites
- L17:710 excites
- L2:1866 excites
- L17:420 excites
- L15:89 excites
- L17:765 suppresses
- L5:890 suppresses
- L16:215 suppresses
- L2:1611 suppresses
- L16:461 excites
- L16:673 excites
- L16:260 excites
- L15:114 excites
- L17:378 suppresses
Feeds later layers
- L20:1630 strongly excites
- L20:121 excites
- L20:147 suppresses
- L19:683 suppresses
- L26:1602 suppresses
- L21:979 suppresses
- L25:2442 suppresses
- L26:2635 suppresses
- L19:2501 suppresses
- L19:2017 suppresses
- L19:2905 suppresses
- L26:1341 suppresses
- L19:754 suppresses
- L20:1486 suppresses
- L22:2379 suppresses
- L20:287 excites
- L19:2787 suppresses
- L26:1054 suppresses
- L19:311 suppresses
- L20:2017 suppresses
- L20:1852 suppresses
- L26:1883 excites
- L20:950 suppresses
- L19:715 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
- eenth
- (proxy
- Braves
- ⏎ ⏎
- reeze
- ⏎ ⏎ ⏎
- Angelo
- indexPath
- puff
- OpenHelper
Suppresses
- ucha
- idity
- ince
- eline
- ariant
- tor
- Lex
- aton
- MEM
- HEN