'in' or 'on' ending a phrase
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. 13 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 small words 'in' or 'on' that come right after a verb or noun and finish a phrase, such as 'hold in', 'moisture in' or 'grid snap on'. Turning it up made the model produce strings of 'and' or 'the' with no meaning.
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.
-
⏎Whether it be driving the kids to extracurriculars, getting up at ungodly hours just to walk the dog or get a quick gym session in, or the daily grind of our careers; interdental cleaning such as f
-
you can change the shape of the curve on the spline.⏎Use the same settings as stated in the Simple Mesh tutorial (16 Grid, Grid snap on).⏎Lets start off by creating a circle with a radius of 64 (
-
in improving your health and or the way you look.⏎Calorie Intake and You⏎We’ll be looking into the idea of energy balance ie:calories in versus calories out. Today we’ll take a look out the calories you expend-energy
-
There were shoals and sand everywhere, blocking nine-tenths of the land altogether, and doing their best to block the other tenth where the great rivers run in.⏎The North Sea coast of the German Empire was the most heavily fortified stretch of
-
Namibia, a former diamond mining town whose population steadily declined after WWI, when diamond prices crashed, and Balestrino, Italy, whose inhabitants were moved in 1953 after several earthquakes brought them to the conclusion that the area
-
eter⏎- Kidney stones⏎- Suppressed immune system⏎- Immobility for a long period⏎- Holding your pee: If you tend to hold in your pee during the day or not empty your bladder completely when you do go to
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 first time I saw the article "Dish 1011112221122230111201112011120111221
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Israel⏎- *Corresponding Author:⏎- Yeongha Wook⏎Department of Pediatrics,⏎Soroka University Medical Center,⏎E-mail: [email protected]⏎Received: 01-Nov-2022, Manuscript No
-
preprocessors are given. It executes them in the following order:⏎removes instances with unknown target values⏎removes empty columns⏎imputes missing
-
2, and we know that,⏎Lines with the same slope are parallel.⏎Hence, the given lines are parallel.⏎Let's consider another set of equations.⏎y = 2x +3 and 4y = –2x+1
-
of most nouns.⏎Add -es if the noun ends in -s, -sh, -ch, or -
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
- L12:893 excites
- L10:217 excites
- L6:32 excites
- L11:3068 excites
- L9:1607 suppresses
- L9:30 suppresses
- L7:712 excites
- L7:21 excites
- L12:707 suppresses
- L12:92 suppresses
- L5:8 excites
- L6:561 excites
- L9:280 suppresses
- L4:2 suppresses
- L7:138 excites 'to' after verbs like prove or time
- L12:431 excites
- L11:300 suppresses
- L12:1468 suppresses
- L12:1112 excites
- L10:105 excites
- L7:116 excites
- L8:9 weakly excites
- L7:19 weakly suppresses
- L7:173 weakly excites
Feeds later layers
- L15:286 strongly suppresses
- L18:93 strongly suppresses
- L19:413 strongly excites
- L14:1091 strongly excites
- L14:348 suppresses
- L14:453 excites
- L19:223 suppresses
- L19:335 suppresses
- L15:1601 excites
- L17:694 excites
- L17:202 excites
- L17:146 excites period after abbreviations and dates
- L16:323 suppresses
- L17:354 suppresses
- L20:79 suppresses
- L17:375 excites
- L16:662 excites
- L20:71 suppresses
- L16:1590 suppresses
- L15:316 suppresses
- L19:182 excites
- L16:2412 suppresses
- L14:143 suppresses 'the' inside journal and institution names
- L22:160 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
- 的最后一
- anned
- emo
- esian
- -lat
- 已久的
- 的第一
- arts
- 的认可
- edral
Suppresses
- "|
- ほうが
- iscopal
- '|
- 続きを読む
- getParameter
- '%(
- respect
- "%(
- indices