"in" before a time span
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. 23 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 word "in" when it introduces how long something takes, like "in a few seconds" or "in under twenty-four hours". When it was turned down, the model got stuck repeating "in", and when turned up it wrote long strings of numbers or code-like text.
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.
-
the maneuvering. All Carpenter did was destroy the group’s paralysis with a vision of where to aim first. They figured out on their own how to get there in the same way they landed the jet five years before.⏎So, there are lessons
-
one of the popular development methodologies that allows development teams to create a working product with high-quality code in a predictable time frame. Below we will describe Agile cycles and also indicate the types
-
shells mark the scene of a few hours illegal fishing, where a whole population of this globally threatened species can be wiped out in a couple of hours.⏎These threats to the species have meant that the freshwater pearl
-
machine Colossus at Bletchley Park, thanks to the efforts of British Intelligence, meant the messages scrambled by the enigma machine could eventually be decoded in under twenty-four hours, leading the allies to an early victory.⏎Laurence Fisher
-
, and traditional educational pathways need to be rethought and redeveloped.⏎Some experts have criticized university curricula for lacking agility. Many programs haven't been updated in years even though the technology refresh cycle is happening at a faster pace. Continuous or
-
requires the user to have a compass and a general idea of the satellite’s location. Slowly turning the terminal will soon indicate signal capture, which can be done in less than a minute for an experienced user with a good signal.⏎Some limitations include
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: in the in the in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ways. Arranging eight different tiems in a line can be done in 8×7×6×5×4×3×2×
-
the history of our planet has EVER created!⏎We can deduce this because counting such arrangements is very simple. Arranging three different items in a line can be done in 3×2×1
-
, how long does it take to learn Hebrew fluently? A conservative estimate might be something like 2 years, though a really motivated and talented student might get there as soon as, say, 18 months.⏎Can I learn Hebrew for
-
learn and Implement⏎PHP is easier to learn than its competitors such as like ASP, CGI, and others. It is so easy, that a beginner can learn php write a program within few
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
- L21:2347 excites
- L22:2338 suppresses
- L19:1917 excites
- L21:1027 excites
- L22:825 suppresses 'for' before a span of time
- L22:2932 excites
- L20:696 excites
- L20:1335 weakly suppresses
- L20:1404 weakly excites
- L22:439 weakly excites
- L21:175 weakly suppresses
- L19:2887 weakly suppresses
- L21:3026 weakly excites
- L19:1764 weakly excites
- L18:2846 weakly suppresses
- L12:734 weakly excites
- L17:2760 weakly excites
- L19:1118 weakly excites
- L20:2041 weakly excites
- L8:1482 weakly excites
- L18:1750 weakly suppresses the blank slot before a year
- L2:108 weakly excites
- L18:291 weakly suppresses
- L11:1604 weakly excites
Feeds later layers
- L24:893 strongly suppresses
- L26:207 strongly excites
- L27:405 strongly suppresses
- L26:825 suppresses
- L25:2608 suppresses
- L27:2429 suppresses
- L25:2282 excites
- L25:1790 suppresses the digit 1 starting a statistic
- L25:2902 excites
- L24:116 weakly excites
- L24:1503 weakly excites
- L26:1401 weakly suppresses
- L25:1369 weakly excites
- L26:584 weakly suppresses
- L27:918 weakly suppresses
- L25:2440 weakly suppresses
- L26:667 weakly excites
- L27:973 weakly suppresses
- L25:421 weakly excites
- L25:2569 weakly excites
- L24:587 weakly excites
- L24:2028 weakly excites
- L24:2950 weakly excites
- L25:2756 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
- ervative
- ato
- igators
- hec
- asons
- rogate
- toHaveLength
- lace
- ató
- requently
Suppresses
- in
- In
- In
- 在
- >In
- _in
- dalam
- 'in
- →in
- -in