'so' in 'and so on'
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 word 'so' in phrases like 'and so on' or 'and so were'. Pushing it up made the model repeat the word 'then' over and over, because the neuron tends to push that word forward. Its negative side is on sentence starts like 'Now' and 'Then' that begin a new step.
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.
-
chemical reaction and so can be reused.⏎- Enzymes are proteins and so are sensitive to temperature
-
Cold War. That was before scientists knew about the oceans on moons around other planets, before they knew about how tough microorganisms get here on Earth (and so maybe in space too?), before they started planning experiments to look for life on Mars
-
in terms of safety, reliability, performance, profitability, and credibility for the affected organization and its representatives.⏎There are standards for quality, safety, reliability, and so on, in industries related to food, drugs, military equipment, manufacturing, computer
-
it to their meaningful contexts, finding it in interesting media and so on. Only when love is strong, memory work can follow.⏎Yesterday, we
-
which you don’t know the return column. For example, depending on the specifications, you want to return the value from either column B, C or D and so on. In such case, we are speaking of 2D lookups or
-
but so is the rest of the Middle East. In Egypt, potential presidential contender Gamal
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: if the question, as stated in the question, says that the problem says that the problem says that, for example, for example, in a problem that is said in the statement, the problem says that, in the problem statement, a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, mull and the covering material as depicted in figure 180. While doing this, make sure that the blade does not reach the first signature. Now cut off the threads as shown in figure 181 and
-
Saint-Galmier (Loire).⏎Seven years later, he found a partnership with Bessac and the painter Joséphus Veyrat. Then, he set up his own business in Grigny, around ten kilometres from
-
at some distant object like a mountain or tree, hold your thumb at arm’s length, close one eye, and then move your thumb to block the object. Now, holding your thumb still, move your head in a small circle. Even though
-
: improving our health and protecting the environment.⏎We’ve addressed the health benefits of choosing a plant-based diet in our blog ‘8 Health Benefits of Going Vegan’. Now let’s take a look at the benefits of a vegan diet for our environment…⏎
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
- L17:845 excites comma after an introductory phrase
- L17:1716 excites
- L11:1370 weakly excites 'then' in step-by-step sequences
- L17:1067 weakly excites
- L17:1426 weakly excites
- L3:1914 weakly suppresses
- L4:42 weakly suppresses
- L10:341 weakly excites "of" inside treaty names
- L13:1108 weakly suppresses
- L16:77 weakly excites
- L2:1494 weakly suppresses
- L11:31 weakly suppresses
- L2:1451 weakly suppresses
- L5:1755 weakly excites
- L5:129 weakly excites
- L3:859 weakly suppresses
- L6:330 weakly excites
- L6:84 weakly excites
- L2:205 weakly suppresses
- L12:2911 weakly suppresses
- L4:592 weakly excites
- L14:2475 weakly suppresses
- L0:1815 weakly excites
- L8:705 weakly excites
Feeds later layers
- L23:221 excites
- L27:1006 excites
- L20:2429 excites
- L20:663 excites
- L22:1056 excites
- L25:1425 excites
- L19:1593 excites
- L19:2501 weakly suppresses
- L19:2110 weakly excites
- L21:1546 weakly excites
- L23:3040 weakly excites
- L19:934 weakly excites
- L27:1964 weakly excites
- L21:2083 weakly excites
- L26:1007 weakly excites
- L24:423 weakly excites
- L25:2171 weakly suppresses
- L20:513 weakly excites
- L19:1128 weakly excites
- L20:1811 weakly excites 'by' before a deadline
- L23:621 weakly excites
- L23:2377 weakly excites
- L27:1750 weakly excites
- L26:304 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
- then
- then
- Then
- Then
- 然后
- →then
- 然後
- THEN
- THEN
- 然后再
Suppresses
- came
- comes
- icular
- store
- くる
- .setProperty
- prü
- VarChar
- filer
- 毫