comma in 'not X, but Y' contrasts
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 lights up on the comma in sentences that set up a 'not this, but that' contrast, such as 'not just a vision of the future, but...'. It seems to be waiting for the second half of the contrast. Turning it up mostly made the writing wander and repeat, so the steering results do not show a clear theme.
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.
-
1960.”⏎New technologies often adopt the terms and design characteristics of the tech that came before them. So this wasn’t simply the television set of tomorrow, it was the radio set of tomorrow — now with TV!⏎This amazing bit of
-
and feeling?⏎Piper urges us to think for the glory of God. He demonstrates from Scripture that glorifying God with our minds and hearts is not either-or, but both-and. Thinking carefully about God fuels passion and affections for God.
-
of dieting could trigger a metabolism gone haywire.⏎In this theory, it’s not stubbornness or a mental disorder that keeps anorexics from eating, it’s their own bodies. The theory could explain why it can be so difficult
-
.⏎It's an experiment that's been in the works for a couple of years at this point—an attempt to solve the problem of too-dark roads not with lights, but with glowing strips of paint. This week, crews painted the Oss stretch
-
it was likely some jobs would be lost but expected it to happen over a long period of time.⏎“It’s not just taking workers out and putting a machine in, it’s going to happen really quite gradually for a long time yet.” Professor At
-
phalia and Bavaria continue to reverberate. They have once again raised our awareness of the fact that climate change is not just a vision of the future, but has already long since begun – that we can no longer afford to continue our
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: not only not a thing that is not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not not
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is also quite a common reason as to why limit calculation can be difficult for a student. A student who has math anxiety wouldn’t not dislike math but would simply feel fear of failure when they see a difficult question/task
-
, has never even been a teacher. He professes to know how teachers should teach in the entire country and what they should teach, and yet, he is not a teacher. Coleman worked for McGraw-Hill (textbook publishing company)
-
How many inventions can you think of that were first made in the first decade of the first century AD? No, I can’t think of any either, and not just because it was a long time ago. There
-
hold.⏎Goats brows hard, leaving little in their wake of appetite. This is not a system I would recommend to sensitive ecosystems that are fully intact, not without a lot of supervision. Tethering goats takes
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
- L16:752 excites
- L10:111 suppresses
- L13:822 excites
- L11:1095 suppresses
- L4:151 excites
- L0:2047 weakly suppresses
- L8:13 weakly suppresses
- L17:443 weakly excites
- L17:1644 weakly suppresses
- L0:2806 weakly excites
- L15:2515 weakly excites
- L7:64 weakly suppresses
- L15:1293 weakly suppresses
- L3:1588 weakly suppresses
- L8:7 weakly excites
- L4:9 weakly excites
- L8:1337 weakly excites
- L11:169 weakly excites
- L16:2137 weakly suppresses
- L7:2977 weakly excites
- L16:735 weakly suppresses
- L10:504 weakly suppresses
- L12:2951 weakly suppresses
- L8:10 weakly excites
Feeds later layers
- L19:1811 strongly excites
- L22:635 excites
- L24:1779 excites
- L19:2544 excites
- L21:1717 excites
- L25:2002 excites
- L22:896 excites
- L23:434 suppresses
- L20:1873 suppresses
- L23:466 excites
- L19:2523 excites
- L26:406 suppresses
- L21:2799 excites
- L19:2171 excites
- L23:2518 excites sentence-initial 'Not' starting a clause
- L21:1974 weakly excites
- L21:544 weakly excites
- L20:3031 weakly excites
- L19:2451 weakly excites
- L19:259 weakly suppresses
- L23:1127 weakly excites
- L21:2310 weakly suppresses
- L21:3046 weakly excites 'at' in 'not at all' and 'at all'
- L19:3025 weakly excites
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
- 而是
- sondern
- anymore
- nor
- 那么简单
- אלא
- nor
- necessarily
- any
- ched
Suppresses
- 并不会
- 不知
- atica
- 不明白
- haven
- 并非
- 没有办法
- olean
- tér
- 憬