Neuron Explorer

Qwen3-0.6B

"but" after a "not just X" contrast

A neuron in layer 3 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron lights up on the word "but" when the sentence has just said what something is not, as in "not just X, but also Y". Turning it up made the model's text drift into unrelated topics, so the steering results do not show a clear effect.

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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

  1. #1 peak

    closer to the original meaning of the term emotions (emovere is Latin for the term "to move"). As we know, emotions are not just passive reflections, but they lead to active changes in the organism; from small changes in attention to behavioral

  2. #2 peak

    be the end in itself, and I think the bigger meaning is missed. The point isn’t just that God did a miracle, but that the miracle was the means to allow Him to be properly worshiped. The

  3. #3 peak

    matter of being nice or PC. It’s a matter of biology and physics. Evolutionary biology shows us we are not governed by the “selfish gene” but by a “hyper-cooperative” nature in which “rec

  4. #4 peak

    Tahira Abdullah expressed her utter sorrow at the passing of another prominent feminist icon in the South Asia region.⏎“It is not just an irreparable loss for India, but also for feminists and peace activists all over South Asia and beyond,” she wrote in

  5. #5 peak

    rs to a universal ideal?⏎The resisters in On the Road to the Wolf’s Lair are not the singular souls doomed to failure by the massive Nazi machinery, but those who emerged from the Third Reich itself—those people whose cultural, administrative,

  6. #6 peak

    real time to the appropriate medical practitioner or the ubiquitous App or other software. The genius of the system is not just that it can detect potentially life threatening conditions, but also that it can monitor,<|endoftext|>This material must not be used for commercial purposes

Steering

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.

turned down a lot
Prompt

Here is something I have been thinking about lately: it's a simple approach to a system that's as easy as it gets. So, it is a simple approach to a system that's as easy as it gets. I will do my own thing, like that. It is

Most negative on

Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.

  1. #1 peak

    Pearsonχ2 test, multivariate logistic regression models and discriminant models were fitted into its comparisons with the non-norovirus group and other pathogens.⏎Norovirus was detected in 22.91% of sampled diarrhea patients.

  2. #2 peak

    at least once a day. Specific germ-prone areas such as bathrooms should also be given more attention.⏎- Clean and disinfect regularly. The flu and norovirus can spread when people touch infected surfaces and then touch their eyes, mouth or

  3. #3 peak

    several meters wide – floating in the Great Pacific Garbage Patch. These ghost nets are known to be a major hazard to vessels<|endoftext|>Norovirus is an important cause of gastroenteritis both in children and adults. In

  4. #4 peak

    by dogs.⏎- Rabies. Rabies is a single strand RNA virus belonging to the Rhabdoviridae family. …⏎- Noroviruses. …⏎- Pasteurella. …⏎- Yersinia

Connections

Show on map

Which neurons in earlier layers feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.

Fed by earlier layers

  1. L1:1464 weakly excites
  2. L2:1118 weakly excites
  3. L2:483 weakly suppresses
  4. L2:1204 weakly excites
  5. L2:899 weakly suppresses
  6. L2:537 weakly suppresses
  7. L2:111 weakly suppresses
  8. L2:36 weakly excites
  9. L2:2692 weakly excites
  10. L2:1951 weakly excites
  11. L0:2407 weakly suppresses
  12. L0:799 weakly suppresses
  13. L2:1729 weakly excites
  14. L2:2710 weakly suppresses
  15. L2:75 weakly suppresses
  16. L1:1037 weakly suppresses
  17. L1:2477 weakly suppresses
  18. L2:2808 weakly suppresses
  19. L0:2563 weakly excites
  20. L1:1099 weakly excites
  21. L2:2236 weakly excites
  22. L2:1642 weakly suppresses
  23. L1:2059 weakly suppresses
  24. L2:503 weakly excites

Feeds later layers

  1. L14:2166 excites
  2. L12:92 excites
  3. L8:1876 excites
  4. L6:3050 weakly excites
  5. L15:833 weakly suppresses
  6. L13:2356 weakly excites
  7. L6:1522 weakly suppresses
  8. L5:2087 weakly suppresses
  9. L7:42 weakly excites
  10. L11:1995 weakly suppresses
  11. L9:563 weakly suppresses
  12. L26:398 weakly suppresses
  13. L7:1230 weakly suppresses
  14. L25:2017 weakly suppresses
  15. L25:953 weakly suppresses
  16. L7:2336 weakly suppresses
  17. L12:1835 weakly excites
  18. L12:2730 weakly excites
  19. L16:1294 weakly excites
  20. L13:1738 weakly suppresses
  21. L26:906 weakly excites
  22. L11:2948 weakly suppresses
  23. L16:1937 weakly excites 'and' joining two time or quantity ranges
  24. L4:2677 weakly suppresses

Activation histogram

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.

1 values between -3.00 and -2.744 values between -1.19 and -0.933 values between -0.93 and -0.687 values between -0.68 and -0.42436 values between -0.42 and -0.162759 values between -0.16 and 0.10658 values between 0.10 and 0.3623 values between 0.36 and 0.617 values between 0.61 and 0.872 values between 0.87 and 1.131 values between 1.13 and 1.391 values between 1.39 and 1.652 values between 1.65 and 1.901 values between 1.90 and 2.161 values between 3.45 and 3.711 values between 4.22 and 4.481 values between 7.06 and 7.32offnegativepositiveoften

Logit lens

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 layers instead.

Promotes

  1. овал
  2. irection
  3. oit
  4. ohl
  5. orical
  6. ocos
  7. רס
  8. едак
  9. ி
  10. emplates

Suppresses

  1. 卫生健康
  2. 健康的
  3. resid
  4. Signup
  5. ties
  6. 계
  7. rique
  8. ieten
  9. EEE
  10. ),"