Neuron Explorer

Qwen3-0.6B

"but" after a concession

A neuron in layer 12 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 fires on "but" when it follows a difficulty or admission and turns to a positive point, such as "it may sound like a pipedream but" or "not easy, but". It is quiet on "even" in the opposite sense. Turning it up made the model write loose, unrelated sentences rather than showing this pattern.

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

    that you<|endoftext|>America's first industry was neither textiles, nor iron, nor arms, but rather glassmaking. Early manufacturing attempts met with hardships and often failure, but in time the business of glassmaking and the demand for glass products became an important

  2. #2 peak

    for them to make new friends, master the teaching material, and be successful in any extracurricular activity. That, of course, is not easy, but fortunately, there are strategies to help one build confidence both in

  3. #3 peak

    even hear a simple word like "mama," "dada," or "ball." You baby isn't able to use these words with real meaning yet – but she's working on it.⏎Your baby's beginning to understand how objects relate to

  4. #4 peak

    the answer to the low cost decarbonisation of heat puzzle, or a flight of fancy? Well, it may sound like a pipedream (sorry!) but try telling that to the 20 members of the Young Persons Network who visited

  5. #5 peak

    could help treat cancer that already has developed. None have been approved for use yet, but a number of experimental therapeutic vaccines have shown promise at helping the body recognize cancer cells

  6. #6 peak

    a complex dilemma leaves you uncertain, there are some tools that can help you work your way to a good decision. They can’t tell you what to do, but they can certainly help you figure it out for yourself.⏎When I’m forced to

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: I have no, and I have not, all of the above, given you a mirror, and I have only just learned something to be a mirror to you, and I am 182, 13.5 weeks

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

    .)⏎It is harder to sort out the chronological data for the period of the kings (between 1000 and 500 B.C.), and even more difficult to be certain about the period between the exile and the New Testament (

  2. #2 peak

    DPFs) and a new generation of substrates designed for diesel engines.⏎An Enabling TechnologyDPFs based on Corning's technology aren't new; in fact, Corning first introduced the cellular ceramic particulate filter to remove soot from

  3. #3 peak

    and UNICEF in 1995 and adopted in over 90 countries and territories. However, few countries have implemented it at scale, and even fewer have effectively adapted their education systems to include health promotion. The new global standards

  4. #4 peak

    offset environmental gains. If a family buys a fuel-efficient vehicle but chooses to drive more miles than they previously did, there is no gain for the environment. Also, while the average U.S. household size is decreasing, Americans are generally choosing to

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. L11:1427 suppresses
  2. L11:2703 suppresses
  3. L10:2215 weakly suppresses
  4. L11:1868 weakly excites
  5. L6:1561 weakly suppresses
  6. L8:384 weakly suppresses
  7. L11:2992 weakly excites
  8. L7:348 weakly suppresses
  9. L1:549 weakly excites "g" in "e.g." examples
  10. L9:1303 weakly suppresses
  11. L8:1920 weakly suppresses
  12. L11:1372 weakly suppresses
  13. L10:2937 weakly excites
  14. L9:2646 weakly excites
  15. L10:1232 weakly excites
  16. L7:1490 weakly suppresses
  17. L3:1906 weakly suppresses
  18. L10:2244 weakly suppresses
  19. L8:2609 weakly suppresses
  20. L10:2304 weakly suppresses
  21. L3:594 weakly suppresses
  22. L4:2264 weakly suppresses
  23. L7:1586 weakly suppresses
  24. L0:2457 weakly excites

Feeds later layers

  1. L13:642 excites
  2. L18:1369 suppresses
  3. L22:1437 suppresses
  4. L20:1960 suppresses
  5. L19:826 excites
  6. L14:111 excites
  7. L16:1747 suppresses
  8. L17:847 suppresses
  9. L17:2025 suppresses
  10. L15:1725 suppresses
  11. L16:2849 suppresses
  12. L23:391 suppresses
  13. L16:295 excites
  14. L15:1484 suppresses
  15. L15:1157 suppresses
  16. L22:289 excites
  17. L15:1669 suppresses
  18. L25:124 suppresses
  19. L13:1447 suppresses
  20. L13:587 excites
  21. L13:924 suppresses
  22. L13:678 excites
  23. L17:778 excites
  24. L21:1045 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 -5.13 and -4.891 values between -4.66 and -4.421 values between -3.72 and -3.481 values between -3.25 and -3.011 values between -2.30 and -2.071 values between -2.07 and -1.831 values between -1.83 and -1.601 values between -1.60 and -1.366 values between -1.36 and -1.136 values between -1.13 and -0.8911 values between -0.89 and -0.6624 values between -0.66 and -0.42169 values between -0.42 and -0.192702 values between -0.19 and 0.05895 values between 0.05 and 0.2857 values between 0.28 and 0.5214 values between 0.52 and 0.753 values between 0.75 and 0.993 values between 0.99 and 1.222 values between 1.22 and 1.461 values between 1.46 and 1.691 values between 1.69 and 1.933 values between 1.93 and 2.161 values between 2.87 and 3.101 values between 3.34 and 3.571 values between 4.04 and 4.28offnegativepositiveoften

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. 准时
  2. playable
  3. tempts
  4. 맞
  5. raries
  6. fortunately
  7. nonetheless
  8. erto
  9. -winning
  10. etection

Suppresses

  1. 匮
  2. inconsist
  3. 凋
  4. 得不到
  5. aside
  6. ath
  7. 不好
  8. bbe
  9. ôme
  10. lacked