Neuron Explorer

Qwen3-0.6B

word-start fragments mid-word

A neuron in layer 26 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the opening piece of a longer word or name that is cut into several tokens, like "Ger" in Gerstenmaier, "Car" in Carole, or "sc" in scorpions. The pattern is broad and not limited to names. Steering did not produce a clear theme: turning it up gave loose, rambling text and turning it down gave repetitive 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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

  1. #1 peak

    some extra time to make a decision. It now hopes to make a launch decision on Monday morning.⏎“We have really two competing objectives,” says NASA associate administrator Bill Gerstenmaier. “One, we want to get the vehicle ready to go

  2. #2 peak

    ners⏎- Visual Supports, Core Words, and Martin Luther King⏎- Core Values: Suggestions for Language Building with AAC Learners⏎This post was written by Carole Zangari<|endoftext|>Australia ranked in the middle of the pack in the recent

  3. #3 peak

    of Text Reinterpretation Skill.⏎Bonitatibus, Gary; Beal, Carole R.⏎Four studies investigated when children became able to detect that the words of

  4. #4 peak

    largest consumer of cocaine and has been leading efforts to eradicate coca production in the Andes for decades.⏎Bolivia is the world's

  5. #5 peak

    a self bailer, a whistle and bow/stern line. It is also a good idea to be prepared with rescue ropes, throw lines, a knife and carabineers.⏎Dress properly for an unexpected cold-water swim and changing weather

  6. #6 peak

    Community. So, what are a few of the events that you should definitely attend? There’s the Wild World of Animals that features 14 interesting animals—from scorpions to legless lizards—sponsored by N.A.T.U.R

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: The first plays a plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays plays

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

    aramidia, along the western shores of the great seaway that separated North America; this landmass hosted a vast array of unique dinosaur species and served as the crucible of evolution for iconic dinosaur groups such as the horn

  2. #2 peak

    erger developed a safety coffin in 1829 that had bell strings attached to both arms and legs of the corpse.⏎In modern times Fabrizio Caselli designed a safety coffin with an emergency alarm, intercom, a light,

  3. #3 peak

    ad region began demanding independence from the southern, Bamako-based government. After forcing the Malian military from the north, however, the separatist forces soon became embroiled in a conflict of their own, between the original Mouvement National pour

  4. #4 peak

    1882. But his absence did little to⏎resolve the theological divide and soon after his return he became embroiled⏎in fresh controversy over his failure to denounce⏎George Higin

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. L20:54 suppresses
  2. L24:1637 suppresses
  3. L21:583 weakly suppresses
  4. L23:13 weakly excites
  5. L25:52 weakly excites
  6. L17:1822 weakly suppresses 'should' in advice and claims
  7. L16:1476 weakly excites
  8. L5:2956 weakly excites
  9. L0:2027 weakly suppresses
  10. L2:828 weakly suppresses
  11. L17:2131 weakly excites
  12. L11:226 weakly excites
  13. L18:927 weakly suppresses
  14. L25:986 weakly excites
  15. L17:473 weakly excites
  16. L25:26 weakly suppresses
  17. L24:488 weakly excites
  18. L10:95 weakly suppresses
  19. L22:1081 weakly excites
  20. L14:1523 weakly excites
  21. L21:592 weakly excites
  22. L19:1427 weakly excites
  23. L19:1331 weakly excites
  24. L12:653 weakly suppresses

Feeds later layers

  1. L27:427 weakly suppresses
  2. L27:515 weakly excites
  3. L27:1166 weakly excites
  4. L27:965 weakly suppresses
  5. L27:1620 weakly excites
  6. L27:2009 weakly suppresses
  7. L27:2146 weakly suppresses
  8. L27:461 weakly suppresses
  9. L27:1789 weakly excites
  10. L27:401 weakly excites
  11. L27:2753 weakly suppresses
  12. L27:419 weakly suppresses
  13. L27:2034 weakly suppresses
  14. L27:1169 weakly excites
  15. L27:216 weakly excites
  16. L27:1363 weakly suppresses
  17. L27:1054 weakly suppresses
  18. L27:1626 weakly suppresses
  19. L27:2698 weakly suppresses
  20. L27:897 weakly excites
  21. L27:890 weakly suppresses
  22. L27:950 weakly excites
  23. L27:647 weakly suppresses
  24. L27:1125 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 -26.75 and -25.521 values between -20.58 and -19.341 values between -19.34 and -18.111 values between -16.88 and -15.643 values between -15.64 and -14.415 values between -14.41 and -13.174 values between -13.17 and -11.9410 values between -11.94 and -10.7015 values between -10.70 and -9.4736 values between -9.47 and -8.2336 values between -8.23 and -7.0070 values between -7.00 and -5.77103 values between -5.77 and -4.53154 values between -4.53 and -3.30259 values between -3.30 and -2.06469 values between -2.06 and -0.831871 values between -0.83 and 0.41469 values between 0.41 and 1.64176 values between 1.64 and 2.8885 values between 2.88 and 4.1143 values between 4.11 and 5.3426 values between 5.34 and 6.5824 values between 6.58 and 7.8111 values between 7.81 and 9.0512 values between 9.05 and 10.287 values between 10.28 and 11.523 values between 11.52 and 12.754 values between 12.75 and 13.981 values between 13.98 and 15.222 values between 15.22 and 16.451 values between 16.45 and 17.691 values between 17.69 and 18.922 values between 18.92 and 20.162 values between 21.39 and 22.63offnegativepositiveoften

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. endcode
  2. mania
  3. ubs
  4. phans
  5. ouble
  6. -sama
  7. 齊
  8. sburgh
  9. blings
  10. nominee

Suppresses

  1. plays
  2. boils
  3. Plays
  4. rests
  5. matches
  6. turns
  7. renders
  8. loses
  9. să
  10. ותר