Neuron Explorer

Qwen3-0.6B

the 'cul' fragment in culprit and culminate

A neuron in layer 24 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 small piece 'cul' when it starts words like 'culprits' or 'culminated', and on similar word starts. When it was turned up, the written text did 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 activation was highest. Amber marks positive activation; the outlined token is the peak. Hover a token for its strength.

  1. #1 peak

    examination of mummified bodies has revealed that ancient Egyptians suffered from hardening of the arteries in surprising frequency, suggesting that blame for heart disease extends beyond the modern culprits of smoking, fast food and the remote control.⏎Among 22

  2. #2 peak

    decline from participating in unwanted situations, many people are saying social media can worsen peer pressure. Sites like Facebook, YouTube, Multiply, and Twitter are the prime culprits for such accusations. But is it true? How can peer pressure be

  3. #3 peak

    and Collaboration.⏎Soon after, I talked to my friend Karen Holmes about it who created an escape room for her Leadership class. Her scenario involves a zombie apocalypse and culminates in students finding the vaccine (she draws on her science background to have

  4. #4 peak

    French king supported the merchants, while the English sided with the artisans.⏎This tension culminated in 1337, over the status of the feudal territories of

  5. #5 peak

    . This two-fold classification of illnesses, based on whether they are caused by over nutrition3 or under nutrition, represents the germinal stages of theorisation that finally culminated in the conception of tridosha (three illness-causing factors). This

  6. #6 peak

    hood. This shift is part of a larger trend: American higher education has swung away from belief in educational access for all into a toxic obsession with meritocracy, culminating in the annual Squid Game frenzy that exploits

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: what if the one single fact that is very important to the world is one single fact that has been very subjectively subjectively subjected to human subjectively subjectively subjectively subjectively subjectively subjectively subjectively subjectively subjectively subjectively

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

    rare disease that often impacts the lungs, affecting breathing. “The first time that I heard about Alpha 1 Antitrypsin Deficiency was on Wednesday, September 4

  2. #2 peak

    the burden of being a gay Asian student in a school they believe did not support him.⏎Miyairi created the new anti-bullying program specifically to address these issues.⏎"It's important for [students] to

  3. #3 peak

    ) used to act, perform, fight and/or kill for the enjoyment of humans. The term encompasses many different forms of entertainment – from circuses to movies to bullfighting.⏎Share information about animals in entertainment.⏎- Animals have been used for

  4. #4 peak

    program provides a range of services in a safe and supportive environment to both youth and adults, and their families, who are dealing with anorexia nervosa, bulimia nervosa and avoidant/restrictive food intake disorder and binge eating

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. L7:2613 suppresses
  2. L2:1008 weakly excites
  3. L9:3011 weakly suppresses
  4. L1:2218 weakly suppresses
  5. L21:1662 weakly suppresses comma introducing a quoted claim
  6. L8:737 weakly excites
  7. L14:609 weakly suppresses
  8. L20:151 weakly excites
  9. L22:2960 weakly suppresses
  10. L17:542 weakly excites
  11. L19:196 weakly excites
  12. L3:1101 weakly suppresses
  13. L21:1189 weakly excites
  14. L22:2138 weakly suppresses
  15. L3:1579 weakly excites
  16. L17:707 weakly suppresses
  17. L21:237 weakly suppresses
  18. L22:316 weakly suppresses
  19. L2:1321 weakly suppresses
  20. L5:9 weakly excites
  21. L23:1207 weakly excites
  22. L9:3055 weakly excites
  23. L9:329 weakly excites
  24. L1:951 weakly excites

Feeds later layers

  1. L26:598 suppresses spot just before a year or century
  2. L25:1452 suppresses
  3. L27:328 suppresses
  4. L25:2383 weakly suppresses
  5. L27:678 weakly excites
  6. L26:880 weakly suppresses
  7. L25:450 weakly excites
  8. L27:1522 weakly suppresses
  9. L27:902 weakly suppresses
  10. L27:180 weakly excites
  11. L25:333 weakly excites
  12. L25:752 weakly suppresses
  13. L27:572 weakly suppresses
  14. L25:621 weakly excites
  15. L25:850 weakly excites
  16. L26:2437 weakly suppresses
  17. L27:1246 weakly suppresses
  18. L27:1656 weakly excites
  19. L25:1895 weakly excites first digit of a year like 2020
  20. L25:1005 weakly suppresses
  21. L27:1215 weakly excites
  22. L25:1749 weakly excites
  23. L25:1733 weakly excites the 'per' piece in perished and per se
  24. L27:2167 weakly excites

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 -13.63 and -12.651 values between -7.77 and -6.803 values between -6.80 and -5.825 values between -5.82 and -4.851 values between -4.85 and -3.875 values between -3.87 and -2.905 values between -2.90 and -1.9232 values between -1.92 and -0.952113 values between -0.95 and 0.031689 values between 0.03 and 1.0121 values between 1.01 and 1.987 values between 1.98 and 2.964 values between 2.96 and 3.935 values between 3.93 and 4.911 values between 4.91 and 5.882 values between 5.88 and 6.863 values between 6.86 and 7.834 values between 7.83 and 8.812 values between 8.81 and 9.781 values between 10.76 and 11.741 values between 12.71 and 13.691 values between 16.61 and 17.591 values between 24.42 and 25.39offnegativepositiveoften

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. over
  2. Cards
  3. Scalars
  4. head
  5. enames
  6. oping
  7. '').
  8. edef
  9. Up
  10. cards

Suppresses

  1. онт
  2. EDIUM
  3. chaud
  4. 巯
  5. fact
  6. arel
  7. ://'
  8. 薷
  9. _DENIED
  10. %"><