Neuron Explorer

Qwen3-0.6B

verbs of reducing or harming

A neuron in layer 16 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron lights up on verbs that describe making something smaller, weaker, or worse, such as 'lower', 'reduces', 'ruin' or 'outweigh', usually just before the thing being harmed is named. Turning it up made the model's writing drift into unrelated, loose text without a clear new 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

    “A declaration from the harmless and innocent people of God, called Quakers, against all sedition, plotters, and fighters in the world, for removing the ground of jealousy and suspicion, from both magistrates and people in the kingdom

  2. #2 peak

    control the ingredients, focusing on whole, unprocessed foods, organic fruits and vegetables, and free-range meats.⏎These common vices can increase blood pressure and even lower your life expectancy! If you smoke, it’s time to work on quitting.⏎

  3. #3 peak

    start building a more confident life?⏎It is essential to overcome negative self-talk if you want to start living a more confident life. Negative self-talk can take control of our lives and make us feel uncomfortable, lost, and scared. In order to

  4. #4 peak

    accountability. There is low trust in the system, causing survivors to feel unsupported by their chain of command. Some members reported being called names or accused of trying to ruin their abuser’s career when they attempted to report abuse.⏎While the Pentagon’s

  5. #5 peak

    ups provide a selection of pictures of possible suspects for the witness to choose from. Showups are done by showing one suspect to the witness.⏎The Fourth Amendment prohibits stopping people without cause unless they consent. To detain someone for a show up in the

  6. #6 peak

    endorphins.⏎These hormones give you a euphoric feeling that calms you down and reduces stress.⏎Stress is an underlying factor in illness as it reduces your immune system’s ability to fight off antigens. This means that the more

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 can write a code that will make use of the fact that I can choose to write a matrix in the following way: for a given value of x, I can choose to make the following matrix: A, B, C,

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

    Society (NCCS). "Childhood cancer patients and families shouldn't have to delay treatment because hospitals and cancer treatment centers ran out of medication."⏎The Fight Against Childhood Cancer Continues⏎The NCCS has provided assistance to more than

  2. #2 peak

    cities inclusive, safe, resilient and sustainable” and SDG 16 seeks to “promote just, peaceful and inclusive societies” with a focus on violence and crime reduction,

  3. #3 peak

    group of young married women. She also joins the Thanatopsis Club, a literary organization, and tries to change the club programs, which are stilted and superficial. Her suggestions make little headway. Her appointment to the library board gives

  4. #4 peak

    no surprise then that commentators including energy academic Dieter Helm and climate sceptic Nigel Lawson have seized upon shale gas as an alternative to wind turbines, which they condemn as expensive and unreliable

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. L15:2820 strongly suppresses
  2. L15:2754 strongly excites
  3. L15:2058 strongly suppresses
  4. L15:1238 excites 'less' before a comparative or negative effect
  5. L15:1935 excites
  6. L15:239 excites
  7. L15:2101 excites
  8. L13:1917 suppresses
  9. L13:2991 suppresses
  10. L12:2852 excites
  11. L15:1669 suppresses
  12. L15:2782 suppresses
  13. L14:1712 suppresses
  14. L12:1049 excites
  15. L14:2390 suppresses
  16. L14:131 suppresses
  17. L14:1900 suppresses
  18. L12:44 excites
  19. L15:2791 excites
  20. L15:1422 excites
  21. L12:2410 suppresses
  22. L12:294 suppresses
  23. L15:1742 suppresses
  24. L10:2024 excites

Feeds later layers

  1. L18:2579 suppresses
  2. L21:1435 suppresses
  3. L17:2530 excites
  4. L18:2176 excites
  5. L17:3029 suppresses
  6. L18:1964 excites saving and reducing in energy talk
  7. L17:2156 excites
  8. L18:1611 excites
  9. L17:847 suppresses
  10. L19:2120 suppresses
  11. L18:1946 suppresses
  12. L20:2182 excites
  13. L21:2034 excites
  14. L21:1397 excites
  15. L19:1674 weakly excites
  16. L17:778 weakly excites
  17. L18:2425 weakly excites
  18. L18:1090 weakly excites unclear: mid-sentence words before adverbs
  19. L18:2030 weakly excites
  20. L22:1578 weakly excites
  21. L19:1301 weakly suppresses
  22. L17:2052 weakly suppresses
  23. L22:725 weakly suppresses
  24. L18:3055 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 -5.41 and -5.051 values between -3.30 and -2.942 values between -2.94 and -2.593 values between -2.59 and -2.243 values between -2.24 and -1.8910 values between -1.89 and -1.5419 values between -1.54 and -1.1942 values between -1.19 and -0.83136 values between -0.83 and -0.48516 values between -0.48 and -0.132488 values between -0.13 and 0.22436 values between 0.22 and 0.5791 values between 0.57 and 0.9347 values between 0.93 and 1.2837 values between 1.28 and 1.6327 values between 1.63 and 1.9819 values between 1.98 and 2.3312 values between 2.33 and 2.684 values between 2.68 and 3.044 values between 3.04 and 3.392 values between 3.39 and 3.742 values between 3.74 and 4.091 values between 4.09 and 4.441 values between 4.44 and 4.802 values between 4.80 and 5.151 values between 6.91 and 7.261 values between 8.31 and 8.66offnegativepositiveoften

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. prohibiting
  2. ANNOT
  3. 不宜
  4. contra
  5. 憎
  6. cene
  7. cern
  8. 健康的
  9. 上下
  10. healthy

Suppresses

  1. 过于
  2. 绍
  3. v
  4. ortal
  5. Echo
  6. ồn
  7. assumed
  8. ReturnType
  9. superficial
  10. assumptions