Neuron Explorer

Qwen3-0.6B

saving and reducing in energy talk

A neuron in layer 18 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 words like 'save', 'reduce' and 'low' when they appear in talk about cutting energy use, emissions or costs. It also has a strong negative response to the word 'plagiarism' in spammy promotional text. Turning it up made the model repeat words like 'size' and 'rain'.

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

    , they provide a transparent benchmark for evaluating national commitments.⏎DDPP research teams study decarbonization in their own countries. Each team defines its own pathways to a low carbon energy system that still provides all the energy services their society needs, taking into

  2. #2 peak

    products, from carpet and cubicles to air fresheners and pesticides, can contribute to a chemical cocktail of indoor pollutants. And as builders and occupants increasingly focus on saving energy,

  3. #3 peak

    reduce its carbon emissions, as well as remain competitive within its own market environment.⏎Energy usage is a consideration for both governments and businesses alike, in the quest to reduce carbon footprints and energy costs.⏎This debate is especially relevant to the 1

  4. #4 peak

    these systems expensive to purchase and maintain, but they also take up significant space in data centers and other critical facilities. Solar mobile power sources offer a more cost-effective and space-efficient alternative to UPS power solar systems, providing reliable backup power without the high

  5. #5 peak

    a deep sense of accomplishment and mastery, but can also be time-consuming and may not be the most effective approach. Smart work, on the other hand, can save time and increase productivity, but may lead to a lack of focus on the details

  6. #6 peak

    put, all the drives are used as one and the data on all of them is identical. Such a setup has two key advantages over using a single drive to save data - the first one is redundancy, so if one drive breaks down, the

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 is the value of the best-balance-balance-what-something-something-something-something-something-something-something-something-something-something-something-something-something-something-something-something-something-s

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

    x7 hours helping students in their Characteristics of Linear Programming related problems. We provide step by step Characteristics of Linear Programming question's answers with 100% plagiarism free content. We prepare quality content and notes for Characteristics of Linear Programming topic under

  2. #2 peak

    ind for assignment help⏎- Higher degree holder and experienced experts network⏎- Punctuality and responsibility of work⏎- Quality solution with 100% plagiarism free answers⏎- Time on Delivery⏎- Privacy of information and details⏎-

  3. #3 peak

    American Beverage Association is quick to pipe up and remind reporters that the studies are inconclusive, and that some studies show drinking diet sodas instead of regular might help weight loss. But just when in doubt, we say limit your sodas, and

  4. #4 peak

    -rise construction to manifold covers in small engines. Two corrosion protection methods can increase hardness and corrosion resistance and prolong the life of steel structures.⏎The two are similar in how they

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. L15:239 strongly suppresses
  2. L16:295 strongly suppresses
  3. L15:407 strongly excites
  4. L15:521 strongly excites
  5. L15:2958 suppresses
  6. L15:1742 excites
  7. L15:2101 suppresses
  8. L14:131 excites
  9. L16:2907 suppresses
  10. L13:1373 excites
  11. L15:2812 excites
  12. L15:1935 suppresses
  13. L15:2058 excites
  14. L15:2820 excites
  15. L16:1507 excites
  16. L15:1721 suppresses
  17. L16:1228 suppresses
  18. L17:2334 suppresses
  19. L13:1784 suppresses
  20. L13:2582 excites
  21. L17:433 suppresses
  22. L17:1761 suppresses
  23. L15:1816 suppresses
  24. L17:1393 excites

Feeds later layers

  1. L19:2120 strongly excites
  2. L22:289 strongly suppresses
  3. L20:995 suppresses
  4. L25:124 excites
  5. L21:1435 excites
  6. L21:2034 suppresses
  7. L20:1609 excites
  8. L19:2792 excites
  9. L21:1191 suppresses
  10. L19:826 suppresses
  11. L19:2114 excites
  12. L21:1841 excites
  13. L27:212 excites
  14. L19:285 excites
  15. L20:2916 excites
  16. L22:725 excites
  17. L24:195 excites
  18. L27:473 excites
  19. L21:3062 excites
  20. L22:845 excites
  21. L21:2234 excites
  22. L19:528 excites
  23. L20:1579 excites
  24. L22:673 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 -6.35 and -6.083 values between -4.18 and -3.911 values between -3.91 and -3.641 values between -3.64 and -3.362 values between -3.36 and -3.091 values between -3.09 and -2.825 values between -2.82 and -2.554 values between -2.55 and -2.288 values between -2.28 and -2.0111 values between -2.01 and -1.7318 values between -1.73 and -1.4622 values between -1.46 and -1.1936 values between -1.19 and -0.9256 values between -0.92 and -0.65135 values between -0.65 and -0.38552 values between -0.38 and -0.101949 values between -0.10 and 0.17785 values between 0.17 and 0.44197 values between 0.44 and 0.7162 values between 0.71 and 0.9821 values between 0.98 and 1.2612 values between 1.26 and 1.535 values between 1.53 and 1.806 values between 1.80 and 2.075 values between 2.07 and 2.341 values between 2.34 and 2.611 values between 2.61 and 2.892 values between 2.89 and 3.162 values between 3.16 and 3.431 values between 3.43 and 3.702 values between 3.70 and 3.971 values between 4.24 and 4.52offnegativepositiveoften

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. 拥堵
  2. 烦
  3. 打扰
  4. 噪声
  5. burden
  6. costly
  7. unwanted
  8. 噪音
  9. 违法行为
  10. stressful

Suppresses

  1. riendly
  2. -free
  3. 删除成功
  4. safe
  5. -safe
  6. 安全感
  7. .getInfo
  8. ameda
  9. безопас
  10. peace