Neuron Explorer

Qwen3-0.6B

"and" joining contrasting pairs

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 lights up on the word "and" when it links two opposite or contrasting things, like "art and the artist" or "truth and lies". When it was turned up, the model's writing drifted into repetitive, loosely related phrases, so the steering did not reveal 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

    of contrasting brilliant color and shade. Some galleries insist that artists stand alongside their works during shows, juxtaposing the art and the artist.⏎Quiz: What is juxtaposition?⏎- The act of placing two

  2. #2 peak

    atism.⏎3. Instead of rigid binary opposition of truth and lies Tetraphilosophy offer an infinite continuum and constant movement from deceit to

  3. #3 peak

    sometimes found in the opposing juxtaposition of different images placed side by side. Artists place the delicate with the blunt, juxtapose complex with simple, or mingle joyful and sad. In one artist’s work, the tension between dark and light colors offers

  4. #4 peak

    God. Piper contends that "thinking is indispensable on the path to passion for God." So how are we to maintain a healthy balance of mind and heart, thinking and feeling?⏎Piper urges us to think for the glory of God. He demonstrates

  5. #5 peak

    women and girls.<|endoftext|>Microcontroller vs FPGA: Understanding the Key Differences⏎When it comes to embedded systems design, two technologies that often come up are microcontrollers and field-programmable gate arrays (FPGAs). While both serve similar purposes

  6. #6 peak

    is no need, however, to sacrifice good acoustics when selecting products for green buildings. Architects and interior designers do not have to choose between sustainability, design and acoustic needs—they can have all

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 list is a result of the system, so I will have to accept it as the next step. So I should put into a higher priority: this is the first option. The sentence: B: C = D = E = F

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

    of a story could support multiple interpretations. Each subject saw two eight-sentence stories of each of four types (no-cause, unbiased, biased-proximal, and biased-distal), designed to support multiple interpretations. After reading or hearing each

  2. #2 peak

    M?⏎A. Can have subnets of different sizes⏎B. Subnets must be in same size⏎C. No required of subnet⏎D. All of above⏎6-3. What does the port number in a TCP connection specify

  3. #3 peak

    = 0.02). Capture and recapture data indicated three infection dynamics: 1) negative–positive-negative infection; 2) positive–negative-positive infection; and 3) positive at all sampling times.⏎These results indicate that

  4. #4 peak

    health department and is not intended to provide diagnosis, treatment, or medical advice.⏎In this article:⏎- Fat-Soluble Vitamins vs. Water-Soluble Vitamins⏎- Vitamin A⏎- Risk Factors for Vitamin A Deficiency⏎

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. L11:1884 suppresses
  2. L9:1449 suppresses
  3. L11:881 excites "with" after combine, mix or react verbs
  4. L11:2631 excites
  5. L10:1232 excites
  6. L3:1253 suppresses
  7. L7:602 suppresses
  8. L11:2772 weakly suppresses
  9. L7:1399 weakly excites "to" between a starting and ending value
  10. L8:2999 weakly suppresses
  11. L7:1156 weakly suppresses
  12. L11:1291 weakly suppresses
  13. L6:1219 weakly excites
  14. L8:1238 weakly suppresses
  15. L1:2463 weakly suppresses
  16. L9:1820 weakly excites
  17. L11:1511 weakly excites
  18. L8:1267 weakly excites
  19. L11:2292 weakly excites 'than' in comparisons
  20. L10:1658 weakly suppresses
  21. L2:2851 weakly suppresses
  22. L5:1256 weakly excites
  23. L11:628 weakly suppresses
  24. L2:1637 weakly excites

Feeds later layers

  1. L13:2356 excites
  2. L16:686 excites
  3. L15:2400 suppresses
  4. L13:1024 excites
  5. L15:1733 excites
  6. L14:1688 weakly excites
  7. L14:2422 weakly excites
  8. L13:758 weakly excites 'and' joining a pair of options
  9. L13:1768 weakly excites
  10. L15:3021 weakly excites
  11. L14:602 weakly suppresses
  12. L18:1801 weakly excites
  13. L15:1134 weakly excites
  14. L15:1280 weakly suppresses
  15. L22:2455 weakly excites
  16. L16:1969 weakly suppresses
  17. L14:2079 weakly excites prepositions after comparison words
  18. L18:2528 weakly suppresses
  19. L14:2820 weakly suppresses "and" or "or" linking two counted items
  20. L13:951 weakly excites
  21. L13:2931 weakly excites
  22. L15:550 weakly suppresses
  23. L14:1134 weakly excites
  24. L18:2542 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 -1.77 and -1.611 values between -1.31 and -1.161 values between -1.16 and -1.002 values between -1.00 and -0.8510 values between -0.85 and -0.7028 values between -0.70 and -0.5556 values between -0.55 and -0.39132 values between -0.39 and -0.24624 values between -0.24 and -0.092360 values between -0.09 and 0.06565 values between 0.06 and 0.2274 values between 0.22 and 0.3721 values between 0.37 and 0.5210 values between 0.52 and 0.675 values between 0.67 and 0.833 values between 0.83 and 0.983 values between 0.98 and 1.131 values between 1.13 and 1.283 values between 1.44 and 1.591 values between 1.74 and 1.891 values between 1.89 and 2.051 values between 2.20 and 2.351 values between 2.35 and 2.503 values between 3.11 and 3.271 values between 4.18 and 4.33offnegativepositiveoften

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. ้อม
  2. (
  3. 和支持
  4. crets
  5. ics
  6. iative
  7. luent
  8. i
  9. ebp
  10. e

Suppresses

  1. tring
  2. ListComponent
  3. LayoutPanel
  4. ableView
  5. 谠
  6. ReadStream
  7. WithTag
  8. 当たり
  9. 로서
  10. ệnh