Neuron Explorer

Qwen3-0.6B

"in" after superlatives and "one of the few"

A neuron in layer 20 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 wakes up on the word "in" when it follows a phrase like "the best prophet", "the foremost intellectuals" or "one of the few maps". When it was turned up, the model got stuck repeating "and, and, and". Turned down, it wrote ordinary 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

    billion tons in the U.S. alone) over the past 30 years through the use of catalytic-converter-equipped vehicles. Virtually every automotive company in the world today relies on a system that is based on a cellular ceramic substrate technology

  2. #2 peak

    These characteristics became his signature during his career as one of the foremost intellectuals in the Afro-Asian anti-colonial movement.⏎Education and Political Activism⏎

  3. #3 peak

    |[Home] [The History of Shirayuki and Sake] [Our Products] Japanese Page|⏎Japan's finest sake in historical perspective⏎Rice was found its way into Japan about 200

  4. #4 peak

    law. The Holy Prophet being the final law-giving prophet who has come for not just one religion, but for all mankind, would therefore be the best prophet in the sight of Allah. This is what we testify to.<|endoftext|>Those aged over

  5. #5 peak

    has been home to many people for centuries and people from all part of India have stayed here and have enriched the city with their ethos and customs. Very few cities in India offer the real spirt of India at one place. Delhi is home to

  6. #6 peak

    scientists have made a surprising discovery: a gigantic aquifer of relatively fresh water trapped in porous sediments lying below the salty ocean.⏎A new book, the second in a series of primers with the Earth Institute imprint, provides an interdisciplinary overview drought

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: if I were to write a book, I would need to have a **displacement difference between the "self" [suggestion, *s} [a] *s [a] *s [a] **s [a]

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 the class in which he or she is enrolled.⏎The HCPS Office of Exceptional Education ensures that individuals employed as homebound teachers hold a Virginia teaching certificate or be determined eligible for one

  2. #2 peak

    nation.⏎According to the Title I Director’s Conference (February 2003), “highly qualified” means that all teachers must have full State certification or have passed teacher licensing exam, and hold a license to teach, and that certification

  3. #3 peak

    Causes of miscarriage can include problems with the mother's uterus or placenta, or abnormalities in the baby's development. In most cases, the

  4. #4 peak

    congenital illness that results in the degradation of the retina. People with LCA slowly lose all vision, and typically are completely blind by their 20s or 30s. The modified virus used in the therapy carries a replacement gene

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. L18:93 suppresses
  2. L13:1046 suppresses 'in' or 'on' ending a phrase
  3. L14:348 suppresses
  4. L15:727 weakly excites
  5. L12:51 weakly excites
  6. L14:1091 weakly excites
  7. L16:245 weakly suppresses
  8. L17:385 weakly suppresses
  9. L15:2502 weakly excites the gap in number ranges
  10. L1:1062 weakly suppresses
  11. L5:5 weakly suppresses digit after a label like paper or type
  12. L11:104 weakly excites
  13. L1:388 weakly excites
  14. L17:317 weakly excites 'of' after a noun phrase
  15. L13:3043 weakly excites
  16. L19:1944 weakly excites
  17. L14:2957 weakly suppresses
  18. L9:29 weakly excites opening parenthesis before an acronym
  19. L10:125 weakly excites
  20. L15:1738 weakly suppresses
  21. L18:565 weakly suppresses
  22. L15:286 weakly excites
  23. L17:62 weakly suppresses
  24. L18:598 weakly suppresses

Feeds later layers

  1. L25:6 suppresses
  2. L24:205 excites
  3. L21:393 excites unclear: relative 'that' and list items
  4. L22:632 excites
  5. L26:347 excites the 're' in self-reliant and similar
  6. L27:289 excites
  7. L27:291 excites
  8. L26:74 suppresses
  9. L23:1798 excites
  10. L21:76 suppresses
  11. L27:358 weakly excites
  12. L25:969 weakly excites
  13. L26:2056 weakly suppresses
  14. L23:379 weakly excites
  15. L27:1364 weakly excites
  16. L27:2924 weakly suppresses
  17. L26:65 weakly suppresses
  18. L22:158 weakly suppresses
  19. L25:2388 weakly suppresses
  20. L26:307 weakly excites
  21. L21:2638 weakly excites
  22. L21:530 weakly suppresses
  23. L21:395 weakly excites
  24. L26:836 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.

2 values between -14.27 and -13.581 values between -12.90 and -12.221 values between -12.22 and -11.542 values between -11.54 and -10.863 values between -10.86 and -10.171 values between -10.17 and -9.491 values between -8.81 and -8.135 values between -8.13 and -7.455 values between -7.45 and -6.767 values between -6.76 and -6.0813 values between -6.08 and -5.4014 values between -5.40 and -4.7231 values between -4.72 and -4.0426 values between -4.04 and -3.3547 values between -3.35 and -2.6774 values between -2.67 and -1.99106 values between -1.99 and -1.31214 values between -1.31 and -0.631762 values between -0.63 and 0.06808 values between 0.06 and 0.74281 values between 0.74 and 1.42171 values between 1.42 and 2.10102 values between 2.10 and 2.7973 values between 2.79 and 3.4756 values between 3.47 and 4.1535 values between 4.15 and 4.8317 values between 4.83 and 5.5115 values between 5.51 and 6.2016 values between 6.20 and 6.884 values between 6.88 and 7.563 values between 7.56 and 8.244 values between 8.24 and 8.922 values between 8.92 and 9.612 values between 9.61 and 10.291 values between 10.29 and 10.971 values between 10.97 and 11.651 values between 11.65 and 12.331 values between 12.33 and 13.02offnegativepositiveoften

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. and
  2. 和
  3. 性和
  4. 和
  5. 、
  6. or
  7. và
  8. 和个人
  9. hoot
  10. and

Suppresses

  1. ernaut
  2. 一一
  3. 新规
  4. 们都
  5. liers
  6. со
  7. irling
  8. sa
  9. 也不例外
  10. chu