Neuron Explorer

Qwen3-0.6B

the '1' starting a century

A neuron in layer 26 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 reacts to the digit 1 at the start of a century, like the 1 in '1800s' or '18th century'. When turned up the model drifted into talk about 'four' and 'middle', and when turned down it repeated the digit 8.

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

    let letter, Reverend Dimmesdale, and burrs to add onto the overall theme of guilt.⏎Nathaniel Hawthorne was a writer in the 1800s, an anti

  2. #2 peak

    Cudnow. Count and Polish writer.⏎In the historical narrative Soplica’s Memoirs (1839), which takes place in the mid-18th century, Rzewuski created a gallery of types from the Polish

  3. #3 peak

    Book Society', founded in 1878.⏎Polled Herefords were developed from the horned Hereford breed which was founded in the mid-18th century by the farmers of Hereford County, England. Among the horn

  4. #4 peak

    the Reformation.⏎John Knox and the Monstrous Regiment of Women | Presbyterian Historical Society.<|endoftext|>The Booth family was an English-American theatrical family of the 19th century. Its most famous and well known members were Edwin Booth, one

  5. #5 peak

    817 and 1826, whose purpose was to educate and train non-white Christian missionaries. The book focuses upon the evangelical movement of the 19th century, and American attitudes on interracial relationships as demonstrated in the unfavorable,

  6. #6 peak

    1760 and 1782. This piece is a excellent example of that cross section of Baroque and Classical styles found in the mid-1700s, and bares stylistic similarities to Haydn's own

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: Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ Ⅷ �

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

    anticipation of the availability of the more secure and complex WPA2, which became available in 2004 and is a common shorthand for the full IEEE 802.11i (or IEEE 802.1

  2. #2 peak

    ⏎Ink painting originated in China during the Tang dynasty (AD 618-907) and was known in Japan by the eighth century.

  3. #3 peak

    heart is resting between beats.⏎Along with high cholesterol and being overweight or obese, a high blood pressure diagnosis, also known as hypertension (over 130/80), is a major risk factor for heart disease or stroke. And because

  4. #4 peak

    a greater optical density than air.⏎Substance Speed of light (m/sec)⏎Vacuum 3 x 108⏎Air 3 x 108⏎Water 2.258 x 108⏎Glass

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. L21:1969 strongly suppresses
  2. L21:1794 strongly suppresses
  3. L25:2029 strongly excites
  4. L25:2947 strongly excites the digit 1 starting a year
  5. L25:2807 excites
  6. L23:903 suppresses
  7. L25:2442 suppresses
  8. L22:83 excites
  9. L23:2992 excites unclear: first pieces of split words
  10. L24:1968 excites
  11. L23:1369 suppresses the '0' in years like 2008
  12. L22:72 suppresses
  13. L24:800 suppresses colon in clock times
  14. L25:2971 excites
  15. L21:2064 excites
  16. L24:2840 excites
  17. L21:563 excites
  18. L25:1702 excites
  19. L22:58 suppresses
  20. L21:1211 excites
  21. L22:1092 excites
  22. L23:2964 suppresses
  23. L18:47 excites
  24. L23:1166 suppresses

Feeds later layers

  1. L27:763 strongly suppresses
  2. L27:785 strongly excites
  3. L27:1087 suppresses
  4. L27:1389 suppresses
  5. L27:1138 weakly suppresses
  6. L27:893 weakly excites
  7. L27:1100 weakly suppresses
  8. L27:572 weakly excites
  9. L27:2123 weakly suppresses
  10. L27:602 weakly suppresses
  11. L27:2182 weakly suppresses
  12. L27:934 weakly suppresses
  13. L27:496 weakly excites
  14. L27:800 weakly excites
  15. L27:1207 weakly suppresses
  16. L27:1738 weakly excites
  17. L27:1904 weakly excites
  18. L27:2918 weakly suppresses
  19. L27:391 weakly excites
  20. L27:1177 weakly excites
  21. L27:1707 weakly suppresses
  22. L27:2958 weakly suppresses
  23. L27:788 weakly suppresses
  24. L27:1713 weakly suppresses

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 -108.50 and -105.581 values between -32.65 and -29.741 values between -23.90 and -20.981 values between -18.07 and -15.151 values between -15.15 and -12.231 values between -12.23 and -9.325 values between -9.32 and -6.406 values between -6.40 and -3.4814 values between -3.48 and -0.563847 values between -0.56 and 2.3529 values between 2.35 and 5.271 values between 5.27 and 8.19offnegativepositiveoften

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. ayah
  2. entially
  3. 隱私權
  4. Player
  5. HEAD
  6. crypt
  7. 裂
  8. elf
  9. seguridad
  10. くれ

Suppresses

  1. 8
  2. eight
  3. 八
  4. eight
  5. Eight
  6. eighty
  7. 8
  8. -eight
  9. 八个
  10. eighth