Neuron Explorer

Qwen3-0.6B

the '9' in years 19xx

A neuron in layer 23 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 fires on the digit 9 in years that start with 19, such as the 9 in 1914 or 1943. Turning it down made the model repeat the word 'divide' over and over.

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

    galvanized by the failure of the international community to prosecute Turkish officials involved in the genocide of approximately one million Armenians between 1914 and 1918.⏎No doubt Lemkin would have been

  2. #2 peak

    XV1 condemned "the genocide of the Jews" during the Jewish Holocaust when six to eight million Jews were killed by a European named Adolf Hitler in the 1930s. However

  3. #3 peak

    <|endoftext|>The Battle of the Kasserine Pass was a battle that took place during the Tunisia Campaign of World War II in February of 1943. It was a series of battles fought around Kasserine Pass,

  4. #4 peak

    laws and regulations imposed by the Nazis and their collaborators to facilitate the annihilation of the Jews.⏎Lemkin’s campaign was crowned with success in December 1948, when the United Nations adopted the Genocide Convention. It defined genocide

  5. #5 peak

    Brittany Hutchinson, assistant curator at the Chicago History Museum. “They did this in 1918.”⏎The Article Describes The Pandemic of 1918

  6. #6 peak

    Raphael Lemkin, who coined that term seventy years ago this autumn, would have been appalled by such abuse of his life’s work.⏎During the 1930s, Lemkin, a Polish Jewish attorney and expert on the development

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: divide the divide, divide the divide, divide the divide, divide the divide... divide the divide divide the divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide divide

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

    - Undesirable things.⏎- Confrontational - Arguments / Dispute.⏎- Dissent - Strong disagreement or dissatisfaction.⏎- Fissures - Split or deep crack.⏎- Condemnation - Act of saying that something is unacceptable.⏎-

  2. #2 peak

    inequalities that exist offline. So it’s no surprise that the majority of unconnected people are women and girls.⏎Connecting these women and girls is critical to closing the digital divide, and essential for boosting economic growth. Internet access allows women to gain a

  3. #3 peak

    and within different regions and different countries when it comes to accessing science, technology and innovation (STI) and enjoying their benefits. To address these disparities and close the existing STI gaps, Open Science is an important step in the right direction.⏎

  4. #4 peak

    /⏎They also say…. ‘ Students with special learning needs may offer the most concrete evidence of the benefits of tablets in the classroom. Advocates, desperate to bridge the educational gaps in children with learning issues

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. L20:2383 weakly suppresses
  2. L14:2612 weakly suppresses
  3. L17:1838 weakly suppresses
  4. L19:2904 weakly excites
  5. L17:1989 weakly suppresses
  6. L12:2834 weakly excites
  7. L5:504 weakly excites
  8. L14:2037 weakly excites
  9. L22:2304 weakly suppresses
  10. L6:2920 weakly excites
  11. L18:1138 weakly excites
  12. L22:2975 weakly suppresses
  13. L12:2852 weakly excites
  14. L15:1337 weakly suppresses
  15. L20:690 weakly excites
  16. L7:2883 weakly excites
  17. L11:2176 weakly excites
  18. L20:570 weakly excites
  19. L7:538 weakly excites
  20. L2:1920 weakly suppresses
  21. L12:2402 weakly suppresses
  22. L4:1382 weakly suppresses
  23. L5:1173 weakly suppresses
  24. L22:1197 weakly excites

Feeds later layers

  1. L24:1232 suppresses
  2. L26:2421 suppresses
  3. L27:1311 suppresses
  4. L27:2099 suppresses
  5. L24:1807 suppresses
  6. L24:2434 suppresses
  7. L26:770 weakly suppresses
  8. L24:1329 weakly suppresses
  9. L25:1452 weakly excites
  10. L25:1884 weakly excites
  11. L26:724 weakly suppresses
  12. L27:296 weakly suppresses
  13. L24:1372 weakly suppresses
  14. L26:1517 weakly suppresses
  15. L25:2453 weakly suppresses
  16. L24:2697 weakly suppresses
  17. L26:1100 weakly suppresses
  18. L27:874 weakly suppresses
  19. L24:621 weakly suppresses
  20. L24:1822 weakly suppresses
  21. L26:2375 weakly excites
  22. L24:1349 weakly suppresses
  23. L26:1468 weakly excites
  24. L27:2577 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 -17.94 and -17.131 values between -14.71 and -13.901 values between -13.09 and -12.281 values between -12.28 and -11.482 values between -11.48 and -10.671 values between -10.67 and -9.864 values between -9.86 and -9.052 values between -9.05 and -8.251 values between -8.25 and -7.443 values between -7.44 and -6.634 values between -6.63 and -5.826 values between -5.82 and -5.029 values between -5.02 and -4.2114 values between -4.21 and -3.4037 values between -3.40 and -2.5947 values between -2.59 and -1.79121 values between -1.79 and -0.981018 values between -0.98 and -0.172340 values between -0.17 and 0.64203 values between 0.64 and 1.4534 values between 1.45 and 2.2517 values between 2.25 and 3.0610 values between 3.06 and 3.878 values between 3.87 and 4.683 values between 4.68 and 5.484 values between 5.48 and 6.293 values between 6.29 and 7.103 values between 7.10 and 7.913 values between 7.91 and 8.712 values between 8.71 and 9.521 values between 9.52 and 10.331 values between 10.33 and 11.141 values between 11.14 and 11.942 values between 13.56 and 14.37offnegativepositiveoften

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. uada
  2. edly
  3. pwd
  4. ildo
  5. đỡ
  6. ila
  7. .badlogic
  8. мом
  9. lint
  10. iant

Suppresses

  1. divide
  2. divides
  3. dividing
  4. division
  5. divide
  6. Divide
  7. divisions
  8. divided
  9. splits
  10. 分