Neuron Explorer

Qwen3-0.6B

"United" in US and UN names

A neuron in layer 24 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 word "United" when it starts names like "United States" or "United Nations". When turned down, the model repeated the word "of" 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

    Naval Facilities Engineering Command (Department of Defense),⏎- Office of Insular Affairs (Department of the Interior),⏎- Office of Naval Intelligence (Department of Defense),⏎- United States Board on

  2. #2 peak

    give it to your baby by mouth, using a medicine dropper, syringe, or open nipple. Treatment begins<|endoftext|>The Government of Indonesia as well as the United Nations have sought support and commitment from parties to the Minamata Convention for a

  3. #3 peak

    Islands, Kiribati, Tuvalu, Nauru, Palau, Federated States of Micronesia, and Republic of the Marshall Islands. The United States builds the capacity and resilience of Solomon Islands to adapt to climate change through regional

  4. #4 peak

    ara, Solomon Islands to provide consular services. The United States and Solomon Islands are committed to working together and improving regional stability, promoting democracy

  5. #5 peak

    areas.⏎The tsunami, which washed away entire communities, caused nearly $10 billion in damage and more casualties than any other tsunami in history, according to the United Nations.⏎Indonesia is on the so-called Ring of Fire, an arc of

  6. #6 peak

    rural and urban poor. Between 1976 and 2001, she worked with the Food and Agriculture Organization (FAO) of the United Nations to identify and support innovative NGO initiatives for development and empowerment of people in South

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: of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of of

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

    the country during various lectures and lessons versus receiving their entire curriculum from local teachers that may or may not have the knowledge to present their content with the most efficiency.⏎Bottom line, if you’re not connected to the Internet, you’re not learning as

  2. #2 peak

    for security from economic gain, and to achieve practical returns. The preferred approach of this dimension is a professional one with a focus on bottom-line results.⏎- The Individualistic Dimension: The Individualistic dimensions deals with one

  3. #3 peak

    comprising 15% of its population. Europe had a total population of nearly 741 million (about 11% of the world population) as of 2016. The European climate is largely affected by hot Atlantic

  4. #4 peak

    President Richard Nixon and Congress responded by creating the Environmental Protection Agency and enacting a wave of laws, including the Clean Air Act, the Clean Water Act and the Endangered Species Act.⏎But Earth Day’s impact extended far beyond the United States.

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. L10:445 weakly excites
  2. L20:1917 weakly suppresses
  3. L4:663 weakly excites
  4. L2:2621 weakly suppresses
  5. L7:2357 weakly excites
  6. L13:1335 weakly suppresses 'of' inside fixed noun phrases
  7. L23:2154 weakly excites
  8. L23:807 weakly suppresses
  9. L7:1092 weakly suppresses
  10. L19:1812 weakly excites
  11. L11:2451 weakly suppresses
  12. L18:569 weakly excites
  13. L19:1234 weakly suppresses
  14. L21:1460 weakly excites
  15. L22:423 weakly suppresses
  16. L11:703 weakly excites
  17. L3:7 weakly excites
  18. L8:641 weakly excites
  19. L11:689 weakly excites
  20. L5:1479 weakly excites
  21. L22:1525 weakly suppresses
  22. L20:2800 weakly suppresses
  23. L20:2599 weakly excites
  24. L17:213 weakly suppresses

Feeds later layers

  1. L26:711 suppresses 'years' after an age number
  2. L27:1962 suppresses
  3. L27:447 suppresses
  4. L25:2020 suppresses
  5. L26:279 weakly excites
  6. L26:920 weakly suppresses
  7. L26:308 weakly suppresses
  8. L26:1274 weakly excites
  9. L25:2769 weakly excites
  10. L26:1094 weakly suppresses
  11. L27:1965 weakly excites
  12. L26:115 weakly excites
  13. L27:1832 weakly excites
  14. L27:893 weakly excites
  15. L27:2218 weakly suppresses
  16. L25:249 weakly suppresses
  17. L25:1710 weakly excites
  18. L25:1421 weakly excites
  19. L25:2029 weakly excites
  20. L25:498 weakly excites
  21. L25:2002 weakly excites
  22. L27:1395 weakly excites
  23. L25:1327 weakly excites
  24. L27:763 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 -21.30 and -20.161 values between -15.59 and -14.452 values between -12.17 and -11.021 values between -11.02 and -9.881 values between -9.88 and -8.743 values between -8.74 and -7.605 values between -7.60 and -6.469 values between -6.46 and -5.325 values between -5.32 and -4.1812 values between -4.18 and -3.0320 values between -3.03 and -1.8975 values between -1.89 and -0.753272 values between -0.75 and 0.39434 values between 0.39 and 1.5326 values between 1.53 and 2.6714 values between 2.67 and 3.817 values between 3.81 and 4.968 values between 4.96 and 6.103 values between 6.10 and 7.241 values between 7.24 and 8.382 values between 8.38 and 9.522 values between 9.52 and 10.663 values between 11.80 and 12.951 values between 23.22 and 24.36offnegativepositiveoften

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. рост
  2. 打ち
  3. Pill
  4. ril
  5. rastructure
  6. utterstock
  7. quets
  8. Stats
  9. ザー
  10. Tits

Suppresses

  1. -of
  2. _of
  3. of
  4. Of
  5. of
  6. _OF
  7. Of
  8. .of
  9. →of
  10. OF