Neuron Explorer

Qwen3-0.6B

opening bracket after an organisation name

A neuron in layer 4 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 opening bracket that comes right after a full name of an organisation or term, such as 'Food and Drug Administration (' before the short form 'FDA'. Turning it up did not produce a clear theme in the model's replies.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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

    stated SIDS share common challenges such as limited resources, long distances to primary markets, and vulnerability to climate change.⏎Although he confirmed the United Nations Industrial Development Organization (UNIDO) has been implementing various programmes to address some of these limitations - including

  2. #2 peak

    focused on the empowerment of the 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

  3. #3 peak

    the national sample, except that the older boys in this sample were rated as more disturbed by their teachers. Teacher ratings of disturbance were associated with perceived Special Educational Needs (SENs), and it was noted that statutory remedial help was said to be

  4. #4 peak

    , making it the most sweeping overhaul of U.S. food safety laws since 1932. FSMA directs the U.S. Food and Drug Administration (FDA) to build a new system of food safety oversight—and with good reason.

  5. #5 peak

    food security as its theme, with downloadable teaching material.⏎While a person might drink between two and five litres of water a day, the UN Food and Agriculture Organisation (FAO) calculates that the water needed to produce a day’s food can range

  6. #6 peak

    One way copies are made is by simply viewing a page on the Internet. This causes a copy of that page to be made and stored in the Random Access Memory (RAM) of your PC. Browsers also make copies so you can return to

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: I will keep. This is just a dreamer. I am so proud to learn. This is a dreamer. I am. I am. I am. I am. I am. I am. I think I will keep going

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

    much lower (1.0 – 10.3 MPN) than in 1992 (1.2 – 165 MPN), and showed no seasonal pattern from January to August, perhaps due to the relative

  2. #2 peak

    aspiration (TBNA) and allows access to more lymph node stations than does mediastinoscopy. EBUS is generally performed as an outpatient procedure under conscious sedation. It affords excellent sampling precision and minimal complications and has been routinely used at

  3. #3 peak

    from the White House due to his support of the temperance movement, and Lucy followed suit.⏎Lucy Ware Webb met Rutherford B. Hayes at Ohio Wesleyan University when she was a young teenager. Her brothers were studying at the University,

  4. #4 peak

    200km south to Deception Island, it no doubt rested on the short-lived futile attempt spent battling the powerful clockwise-circulating Antarctic Circumpolar Current and winds.⏎Elephant Island is rarely visited by cruise ships as strong currents

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. L1:1777 weakly suppresses
  2. L3:189 weakly excites
  3. L1:2654 weakly excites
  4. L3:256 weakly excites
  5. L1:861 weakly suppresses
  6. L0:363 weakly suppresses
  7. L1:1958 weakly suppresses
  8. L2:889 weakly excites
  9. L1:2308 weakly excites
  10. L1:3045 weakly excites
  11. L3:2797 weakly suppresses
  12. L3:278 weakly excites
  13. L2:1543 weakly suppresses
  14. L0:1636 weakly excites
  15. L0:1159 weakly excites
  16. L1:976 weakly suppresses
  17. L2:190 weakly suppresses
  18. L2:1628 weakly excites
  19. L0:1336 weakly suppresses
  20. L2:429 weakly excites
  21. L3:1389 weakly suppresses 'back' after 'dates' or 'dating'
  22. L3:2150 weakly excites
  23. L0:2483 weakly suppresses
  24. L3:1027 weakly suppresses

Feeds later layers

  1. L20:938 excites
  2. L9:1181 weakly excites
  3. L23:649 weakly suppresses
  4. L9:2898 weakly excites
  5. L11:218 weakly excites
  6. L14:731 weakly suppresses
  7. L15:10 weakly excites
  8. L24:1490 weakly excites
  9. L6:139 weakly excites
  10. L11:1635 weakly excites
  11. L6:804 weakly excites opening bracket before an acronym
  12. L7:438 weakly excites
  13. L6:1989 weakly excites
  14. L14:2498 weakly suppresses
  15. L12:1115 weakly excites
  16. L8:592 weakly excites
  17. L8:293 weakly suppresses
  18. L10:863 weakly suppresses
  19. L15:676 weakly suppresses
  20. L19:1102 weakly suppresses
  21. L6:536 weakly suppresses
  22. L20:1884 weakly suppresses
  23. L6:1478 weakly excites
  24. L6:19 weakly suppresses second half of proper names

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.

3 values between -0.89 and -0.743 values between -0.74 and -0.598 values between -0.59 and -0.4429 values between -0.44 and -0.29340 values between -0.29 and -0.141640 values between -0.14 and 0.011516 values between 0.01 and 0.16302 values between 0.16 and 0.3143 values between 0.31 and 0.467 values between 0.46 and 0.616 values between 0.61 and 0.761 values between 0.76 and 0.911 values between 0.91 and 1.062 values between 1.06 and 1.211 values between 1.21 and 1.361 values between 1.66 and 1.811 values between 1.81 and 1.962 values between 1.96 and 2.111 values between 2.26 and 2.401 values between 4.95 and 5.10offnegativepositiveoften

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. implicitly
  2. ties
  3. conjunction
  4. imply
  5. ᴠ
  6. conclude
  7. nu
  8. cite
  9. 再说
  10. نت

Suppresses

  1. беж
  2. ôme
  3. odon
  4. pré
  5. 落ち
  6. ycastle
  7. ética
  8. 山庄
  9. angered
  10. MLElement