Neuron Explorer

Qwen3-0.6B

middle of fixed names and phrases

A neuron in layer 10 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

The neuron fires inside well-known fixed phrases and names, like 'too good to be true', 'Centers for Disease Control and Prevention' or 'National Aeronautics and Space Administration', where the next word is almost certain. The strongest examples are the middle words of these set phrases. Turning it up did not produce any clear theme in the model's writing.

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

    ask a parent or adult before you fill out any forms, or give out money or credit card details.⏎- Don't accept any offers that seem too good to be true - they probably are.⏎Remember to always be

  2. #2 peak

    when you're online to keep your identity safe.⏎- Only Make Purchases On Trusted Sites⏎When deals seem too good to be true, they just may be - you might be paying as a victim of identity

  3. #3 peak

    GAO found that: (1) DOE plans to continue RTG assembly and testing at its Mound facility through the completion of the National Aeronautics and Space Administration

  4. #4 peak

    than usual and without the energy you normally have during the warm months to get things done.⏎Other symptoms of SAD include the following:⏎Often, overcoming seasonal affective disorder can be done at home. However, if you choose to not give S

  5. #5 peak

    eroding or shaping the landscape,” says Corbett.<|endoftext|>Seasat image of sea ice, 1978.⏎© National Aeronautics & Space Administration / Science & Society⏎This map was prepared using altimeter data from

  6. #6 peak

    laxatives. Some symptoms of anorexia nervosa are:⏎Loss of a period⏎Lightheadedness, dizziness

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: the past is not fully due to the individual, and the ... well, this being proposed ... to the ... well-informed society, it is ... well-instructed ... by the ... well-known ... and the ... well-known ... and

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

    alopathy-PML. Updated January 8, 2016. Accessed September 26, 2016.⏎Progressive multifocal leukoencephalopathy (PML). AETC National Resource Center website

  2. #2 peak

    defense.⏎Then and now:⏎- View a streetscape panel displayed on Roybal campus showing the original and expanded areas of CDC’s campus.<|endoftext|>Woods Hole Oceanographic Institution (WHOI), renowned for its role in the discovery of the Titanic

  3. #3 peak

    , Action, Awareness⏎NORD—National Organization for Rare Disorders⏎Canadian AIDS Society⏎CORD—Canadian Organization for Rare Disorders⏎NINDS progressive multifocal leukoencephalopathy information page. National Institute of Neurological Disorders and Stroke

  4. #4 peak

    this spectacle, drawing on evidence gathered by NASA’s Galaxy Evolution Explorer and the Pan-STARRS1 telescope located in Hawaii. Here’s how NASA describes what you’re seeing

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. L5:297 strongly excites
  2. L0:1548 excites
  3. L8:849 suppresses
  4. L3:607 excites
  5. L2:693 suppresses
  6. L3:1302 excites
  7. L2:1840 excites
  8. L5:67 excites
  9. L1:328 excites
  10. L0:245 excites
  11. L0:1902 excites
  12. L3:521 suppresses
  13. L1:48 excites
  14. L9:746 suppresses
  15. L6:2612 excites
  16. L4:335 suppresses
  17. L0:88 excites
  18. L9:845 excites
  19. L8:1183 excites
  20. L0:1141 excites
  21. L1:2511 excites
  22. L8:2388 suppresses
  23. L9:2153 excites
  24. L9:2656 excites

Feeds later layers

  1. L17:892 suppresses
  2. L13:1513 suppresses
  3. L14:2169 suppresses
  4. L13:2328 excites
  5. L19:2054 suppresses
  6. L11:171 excites
  7. L16:1938 suppresses
  8. L19:1981 suppresses
  9. L16:1129 suppresses
  10. L14:2556 suppresses
  11. L13:1001 excites
  12. L11:122 excites
  13. L20:38 excites
  14. L12:576 suppresses 'of' in fixed phrases like table of contents
  15. L14:1403 suppresses
  16. L16:882 suppresses
  17. L13:450 excites
  18. L14:1090 suppresses
  19. L19:1432 suppresses
  20. L15:1317 suppresses
  21. L14:1641 excites
  22. L16:2166 suppresses
  23. L16:1246 suppresses
  24. L15:1299 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 -4.29 and -4.121 values between -2.01 and -1.833 values between -1.65 and -1.486 values between -1.30 and -1.131 values between -1.13 and -0.953 values between -0.95 and -0.773 values between -0.77 and -0.6011 values between -0.60 and -0.4278 values between -0.42 and -0.25690 values between -0.25 and -0.071735 values between -0.07 and 0.111053 values between 0.11 and 0.28276 values between 0.28 and 0.4636 values between 0.46 and 0.631 values between 0.63 and 0.812 values between 0.81 and 0.991 values between 0.99 and 1.161 values between 1.16 and 1.341 values between 1.51 and 1.691 values between 2.04 and 2.222 values between 2.39 and 2.571 values between 2.57 and 2.74offnegativepositiveoften

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. enerate
  2. lian
  3. [,]
  4. ReadStream
  5. enerative
  6. 馨
  7. LJ
  8. rarian
  9. istency
  10. [^

Suppresses

  1. >>&
  2. zeichnet
  3. usted
  4. 很好地
  5. ECH
  6. .docs
  7. för
  8. "..
  9. 主场
  10. 者