Neuron Explorer

Qwen3-0.6B

zeros in large round numbers

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 zero digits inside big round numbers like 2,000, 10,000 or 127,000. When it was turned up, the model drifted into long strings of zeros.

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

    curriculum that included linguistics, arts and crafts, medicine, logic, and spirituality. People in India continue to practice two ancient systems developed over 2,000 years ago to maintain health: Ayurvedic medicine, described by the physician

  2. #2 peak

    ancient Indian holistic healing system. The technique has evolved over a period of more than 3,000 years. Contrary to what many people believe in the US, Ayurved

  3. #3 peak

    to humanity. It has been reported that the cultivation of hemp goes back more than 10,000 years having its roots in Asia. The Chinese discovered many uses for hemp including

  4. #4 peak

    PCs running 6th Generation AMD A-Series APU could save 4.9 million kilowatt hours of electricity – about $495,000 – and 3,350 metric tons of Carbon Dioxide over

  5. #5 peak

    ed by the socioeconomic benefits accrued to translocated farmers," they write. "At an average start-up resettlement cost of at least US$12,000 per family, this is an expensive development program."⏎From a macroeconomic standpoint

  6. #6 peak

    that enhance public safety, and provides lifelong learning and technical exchange opportunities benefiting the global engineering and technology community. ASME has more than 127,000 members worldwide.⏎ASME′S VISION, MISSION STATEMENT AND

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: there is a 55.55555555555555555555555555555555555555555

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

    and even life-threatening inflammation of the brain (encephalitis).⏎Treatment --- BEFORE YOU TAKE ANY ACTION ON YOUR OWN YOU MUST SEE aND CONSULT A DOCTOR.-⏎Herpetic gingivostomatitis is the most common specific

  2. #2 peak

    , WIND, (shelter?) Freedom from: weeds, pests, disease.⏎- WATER: tap, butts, drainage (water well weekly)⏎- SOIL type, texture, structure, pH, organic matter content, fertility;⏎-

  3. #3 peak

    Easter itself, the observances properly start on the Friday before Easter (Good Friday). Historically, people ate<|endoftext|>Last modified: 2002-09-28 by rick wyatt⏎Keywords: arapaho |

  4. #4 peak

    them to acquire values that can guide them to make appropriate choices in life (DOE, 2006).⏎The literature also explains values as everything from

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. L19:843 suppresses
  2. L6:2788 excites
  3. L17:213 suppresses
  4. L17:1422 excites zeros in rounded thousands (e.g. 20,000)
  5. L2:1795 suppresses
  6. L13:2379 suppresses the digit zero in numbers
  7. L20:2681 suppresses
  8. L20:626 excites
  9. L19:996 suppresses the digit 5 inside numbers
  10. L16:177 excites zero digit inside dates and codes
  11. L4:498 suppresses
  12. L23:900 excites
  13. L20:217 excites
  14. L18:112 suppresses 'over' meaning time or duration
  15. L6:1322 excites the last digit of decade years
  16. L16:184 weakly suppresses
  17. L18:1155 weakly excites the digit 0 in dates and numbers
  18. L18:778 weakly suppresses
  19. L14:246 weakly suppresses
  20. L20:212 weakly suppresses
  21. L19:1370 weakly excites
  22. L9:977 weakly suppresses the last digit of a decade year
  23. L18:805 weakly excites
  24. L18:127 weakly suppresses

Feeds later layers

  1. L27:153 strongly suppresses the spot just before a historical number
  2. L25:473 strongly excites zeros inside large round numbers
  3. L26:413 strongly excites
  4. L27:862 strongly excites
  5. L27:1153 strongly excites
  6. L25:92 excites
  7. L25:543 suppresses
  8. L27:396 excites
  9. L27:132 excites
  10. L27:1202 excites
  11. L25:1895 excites first digit of a year like 2020
  12. L25:324 suppresses
  13. L26:848 excites
  14. L27:678 weakly suppresses
  15. L27:364 weakly suppresses
  16. L27:391 weakly suppresses
  17. L26:2481 weakly suppresses
  18. L26:682 weakly suppresses
  19. L27:992 weakly suppresses
  20. L27:205 weakly excites
  21. L27:435 weakly suppresses
  22. L25:2312 weakly suppresses
  23. L25:480 weakly suppresses
  24. L26:1554 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.

2 values between -20.03 and -17.761 values between -13.22 and -10.951 values between -10.95 and -8.682 values between -8.68 and -6.414 values between -6.41 and -4.1316 values between -4.13 and -1.863531 values between -1.86 and 0.41287 values between 0.41 and 2.6820 values between 2.68 and 4.959 values between 4.95 and 7.2210 values between 7.22 and 9.494 values between 9.49 and 11.761 values between 11.76 and 14.044 values between 14.04 and 16.312 values between 16.31 and 18.581 values between 18.58 and 20.855 values between 23.12 and 25.392 values between 25.39 and 27.661 values between 27.66 and 29.931 values between 43.56 and 45.831 values between 45.83 and 48.101 values between 54.92 and 57.191 values between 61.73 and 64.001 values between 68.54 and 70.81offnegativepositiveoften

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. 0
  2. 0
  3. ₀
  4. zero
  5. _zero
  6. -zero
  7. zer
  8. 零
  9. zero
  10. Zero

Suppresses

  1. atee
  2. udiante
  3. uxtap
  4. 🏻
  5. 🏽
  6. sheets
  7. uced
  8. eut
  9. chnitt
  10. ushing