Neuron Explorer

Qwen3-0.6B

empty slot before a decade

A neuron in layer 21 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 blank spot right after 'early' in dates like 'the early 1970s', where the next thing is a decade number. Its negative side fires on colons and speaker labels such as 'says:' or 'A:'. Turning it up gave some unrelated text, so the steering does not clearly match the trigger.

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

    the Saffir-Simpson Hurricane scale, developed as a classification system for Western Hemisphere tropical cyclones in the late 1960s and early 70s by structural engineer Herbert Saffir and his friend, meteorologist

  2. #2 peak

    frames and technological implications’ organized by the Higher School of Economics on the 7th Internet Governance Forum (Baku, Azerbaijan, 6-9 November, 2012).⏎The emergence of the Internet is a turning point in human

  3. #3 peak

    warmer history. These insights into the Martian past demonstrated the science potential of investing in imaging of the planets from orbit. By the end of their missions in the early 1980s, the Viking landers and orbiters had accumulated over

  4. #4 peak

    company also approached university bookstores. Its Ski and Hike day pack sold like hotcakes. Other college bookstores began selling it as well.⏎In the early 1970s, JanSport teamed up with a

  5. #5 peak

    . This measurement station is our ‘national atmospheric observatory’ and is celebrating its 50th anniversary this year. KNMI built the tower in the early 1970s at the centre of the Amsterdam-Utrecht-Rotterdam

  6. #6 peak

    to the region in late 19th century. The site was excavated by the American Foundation for the Study of Man (AFSM) in the early 1950s and by the Italian Mission to Oman (IMTO)

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: every time some new business each business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every business every

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

    good health can help to lower your premiums - another reason to reach for the nutritious stuff.⏎Expert dietitian and nutritionist, Annemarie Aburrow says: "Basically the more you eat the better. The five-a-day message was chosen

  2. #2 peak

    - More about:⏎- First World War⏎- Imperial War Museum⏎- Lipstick And Other Lip Makeup<|endoftext|>Q: Why is it important to recycle?⏎A: With the involvement and enthusiasm of people like you, recycling is back and so are

  3. #3 peak

    a religious or educational one.⏎D<|endoftext|>Administration of Justice 104:⏎Unit Objectives: At the completion of

  4. #4 peak

    it inland, with potentially devastating effects to marshlands and local wildlife.⏎Dr. Piers Champan, Head of the Oceanography Department at Texas A&M University says that if a storm surge were to move inland, the oil would be left behind

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. L18:1808 suppresses
  2. L17:2101 excites
  3. L20:1832 excites 'as' introducing a reason
  4. L17:2684 weakly suppresses
  5. L19:1197 weakly suppresses
  6. L19:1580 weakly excites
  7. L12:1934 weakly suppresses
  8. L17:689 weakly excites
  9. L19:1972 weakly suppresses
  10. L8:475 weakly suppresses
  11. L19:917 weakly excites
  12. L19:2058 weakly excites
  13. L20:1959 weakly suppresses
  14. L17:1420 weakly excites
  15. L17:2397 weakly excites
  16. L18:1519 weakly excites
  17. L18:229 weakly suppresses
  18. L20:850 weakly excites
  19. L14:1876 weakly excites
  20. L18:2434 weakly excites
  21. L18:227 weakly suppresses
  22. L14:1128 weakly excites
  23. L3:144 weakly excites
  24. L14:978 weakly excites

Feeds later layers

  1. L27:609 weakly suppresses
  2. L25:454 weakly suppresses
  3. L26:715 weakly excites
  4. L24:545 weakly excites
  5. L27:460 weakly suppresses
  6. L24:2884 weakly excites
  7. L27:2203 weakly excites
  8. L25:2287 weakly excites
  9. L23:24 weakly excites
  10. L25:1260 weakly suppresses
  11. L27:789 weakly suppresses
  12. L23:665 weakly excites
  13. L22:892 weakly excites
  14. L23:96 weakly excites
  15. L25:1586 weakly excites
  16. L22:835 weakly suppresses words before water, lava or magnetic fields
  17. L24:1439 weakly excites
  18. L24:1952 weakly excites
  19. L22:626 weakly excites
  20. L22:2726 weakly excites
  21. L25:412 weakly excites
  22. L23:988 weakly excites
  23. L24:2764 weakly excites
  24. L27:685 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 -20.64 and -19.832 values between -17.39 and -16.571 values between -15.76 and -14.953 values between -14.95 and -14.134 values between -14.13 and -13.322 values between -13.32 and -12.516 values between -12.51 and -11.704 values between -11.70 and -10.885 values between -10.88 and -10.0714 values between -10.07 and -9.2612 values between -9.26 and -8.4413 values between -8.44 and -7.6319 values between -7.63 and -6.8215 values between -6.82 and -6.0019 values between -6.00 and -5.1919 values between -5.19 and -4.3828 values between -4.38 and -3.5634 values between -3.56 and -2.7552 values between -2.75 and -1.9479 values between -1.94 and -1.12524 values between -1.12 and -0.312832 values between -0.31 and 0.51132 values between 0.51 and 1.3242 values between 1.32 and 2.1316 values between 2.13 and 2.9511 values between 2.95 and 3.767 values between 3.76 and 4.572 values between 4.57 and 5.384 values between 5.38 and 6.201 values between 6.20 and 7.012 values between 7.82 and 8.642 values between 10.26 and 11.081 values between 11.08 and 11.89offnegativepositiveoften

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. 请您
  2. we
  3. wich
  4. 我们必须
  5. *\
  6. STOP
  7. *",
  8. →sh
  9. تي
  10. sh

Suppresses

  1. cers
  2. 每个人的
  3. 每个人
  4. 不同的
  5. 适当的
  6. 不同
  7. 同一个
  8. 每个
  9. niekt
  10. veys