Neuron Explorer

Qwen3-0.6B

the digit 9 in years 1900s

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 '9' in years written like 1960, 1947 or 1990, where the year is split so the model has to guess the next digits. Its opposite side is weak and unrelated. Turning it up gave no clear theme in the output.

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

    . military to experiment with containers. First used by military traffic commands, containers were extended to commercial use in the mid-1960s. Container use has continued to grow steadily in the United States and

  2. #2 peak

    Mountain states acting as a neutral buffer zone between the two.⏎Set in the early 1960s, fifteen years after the Axis victory in 1947, much of the book takes place in San Francisco where we see Robert

  3. #3 peak

    lags behind the industry high of 55 million boxes, reached in the mid 1990's. However, according to Ocean Spray's Grapefruit Juice Marketing Manager

  4. #4 peak

    instruction. In the mid-1930s the central government tried to culturally assimilate the local population through the

  5. #5 peak

    00,000 are ethnic Tibetans. In a survey carried out in Lhasa in July 1993 on the southern Lingkor, a street parallel to the Barkhor,

  6. #6 peak

    CIA⏎The US Congress has had oversight responsibilities over the CIA since the Agency was established in 1947. However, prior to the mid-1970s, oversight responsibilities resided in the Armed Services Committees of both chambers

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: a company that a person has had for a company. The building that a person has had for a company. The building that a person has had for a company. The building that a person has had for a company. The building that a

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

    N.F.) is the third supporting idea. * Amerindians forced to do unpaid labor to build roads, buildings, and other infrastructure in towns through the Encomienda System (S.W.) * The N.E. Confederation made specifically for

  2. #2 peak

    age was a field hand on an Alabama cotton plantation. Four times he attempted escape and finally, on his fifth try in 1

  3. #3 peak

    River.⏎It should surprise no one that women played an important role in the industrialization of the United States. Homage to Women celebrates their role in the textile mills that were built along the Merrimack River in Lowell. Kaufman had this

  4. #4 peak

    early 17th c. War resulted in a boundary being defined between the Indians and English.⏎The Spanish Southwest, New York and New France: the Encomienda

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. L3:631 excites
  2. L3:960 excites
  3. L3:142 suppresses
  4. L3:2628 weakly excites
  5. L1:1142 weakly suppresses
  6. L1:1229 weakly excites
  7. L1:1387 weakly suppresses
  8. L3:108 weakly suppresses
  9. L2:2479 weakly suppresses
  10. L1:709 weakly suppresses
  11. L3:1588 weakly excites
  12. L3:115 weakly suppresses
  13. L2:1957 weakly excites
  14. L0:2743 weakly suppresses
  15. L3:1864 weakly suppresses
  16. L0:2042 weakly suppresses
  17. L0:1731 weakly suppresses
  18. L3:408 weakly suppresses
  19. L2:1575 weakly suppresses
  20. L1:955 weakly excites
  21. L3:2684 weakly suppresses
  22. L3:620 weakly excites
  23. L3:1474 weakly excites
  24. L1:2593 weakly excites

Feeds later layers

  1. L18:991 weakly excites
  2. L5:2414 weakly suppresses
  3. L5:267 weakly excites
  4. L18:2807 weakly suppresses
  5. L17:560 weakly excites
  6. L18:265 weakly excites first half of institution names
  7. L8:394 weakly excites
  8. L16:1612 weakly suppresses
  9. L6:1405 weakly suppresses
  10. L18:2756 weakly suppresses
  11. L16:211 weakly excites
  12. L5:1620 weakly suppresses
  13. L21:2931 weakly suppresses
  14. L6:509 weakly suppresses
  15. L23:1023 weakly suppresses
  16. L20:1210 weakly suppresses
  17. L16:256 weakly excites number right after 'age of' or 'ages'
  18. L10:1091 weakly excites
  19. L6:579 weakly excites
  20. L8:457 weakly excites year digits in copyright and dates
  21. L15:42 weakly excites
  22. L12:2733 weakly suppresses
  23. L10:1083 weakly suppresses
  24. L10:220 weakly excites hyphen inside compound words

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 -2.57 and -2.292 values between -2.29 and -2.002 values between -1.71 and -1.433 values between -1.43 and -1.147 values between -1.14 and -0.8513 values between -0.85 and -0.5786 values between -0.57 and -0.282042 values between -0.28 and 0.011542 values between 0.01 and 0.29116 values between 0.29 and 0.5832 values between 0.58 and 0.8721 values between 0.87 and 1.157 values between 1.15 and 1.446 values between 1.44 and 1.734 values between 1.73 and 2.014 values between 2.01 and 2.303 values between 2.30 and 2.593 values between 2.87 and 3.163 values between 3.16 and 3.451 values between 3.45 and 3.731 values between 4.02 and 4.311 values between 4.31 and 4.591 values between 5.45 and 5.741 values between 5.74 and 6.032 values between 7.46 and 7.742 values between 8.03 and 8.321 values between 8.60 and 8.89offnegativepositiveoften

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. writable
  2. SYNC
  3. 螵
  4. 起飞
  5. odox
  6. ゃ
  7. TYPO
  8. ">{{
  9. ikki
  10. وى

Suppresses

  1. ements
  2. quil
  3. unnel
  4. Answers
  5. ein
  6. Answer
  7. دول
  8. 天河
  9. ifi
  10. 槎