Neuron Explorer

Qwen3-0.6B

blank before a number word

A neuron in layer 23 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 an empty gap just before a number, such as "for [ ]40 years" or "the last [ ]16 years", where the missing word is a count. Turning it up made the model write long strings of zeros and percentages.

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

    sound of the Antarctic break. Their work makes use of data gathered by University of Chicago glaciologist Douglas MacAyeal, who has studied Antarctic ice for 40 years. MacAyeal records seismographic data of glaciers to

  2. #2 peak

    China had managed to sustain such high long-term growth and transformation. Frankly, I was stumped. I have spent over four decades in Asia, the last 16 years living in China, and no one had ever asked me that question

  3. #3 peak

    grew up and his parents still own a ranch. They recently had to sell half their herd when their pasture couldn’t provide enough feed.⏎“They’ve spent the last 15 years grooming that herd for organic beef stock,” he said.⏎Bled

  4. #4 peak

    tailor their research to meet the needs of field-based researchers and conservation practitioners.”⏎With funding from NASA to lead the study, Rose and his co-authors brought together 32 thought leaders from both the conservation and remote-sensing communities. The participants

  5. #5 peak

    0 families as research partners to accelerate the scientific understanding of autism.⏎I’m a co-leader of the OHSU SPARK site, now one of just

  6. #6 peak

    6 percent). Only one state, New Hampshire, had a poverty rate of less than ten percent for women, at 8.9 percent. In the other 47 states and the District of Columbia, between 10 and 2

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 start a new college college right after college beginning into a new university where I begin two weeks into a semester right beginning right beginning. I start a new two weeks into a semester and begin a new college right on campus. I begin 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

    ⏎- other public and private sources.⏎The official seal of the CIA, however, may NOT be copied without permission as required by the CIA Act of 1949 (50 U.S.C. section 403m).

  2. #2 peak

    the elderly. On the same day, the Consumer Financial Protection Bureau (CFPB) released a study reviewing financial education research priorities for America’s youth.⏎Congressmen Bill Foster (D-IL) and French Hill (R-AR) both introduced

  3. #3 peak

    at the Capitol Hill exhibition, we’re coming across new questions as fast as we can answer the original ones. There’s a long way to go yet.⏎Senator Bill Nelson (D-FL), in his keynote address, highlighted the crucial importance of

  4. #4 peak

    Pyongyang Times newspaper are on display. One reports on warming relations under North Korea’s late leader Kim Jong Il and U.S. President Bill Clinton during the 1990s. A more recent issue of the newspaper shows a photo of Trump

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. L21:2688 excites
  2. L19:710 excites
  3. L20:2496 excites
  4. L10:716 weakly suppresses
  5. L18:122 weakly excites
  6. L19:885 weakly excites
  7. L8:1169 weakly excites
  8. L10:2374 weakly excites
  9. L21:134 weakly excites
  10. L16:3021 weakly suppresses
  11. L10:333 weakly excites
  12. L10:1060 weakly suppresses
  13. L12:2323 weakly suppresses
  14. L1:1552 weakly excites
  15. L1:2813 weakly suppresses
  16. L15:2600 weakly suppresses
  17. L4:1311 weakly suppresses
  18. L16:1156 weakly suppresses
  19. L9:1887 weakly excites
  20. L22:318 weakly suppresses
  21. L14:2406 weakly excites
  22. L12:1979 weakly excites
  23. L0:1286 weakly suppresses
  24. L19:1568 weakly excites

Feeds later layers

  1. L25:2915 suppresses
  2. L24:126 weakly excites
  3. L24:2842 weakly suppresses
  4. L27:1894 weakly suppresses
  5. L25:3002 weakly suppresses
  6. L24:2519 weakly suppresses
  7. L27:1090 weakly excites
  8. L26:1574 weakly excites
  9. L24:2948 weakly suppresses
  10. L24:198 weakly excites
  11. L24:2512 weakly excites
  12. L25:1155 weakly suppresses
  13. L26:239 weakly excites
  14. L26:1852 weakly suppresses
  15. L24:2441 weakly excites
  16. L26:2653 weakly suppresses
  17. L27:287 weakly excites
  18. L24:2326 weakly suppresses
  19. L24:20 weakly excites
  20. L24:849 weakly suppresses
  21. L27:627 weakly excites
  22. L25:1628 weakly excites
  23. L24:2823 weakly suppresses
  24. L27:1651 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.

1 values between -17.06 and -15.231 values between -15.23 and -13.394 values between -11.55 and -9.717 values between -9.71 and -7.889 values between -7.88 and -6.0415 values between -6.04 and -4.2049 values between -4.20 and -2.36264 values between -2.36 and -0.533431 values between -0.53 and 1.3160 values between 1.31 and 3.1524 values between 3.15 and 4.996 values between 4.99 and 6.837 values between 6.83 and 8.661 values between 8.66 and 10.505 values between 10.50 and 12.341 values between 12.34 and 14.181 values between 14.18 and 16.011 values between 16.01 and 17.852 values between 17.85 and 19.693 values between 19.69 and 21.531 values between 21.53 and 23.362 values between 23.36 and 25.202 values between 25.20 and 27.043 values between 28.88 and 30.711 values between 30.71 and 32.552 values between 32.55 and 34.391 values between 38.06 and 39.901 values between 45.41 and 47.251 values between 49.09 and 50.931 values between 52.76 and 54.601 values between 54.60 and 56.44offnegativepositiveoften

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. ocate
  2. яз
  3. Culture
  4. ConnectionFactory
  5. bound
  6. Found
  7. oose
  8. yclopedia
  9. ellas
  10. ington

Suppresses

  1. subcategory
  2. ministers
  3. Thom
  4. isque
  5. ade
  6. ]+'
  7. .getServer
  8. adem
  9. 沐
  10. apartments