Neuron Explorer

Qwen3-0.6B

digits in population and money figures

A neuron in layer 12 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 a single digit inside large numbers, mostly city and county population counts such as '80,000' or '50,845'. It goes strongly negative on big dollar amounts in billions or trillions, such as '$20 billion' or '$2.25 trillion'. Turning it up did not give a 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

    the Kurds. The major urban center is Efrîn (‘Afrin in Arabic), with an urban population of approximately 80,000. The city of Efrîn and the surrounding region have a population of

  2. #2 peak

    with Qamişlo (al-Qamishli in Arabic) being the largest city of that region with an urban population of 83,000 (more than 200,000 in the greater Qami

  3. #3 peak

    is a county located in the U.S. state of Texas. As of the 2010 census, its population was 50,845 . The county seat is Rusk. The county was named for the Cherokee,

  4. #4 peak

    in 2002, the 26th largest US city. Its estimated 2002 population was 560,415, but its metropolitan area (including Boulder and Greeley) exceeded 2,

  5. #5 peak

    north of San Francisco and south of the Oregon border. At the 2010 census, the population of the city was 27,191, and the population of Greater Eureka was 45,034

  6. #6 peak

    on foods similar to what might have been eaten during the Paleolithic era, which dates back from about 2.5 million to 10,000 years ago.⏎Food used to be obtained through hunting and gathering. The paleo

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 have been thinking about everything that is in the form of the words that are in the form of the word that is in the form of the word that is in the form of the word that is in the form of the word that

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

    awarded, the Pentagon selected the Lockheed Martin Corp. to build the Joint Strike Fighter (JSF). The contract had a potential total value of over $200 billion during the life of the program. The U.S. planned to buy nearly

  2. #2 peak

    risks climate change poses to federal assets, budgets, and investments. Presumably, this order also applies to the American Jobs Act, Biden’s $2.25 trillion plan to electrify transportation, build millions of energy-efficient homes, and upgrade

  3. #3 peak

    agency announced that it has awarded Elon Musk’s company a contract to build a lunar lander for its Artemis mission.⏎The contract, which is worth $2.9 billion, means that a SpaceX vehicle will be responsible for putting humans back on

  4. #4 peak

    with uncontrolled cell growth. It's one of the most common causes of death in the United States. The financial costs of cancer are estimated to be over $150 billion per year nationally. Thus, the research of cancer has always been

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. L11:2659 suppresses
  2. L11:1552 excites
  3. L11:2510 suppresses
  4. L11:2203 suppresses
  5. L11:1683 suppresses
  6. L9:1222 excites
  7. L5:1885 suppresses
  8. L9:1845 weakly suppresses
  9. L10:3022 weakly suppresses
  10. L11:1637 weakly excites
  11. L5:206 weakly suppresses
  12. L11:1063 weakly suppresses
  13. L11:1698 weakly excites
  14. L3:2866 weakly suppresses
  15. L11:2544 weakly excites
  16. L11:98 weakly suppresses
  17. L9:3037 weakly suppresses
  18. L10:2300 weakly excites
  19. L5:1926 weakly excites
  20. L3:238 weakly excites
  21. L11:1329 weakly suppresses
  22. L8:465 weakly suppresses
  23. L7:2263 weakly excites
  24. L9:1469 weakly suppresses

Feeds later layers

  1. L13:1301 excites
  2. L14:2257 excites
  3. L15:2607 excites
  4. L13:2143 excites
  5. L14:929 excites
  6. L13:1530 excites
  7. L13:2265 suppresses
  8. L13:1386 excites
  9. L20:520 suppresses
  10. L13:2006 excites
  11. L23:2655 suppresses
  12. L13:2726 excites
  13. L13:2314 weakly suppresses
  14. L18:2906 weakly excites
  15. L13:2845 weakly excites
  16. L14:3044 weakly suppresses
  17. L14:1809 weakly excites
  18. L13:3003 weakly excites
  19. L13:2869 weakly excites
  20. L14:253 weakly suppresses
  21. L21:2609 weakly excites
  22. L14:536 weakly suppresses
  23. L22:1985 weakly excites
  24. L24:1454 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 -3.03 and -2.861 values between -2.86 and -2.682 values between -2.51 and -2.331 values between -2.33 and -2.161 values between -2.16 and -1.991 values between -1.81 and -1.641 values between -1.46 and -1.294 values between -1.29 and -1.115 values between -1.11 and -0.946 values between -0.94 and -0.769 values between -0.76 and -0.5921 values between -0.59 and -0.41125 values between -0.41 and -0.24873 values between -0.24 and -0.062275 values between -0.06 and 0.11492 values between 0.11 and 0.2956 values between 0.29 and 0.4611 values between 0.46 and 0.646 values between 0.64 and 0.812 values between 0.81 and 0.993 values between 0.99 and 1.162 values between 1.16 and 1.331 values between 1.51 and 1.682 values between 1.68 and 1.861 values between 1.86 and 2.031 values between 2.03 and 2.212 values between 2.56 and 2.731 values between 2.91 and 3.081 values between 3.26 and 3.431 values between 3.78 and 3.96offnegativepositiveoften

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. 程序
  2. 对她
  3. 交互
  4. 定了
  5. 她的
  6. 小组
  7. UI
  8. elif
  9. elli
  10. iciency

Suppresses

  1. regar
  2. ฏ
  3. Jinping
  4. 扫
  5. 万亿
  6. tot
  7. millions
  8. امة
  9. 大火
  10. .wikipedia