Neuron Explorer

Qwen3-0.6B

'in' in set phrases like 'in mind'

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

Fires on the word 'in' when it sits inside a fixed phrase like 'in mind', 'in place' or 'in the forefront'. Turning it down made the model repeat the word 'presence' again and again, though the link to these phrases is only loose.

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

    't call it a sustainable fuel source unless we harvest forests in a sustainable way. It's important for any harvesting of the forest to keep the big picture of sustainability in mind. At what level can we remove wood that's going to be sustainable into

  2. #2 peak

    of Hugh Redwood: 'The true imperial⏎import of Britain's world mission was the spread of God's eternal word.'"⏎"Here is a<|endoftext|>Head Injuries can occur in any accident⏎They are often the result of a fall at

  3. #3 peak

    a secret code that can only be understood by authorised individuals. This helps prevent unauthorised access and keeps your data secure.⏎Cloud service providers have robust security measures in place to protect the physical infrastructure where your data is stored. They employ advanced security

  4. #4 peak

    The⏎Halacha section will primarily be a review of the lessons from the previous⏎cycles while the Aggadah section will be new material. Please bear in mind⏎that the purpose of learning the laws of Shmiras haloshon

  5. #5 peak

    basic concept of an operating system is the process. If we think of the program as the file stored on the hard disk or floppy and the process as that program in memory, we can better understand the difference between a program and a process. Although

  6. #6 peak

    of 2001 and the high school exit examinations in use in at least 23 of the 50 states, high-stakes testing is in the forefront of the subject matter taught in schools. Whether or not these institutions have

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: the presence of a presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in the presence in

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

    works, you are close to understanding how allergy shots work! A small amount of your allergen is injected in slowly increasing amounts, which will help you to develop a tolerance or immunity to the

  2. #2 peak

    result, ER-positive tumors can be treated in part with hormone-blocking therapies to restrict their growth. ER-negative tumors are considered harder to treat since they will grow even in the absence of estrogen.⏎Women in the U.S. have around a 1

  3. #3 peak

    ema. Immunotherapy can also be used to treat stinging insect allergies. Small amounts of the allergen (the substance that patients are allergic to) are injected in regular intervals over several years to desensitize patients to those substances. Immunotherapy

  4. #4 peak

    forms of fuel cells and technology used in the production of fuel cells. A very simplistic description of the technology is that a fuel, typically hydrogen, is oxidized in the first segment of the fuel cell to create positively charged ions and

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. L21:1935 suppresses
  2. L20:1574 excites
  3. L4:289 excites
  4. L17:706 excites
  5. L22:535 excites
  6. L18:869 suppresses
  7. L5:490 excites
  8. L8:351 suppresses 'in' after 'high' or 'dressed'
  9. L14:1625 excites the word 'in' after a noun
  10. L10:262 suppresses
  11. L16:1557 excites 'in' meaning 'in this case' or 'in between'
  12. L19:704 excites
  13. L20:1905 excites
  14. L9:1025 weakly excites
  15. L16:1807 weakly suppresses 'in' after a motion verb
  16. L6:2417 weakly suppresses
  17. L23:1366 weakly excites
  18. L6:846 weakly suppresses
  19. L19:1986 weakly excites 'into' after a verb or 'well'
  20. L14:1122 weakly suppresses
  21. L16:2120 weakly suppresses
  22. L20:2850 weakly excites
  23. L19:613 weakly excites
  24. L18:2356 weakly excites

Feeds later layers

  1. L27:2462 suppresses
  2. L26:2278 weakly suppresses
  3. L26:207 weakly excites
  4. L25:658 weakly suppresses
  5. L26:2282 weakly suppresses
  6. L26:1299 weakly suppresses
  7. L25:1345 weakly excites
  8. L27:1909 weakly excites
  9. L26:2165 weakly suppresses
  10. L26:1089 weakly suppresses
  11. L25:184 weakly excites
  12. L25:1156 weakly excites
  13. L27:1179 weakly suppresses
  14. L27:405 weakly suppresses
  15. L25:2028 weakly suppresses
  16. L25:1987 weakly suppresses
  17. L26:2782 weakly excites
  18. L27:2866 weakly suppresses
  19. L26:1443 weakly suppresses
  20. L26:772 weakly excites
  21. L27:1229 weakly excites
  22. L27:1493 weakly suppresses
  23. L27:2228 weakly suppresses
  24. L25:2325 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.

2 values between -22.23 and -20.811 values between -20.81 and -19.381 values between -19.38 and -17.951 values between -17.95 and -16.522 values between -13.66 and -12.244 values between -12.24 and -10.814 values between -10.81 and -9.383 values between -9.38 and -7.958 values between -7.95 and -6.5210 values between -6.52 and -5.099 values between -5.09 and -3.6620 values between -3.66 and -2.24104 values between -2.24 and -0.813471 values between -0.81 and 0.62200 values between 0.62 and 2.0526 values between 2.05 and 3.4813 values between 3.48 and 4.9112 values between 4.91 and 6.346 values between 6.34 and 7.765 values between 9.19 and 10.622 values between 10.62 and 12.052 values between 13.48 and 14.911 values between 14.91 and 16.341 values between 33.48 and 34.91offnegativepositiveoften

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. uste
  2. insic
  3. apesh
  4. TOCOL
  5. влад
  6. 绩
  7. erto
  8. 又被
  9. LOOR
  10. karış

Suppresses

  1. presence
  2. Presence
  3. présence
  4. presence
  5. presença
  6. Presence
  7. 存在
  8. exists
  9. 是否存在
  10. company