Neuron Explorer

Qwen3-0.6B

in-person versus online delivery

A neuron in layer 24 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the word that sets up a choice between ways of meeting or delivering something, like 'in' in 'in-person', 'online and in-class', or 'telephone or in person'. Turning it up made the model repeatedly write about 'face' and 'in-person', which fits what it reacts to.

Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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 new study, published online in Critical Care Medicine, researchers sought to measure the impact of telemedicine consultations compared to other modes of treatment, such as telephone consultations or treatment without consultations.⏎The study included 320 seriously ill or injured patients

  2. #2 peak

    independent, unbiased emergency medicine physician experts.⏎Overall, cases involving a telemedicine consultation received significantly higher quality-of-care scores than did those involving a telephone consultation or no consultation. In addition, rural emergency room physicians were far more

  3. #3 peak

    , fun online discussions, this site has a community forum all about Mental Health located here -⏎I hope to hear from you sometime soon, either in the forum or in response to this email message. I appreciate your feedback!⏎Have fun passing this message

  4. #4 peak

    - Missing Children’s Citizens Award⏎- Missing Children’s Law Enforcement Award⏎Due to COVID-19 pandemic, this year’s ceremony was unable to take place in-person. To learn more about this year’s awardees and virtual commemoration

  5. #5 peak

    is offered as a semester-long introductory logic course. In addition to working through the online material at a specified rate (approximately one chapter per week), the class meets in small groups once a week for an instructor-led discussion session

  6. #6 peak

    are using their websites to communicate important upcoming events, rules and expectations to parents and their communities.⏎- Communicate regularly with teachers and staff via email, phone or in person to let them know you are involved in your child’s education. By being

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 key to a successful working environment is to let employees have the freedom to work where they want to, work from anywhere, work on any time, work on any size, work on any product or service, and work on any socialization

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

    ers was responsible for tallying the number and different species of birds in the area as part of the Audubon Society’s 121st annual Christmas Bird Count.⏎Since 1900, an annual census of birds in the

  2. #2 peak

    we invented the an air monitoring system we call the "Drift Catcher." It's affordable, robust and based on the same technology used by the California Air Resources Board in its studies.⏎The big idea? Put

  3. #3 peak

    697-707, and was one of the great poets of the early days of Japan; he is known as the rival of Akahito Yamabe (see next verse), and after death was deified as a God of

  4. #4 peak

    mid-May, often coinciding with Global Big Day. For more information about this year's count, contact Suzy Santaella at firstname.lastname@example.org⏎Christmas Bird Count *⏎December 14–January 5. The nation's longest-running

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. L21:1631 suppresses
  2. L22:2742 excites
  3. L21:1089 suppresses
  4. L21:2192 weakly suppresses
  5. L10:2368 weakly suppresses
  6. L18:915 weakly suppresses
  7. L23:2810 weakly suppresses
  8. L9:621 weakly excites
  9. L3:1040 weakly suppresses
  10. L14:1434 weakly suppresses
  11. L14:2834 weakly excites
  12. L17:2650 weakly excites
  13. L15:570 weakly excites
  14. L5:1277 weakly suppresses
  15. L23:1050 weakly excites
  16. L14:918 weakly excites
  17. L11:2531 weakly suppresses
  18. L19:2365 weakly excites
  19. L16:1963 weakly excites
  20. L10:876 weakly excites
  21. L8:2510 weakly suppresses
  22. L19:1897 weakly excites
  23. L12:1861 weakly suppresses
  24. L14:642 weakly suppresses

Feeds later layers

  1. L27:2160 strongly excites
  2. L27:2339 excites
  3. L26:2349 excites
  4. L25:2554 weakly suppresses
  5. L26:207 weakly suppresses
  6. L26:942 weakly suppresses
  7. L25:2608 weakly excites
  8. L27:405 weakly excites
  9. L27:2876 weakly excites
  10. L27:1239 weakly excites
  11. L27:1125 weakly suppresses
  12. L26:2492 weakly excites
  13. L27:911 weakly suppresses
  14. L26:2089 weakly excites
  15. L25:831 weakly suppresses
  16. L27:1215 weakly excites
  17. L25:771 weakly suppresses
  18. L25:1248 weakly suppresses
  19. L25:2031 weakly excites
  20. L26:1390 weakly excites
  21. L25:2588 weakly excites
  22. L25:3041 weakly suppresses
  23. L26:2347 weakly suppresses
  24. L27:1297 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.

4 values between -6.77 and -5.865 values between -5.86 and -4.949 values between -4.94 and -4.032 values between -4.03 and -3.1216 values between -3.12 and -2.2127 values between -2.21 and -1.29580 values between -1.29 and -0.382788 values between -0.38 and 0.53332 values between 0.53 and 1.4555 values between 1.45 and 2.3631 values between 2.36 and 3.2714 values between 3.27 and 4.198 values between 4.19 and 5.108 values between 5.10 and 6.019 values between 6.01 and 6.935 values between 6.93 and 7.841 values between 7.84 and 8.754 values between 8.75 and 9.662 values between 9.66 and 10.581 values between 10.58 and 11.492 values between 11.49 and 12.401 values between 14.23 and 15.141 values between 16.06 and 16.971 values between 23.36 and 24.271 values between 27.01 and 27.921 values between 28.84 and 29.75offnegativepositiveoften

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. face
  2. 面
  3. face
  4. Face
  5. Face
  6. faces
  7. (face
  8. .face
  9. ⾯
  10. _face

Suppresses

  1. [Unit
  2. HeaderValue
  3. Bose
  4. 佑
  5. dition
  6. ression
  7. Giov
  8. lings
  9. unserialize
  10. gression