Neuron Explorer

Qwen3-0.6B

'and' joining clauses after a comma

A neuron in layer 1 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 word 'and' when it comes right after a comma and joins two full statements, such as 'help benefit individuals, and their families'. The negative side is weak and fires on 'too' after a comma. Turning it up gave wandering text that did not clearly show this pattern.

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

    ressor framework will help guide management of stream fishes, and provide a better understanding of the mechanisms underlying responses of different species.<|endoftext|>The Liber

  2. #2 peak

    new light on our understanding of autism genetics. Importantly, these efforts are already sparking new actions that will help benefit individuals with autism, and their families, in the years to come.⏎April is World Autism Awareness Month,

  3. #3 peak

    physical ticket sales offices. Most states have laws regulating the operation of lotteries, and some even require retailers to be licensed before selling them.⏎The history of the lottery

  4. #4 peak

    test results to identify available treatment options. Artificial intelligence helps fund managers identify opportunities by utilizing data to evaluate statistical models and macroeconomic, and geographical trends.⏎Judges and lawyers can benefit from AI’s ability to analyze information

  5. #5 peak

    Black and white illustrations are essential as a newborn’s retina can only distinguish between light and dark contrasts. Research shows that black and white images promote optic nerve growth, and help the visual part of a baby’s brain to thrive and develop by leaps and

  6. #6 peak

    of time and creating high operations costs. Terraform is an effective solution to this issue, allowing administrators to rapidly deploy and reconfigure infrastructure all from one place, and all at the same time. It is this ability especially that makes Terraform so

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 don't know how to be the right one, but I don't know how to be the right one." — That is a very good word, but I don't know how to be the right one." — That is 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

    baby warm after birth," said a CHW. "Embracing a newborn against the mother's warm skin is so simple, and the mothers like it, too."⏎With nearly nine out of 10 births in Afghanistan still occurring in the

  2. #2 peak

    replied “don’t know.”⏎As the lives of English-language learners change because of attending U.S. schools and interacting with other children in this country, so, too<|endoftext|>There has been a catastrophic event in 2010 in the Gulf

  3. #3 peak

    long shadows produced by the low-angle sun to lead the viewer's eye into the photo.⏎Shadows extending toward the camera can create leading lines, too.⏎Shadows are more than just black blobs that mimic the shapes of things casting

  4. #4 peak

    from proteins, too.⏎In a healthy body, carbs are all broken down into glucose. That glucose

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. L0:27 suppresses
  2. L0:313 weakly suppresses "and" joining paired nouns
  3. L0:1224 weakly suppresses
  4. L0:1301 weakly suppresses
  5. L0:178 weakly suppresses
  6. L0:97 weakly excites
  7. L0:385 weakly excites
  8. L0:251 weakly suppresses
  9. L0:1909 weakly excites
  10. L0:838 weakly suppresses
  11. L0:274 weakly excites
  12. L0:1624 weakly suppresses
  13. L0:1514 weakly suppresses
  14. L0:1259 weakly excites
  15. L0:1560 weakly excites
  16. L0:271 weakly excites
  17. L0:692 weakly suppresses
  18. L0:1335 weakly excites
  19. L0:2921 weakly suppresses
  20. L0:2659 weakly excites
  21. L0:2907 weakly excites
  22. L0:1493 weakly suppresses
  23. L0:386 weakly suppresses
  24. L0:1097 weakly suppresses

Feeds later layers

  1. L6:1349 weakly suppresses
  2. L4:1572 weakly suppresses
  3. L19:851 weakly excites comma inside numbers with thousands
  4. L23:676 weakly suppresses the 'm' in a.m. and p.m
  5. L4:1351 weakly suppresses
  6. L6:798 weakly excites
  7. L6:1978 weakly suppresses
  8. L7:498 weakly suppresses
  9. L4:1642 weakly suppresses
  10. L21:1947 weakly suppresses
  11. L10:2804 weakly excites
  12. L13:609 weakly suppresses
  13. L5:394 weakly suppresses
  14. L23:199 weakly suppresses
  15. L23:993 weakly excites
  16. L24:880 weakly suppresses
  17. L3:621 weakly suppresses
  18. L5:668 weakly suppresses
  19. L23:2871 weakly suppresses
  20. L2:428 weakly suppresses
  21. L20:2717 weakly suppresses
  22. L4:1256 weakly suppresses
  23. L12:1235 weakly suppresses commas inside numbers and lists
  24. L13:1436 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.

6 values between -0.89 and -0.7124 values between -0.71 and -0.5489 values between -0.54 and -0.36150 values between -0.36 and -0.191904 values between -0.19 and -0.011620 values between -0.01 and 0.1675 values between 0.16 and 0.343 values between 0.34 and 0.521 values between 0.52 and 0.694 values between 0.69 and 0.872 values between 0.87 and 1.042 values between 1.04 and 1.222 values between 1.22 and 1.391 values between 1.39 and 1.572 values between 1.57 and 1.741 values between 1.92 and 2.091 values between 2.27 and 2.441 values between 2.79 and 2.973 values between 2.97 and 3.141 values between 3.14 and 3.321 values between 3.32 and 3.493 values between 3.49 and 3.671 values between 3.67 and 3.841 values between 3.84 and 4.021 values between 4.19 and 4.374 values between 4.37 and 4.541 values between 5.07 and 5.252 values between 5.25 and 5.421 values between 5.60 and 5.771 values between 5.95 and 6.12offnegativepositiveoften

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. undler
  2. ersonic
  3. recommend
  4. ировки
  5. …⏎⏎⏎
  6. érique
  7. antino
  8. …⏎⏎
  9. 退税
  10. ATERIAL

Suppresses

  1. казал
  2. ffield
  3. ᴗ
  4. ;"><
  5. IRTH
  6. eresa
  7. 膝
  8. her
  9. irth
  10. �