Neuron Explorer

Qwen3-0.6B

'with' after exchange nouns

A neuron in layer 9 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 'with' when it follows a noun about exchanging or relating to someone, such as letters with a friend, talks with a leader, or an agreement with a group. When it was turned up, the model's replies did not show a clear change in theme.

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

    correspondence, one of the best- preserved collections of letters from the period, shows that he exchanged letters constantly with his friends elsewhere, keeping them up to date with the latest political machinations,

  2. #2 peak

    your choice, fill out an order form for merchandise as if you⏎intended to place an order. Share the completed form with your counselor and discuss⏎- With your parent's permission, locate at least five Web sites that are helpful for⏎

  3. #3 peak

    delegated to the United States, in Congress assembled.⏎Provide for common defence was:⏎Article III. The said States hereby severally enter into a firm league of friendship with each other, for their common defense, the security of their liberties, and their

  4. #4 peak

    -known material from the UK National Archives. Includes James' own description of the Battle of Edgehill, his reasons for his conversion to Catholicism, and his correspondence with William of Orange.⏎A Court in Exile: The Stuarts in France

  5. #5 peak

    breed in aquarium environments.<|endoftext|>Donald Trump became the first serving American president to meet with a North Korean leader in June 2018.⏎Trump’s talks with North Korean Leader Kim Jong Un in Singapore led to an agreement in which the two

  6. #6 peak

    region and own and manage the remaining 2.7% of the Watershed.⏎The Yukon has several First Nation Final Agreements (“Final Agreements”) with various Yukon First Nations. These Final Agreements adopt terms from what is called

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 it, and I have been thinking about it, and I have been thinking about it, and I have been thinking about it, and I have been thinking about it, and I have been thinking about it, and

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

    patients. Based at University of Chicago, they hypothesize that it’s the ‘abnormal’ gut microbes, those that survive the extreme amount of antibiotics thrown at these patients that eventually turn deadly.⏎“Nobody had asked what happens to bacteria in

  2. #2 peak

    to aim your stream of water directly onto the soil. Watering at the top of the plant will not get the roots

  3. #3 peak

    or dead.⏎When the king killed Arbsima, the virgin, and all the virgins who were with her, he also ordered that their bodies be cast away on the mountains. Later

  4. #4 peak

    which can eventually lead to a heart attack. The purpose of an angioplasty is to clear plaque and cholesterol build up in the coronary arteries to improve blood flow to the heart.⏎The procedure for performing an angioplasty involves the use of a

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. L5:910 excites
  2. L8:1862 suppresses
  3. L8:1627 excites
  4. L6:2842 excites
  5. L4:1330 suppresses
  6. L6:800 excites
  7. L7:2766 suppresses
  8. L8:45 suppresses
  9. L6:1477 suppresses
  10. L4:328 suppresses 'to' after 'so as' and infinitive cues
  11. L6:1828 suppresses
  12. L5:1624 excites 'from' after 'Retrieved' and source phrases
  13. L3:2669 excites
  14. L5:2877 excites
  15. L8:2632 suppresses
  16. L6:1534 suppresses
  17. L5:1535 suppresses
  18. L4:882 weakly suppresses 'with' after provide or familiarize
  19. L6:1956 weakly suppresses "for" introducing a purpose or source
  20. L7:1959 weakly excites
  21. L6:742 weakly excites
  22. L2:863 weakly suppresses
  23. L7:2748 weakly suppresses
  24. L8:216 weakly excites

Feeds later layers

  1. L13:2116 excites
  2. L18:1857 excites
  3. L15:2247 excites
  4. L10:297 excites
  5. L16:1999 excites
  6. L10:2278 excites
  7. L11:437 excites
  8. L19:1434 excites
  9. L11:881 excites "with" after combine, mix or react verbs
  10. L20:1794 excites
  11. L11:2292 excites 'than' in comparisons
  12. L12:2014 excites
  13. L15:799 excites
  14. L16:1469 excites
  15. L16:490 excites
  16. L23:316 excites
  17. L17:2197 excites
  18. L19:1024 excites
  19. L17:2591 excites
  20. L16:564 excites
  21. L14:2508 excites
  22. L18:56 excites
  23. L15:1305 excites
  24. L19:2845 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.

12 values between -0.82 and -0.6273 values between -0.62 and -0.42738 values between -0.42 and -0.222009 values between -0.22 and -0.03937 values between -0.03 and 0.1776 values between 0.17 and 0.3716 values between 0.37 and 0.577 values between 0.57 and 0.778 values between 0.77 and 0.974 values between 0.97 and 1.166 values between 1.16 and 1.361 values between 1.36 and 1.563 values between 1.56 and 1.764 values between 1.96 and 2.151 values between 2.35 and 2.552 values between 2.55 and 2.752 values between 2.75 and 2.952 values between 3.34 and 3.542 values between 3.54 and 3.741 values between 3.74 and 3.943 values between 4.14 and 4.331 values between 6.91 and 7.11offnegativepositiveoften

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. gegenüber
  2. 朋友
  3. entifier
  4. peers
  5. 当地人
  6. 别人的
  7. 队友
  8. 对方
  9. 当地政府
  10. 别人

Suppresses

  1. ecure
  2. case
  3. Ping
  4. casts
  5. ATCH
  6. ews
  7. 싱
  8. Ping
  9. ело
  10. esthesia