Neuron Explorer

Qwen3-0.6B

'City' in 'New York City'

A neuron in layer 4 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 responds to the word 'City' when it follows 'New York', and also to a few other place-name endings. Its strongest negative responses are on 'even' in phrases like 'even if' and 'even though'. Turning it up or down did not produce a clear, consistent change in the written 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

    a storage tank, which would connect to a drip irrigation system to water fruits and vegetables.⏎Our 3D viewer requires a WebGL enabled browser http://get.webgl

  2. #2 peak

    6, the Soviet probe Luna 9 became the first manmade object to make a soft landing on the moon.⏎In 1971, New York City police officer Frank Serpico, who had charged there was widespread corruption in the

  3. #3 peak

    ’t,” he said.⏎Trujillo enlisted in the U.S. Army in <|endoftext|>On June 12, 1909, New York City began an eight-day celebration of the connection of the East Side of Manhattan with Long

  4. #4 peak

    is the Associate Professor of Stage Management and Production Manager at Brooklyn College. He has been production stage manager for numerous Broadway and regional productions, as well as New York City Opera. His diverse career in live performance presentations includes: Entertainment Operations Manager for the

  5. #5 peak

    were circumcised by the same person, the report said. The New York City Commissioner of Health has issued a directive ordering that person to stop performing direct orogen

  6. #6 peak

    on display. However, while Fort Hamilton is an active Army base, the defense of New York City is now the responsibility of the Navy and the Air Force.⏎Part of the Installation

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: a problem. So, if you need to get into the fun part: a challenge. And if you want to get the adventurous: a task. So, if you want to get the even more adventurous: a challenge. So, if

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

    course, here too, you want to redirect the output.⏎svn diff > ~/patch.diff⏎There are useful variants too (shown here without redirection)⏎Note: even if svn can make the difference of another directory (svn diff mydirectory), it

  2. #2 peak

    levels below that standard, it is generally considered safe to drink. People are strongly advised to take special care if a water supply contains two VOCs, however, even if they are both below the maximum allowable concentration.⏎Bioworx has just

  3. #3 peak

    not quite true. Even though they represent one family of plants, there are some distinctive features between the two

  4. #4 peak

    by buying tracts of land, but which often fail to address the needs or rights of local Indigenous peoples. Uninhabited forest areas often go unprotected, even if purchased through a buy-an-acre program. It is not uncommon for loggers

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. L3:46 suppresses
  2. L3:456 weakly suppresses
  3. L3:2720 weakly suppresses
  4. L3:50 weakly excites
  5. L3:2280 weakly excites
  6. L0:2846 weakly excites
  7. L3:464 weakly excites
  8. L3:703 weakly suppresses
  9. L3:59 weakly excites
  10. L2:977 weakly excites
  11. L3:421 weakly suppresses
  12. L2:2897 weakly excites
  13. L3:1936 weakly excites
  14. L1:1053 weakly suppresses
  15. L2:793 weakly suppresses
  16. L0:872 weakly excites
  17. L2:2912 weakly excites
  18. L1:2036 weakly suppresses
  19. L0:1042 weakly suppresses "This" starting a labelled sentence
  20. L2:2073 weakly suppresses
  21. L3:691 weakly excites
  22. L2:2213 weakly excites
  23. L2:2679 weakly excites
  24. L3:122 weakly excites

Feeds later layers

  1. L5:2775 suppresses
  2. L5:1877 excites
  3. L9:1328 suppresses
  4. L23:851 weakly suppresses
  5. L6:2739 weakly suppresses
  6. L5:2197 weakly excites
  7. L5:813 weakly excites
  8. L11:1472 weakly suppresses
  9. L7:1658 weakly excites
  10. L25:2634 weakly excites
  11. L24:524 weakly excites
  12. L8:1020 weakly excites
  13. L5:671 weakly suppresses
  14. L15:642 weakly suppresses
  15. L8:1070 weakly excites
  16. L22:1149 weakly suppresses
  17. L12:1195 weakly excites
  18. L10:845 weakly excites
  19. L14:1945 weakly suppresses
  20. L16:2698 weakly suppresses 'even' after 'but' or 'while'
  21. L22:2947 weakly suppresses
  22. L23:2297 weakly excites
  23. L23:1582 weakly excites
  24. L14:1999 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 -11.37 and -11.071 values between -10.18 and -9.891 values between -9.89 and -9.591 values between -8.70 and -8.411 values between -6.93 and -6.631 values between -1.60 and -1.303 values between -0.71 and -0.42354 values between -0.42 and -0.123412 values between -0.12 and 0.18133 values between 0.18 and 0.47offnegativepositiveoften

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. 'er
  2. 鞅
  3. ellers
  4. ductor
  5. rift
  6. clid
  7. InMillis
  8. reator
  9. ufreq
  10. ях

Suppresses

  1. sterdam
  2. assigned
  3. .Configuration
  4. JSImport
  5. 防卫
  6. Báo
  7. borg
  8. 皲
  9. mented
  10. nie