Neuron Explorer

Qwen3-0.6B

word fragments before the word ends

A neuron in layer 27 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on a piece of a word that is split in the middle, like the start of a name or a long word, right before the rest of it arrives. When it was turned up, the model kept writing the single word 'accept' over and over.

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

    depending on where you are. I set up my camera set my lens to its lowest aperture (for my 18mm kit lens it is f/3.5) and then focused the image. Most advise using a depth

  2. #2 peak

    the<|endoftext|>Would you be able to live in a house that was just four feet wide? That’s right – four feet wide! Well, Polish architect Jakub Szczesny began designing the Keret House about three years ago when he discovered

  3. #3 peak

    who have only the slightest signs of potential heart problems, according to a January 2019 article in the International Journal of Cardiology. Dr. Claudia Szlejf, a researcher at HU-USP, examined data on 1

  4. #4 peak

    in which case raising the hands is regarded as an additional action (and should not be done). Examples of such acts of worship include salaah (prayer), khutbah (sermon), tawaaf (circumambulation of the

  5. #5 peak

    hailed as a conquering hero, greeted with palm branches and loud hosannas. But Holy Week reaches its climax on Good Friday, when Jesus is crowned with thorns and led out of the city like a common criminal, on to crucifix

  6. #6 peak

    ment on body surfaces. Penicillin was finally isolated in 1939, and in 1944 Selman Waksman and Albert Schatz, American microbiologists, isolated streptomycin

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: What would you do with platforms centered around platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms platforms

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

    if a certain behavior is found in the Old Testament narratives, and that behavior is not specifically condemned in those same narratives, then the narratives are teaching that this behavior is morally acceptable.⏎Is that how we should understand the narratives in

  2. #2 peak

    , these groups are simply subject to laws against adultery or illegal cohabitation – laws that are not universally enforced because they also criminalize other behaviors that are otherwise socially sanctioned. However, some “fundamentalist” polygamists marry off girls

  3. #3 peak

    two alternatives gives rise to a highly confusing situation. People will lose all faith in the karmas prescribed by the śāstras. The third alternative has to be ruled out since it is not possible for a jīva (individual soul

  4. #4 peak

    1473 and as a young man became fascinated by math, medicine and science, particularly astronomy. His studies led him to challenge what was then a widely accepted belief — that the Earth stood still in outer space while the sun and other

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. L23:2412 excites 'well' or 'get' before a verb phrase
  2. L24:2299 excites
  3. L23:2079 suppresses
  4. L21:2132 excites
  5. L22:2050 excites
  6. L24:2714 excites
  7. L25:2522 suppresses
  8. L24:1255 suppresses
  9. L22:2013 excites
  10. L15:1089 suppresses
  11. L17:2903 weakly suppresses
  12. L23:249 weakly suppresses
  13. L2:49 weakly suppresses
  14. L21:1137 weakly suppresses
  15. L21:3018 weakly excites
  16. L21:2319 weakly excites
  17. L20:2254 weakly excites
  18. L16:2589 weakly excites
  19. L5:1986 weakly excites
  20. L23:1804 weakly suppresses
  21. L18:2868 weakly suppresses
  22. L24:2134 weakly excites
  23. L18:2484 weakly excites
  24. L22:2102 weakly excites

Feeds later layers

None: this is the last layer.

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 -17.39 and -16.481 values between -16.48 and -15.581 values between -11.04 and -10.131 values between -7.41 and -6.501 values between -6.50 and -5.591 values between -5.59 and -4.682 values between -3.77 and -2.872 values between -2.87 and -1.9662 values between -1.96 and -1.052069 values between -1.05 and -0.141718 values between -0.14 and 0.7731 values between 0.77 and 1.676 values between 1.67 and 2.582 values between 2.58 and 3.491 values between 3.49 and 4.402 values between 4.40 and 5.311 values between 7.12 and 8.031 values between 8.03 and 8.941 values between 9.84 and 10.751 values between 10.75 and 11.661 values between 14.38 and 15.291 values between 16.20 and 17.111 values between 18.01 and 18.92offnegativepositiveoften

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. accept
  2. accept
  3. acceptance
  4. Accept
  5. accepting
  6. accepted
  7. accepts
  8. Accept
  9. ACCEPT
  10. accepted

Suppresses

  1. ingham
  2. pell
  3. ixer
  4. =count
  5. PLATFORM
  6. QText
  7. 嗽
  8. uru
  9. arro
  10. (entities