Neuron Explorer

Qwen3-0.6B

capital G starting names

A neuron in layer 24 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 a capital G that begins a person's surname or a middle initial, like the G in 'Georges Gonthier' or 'P. G. Pessler'. It goes quiet or negative on the 'Gu' in place names such as Guiana and Guwahati. Turning it up produced repeated quotation marks, which doesn't obviously match the trigger.

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

    ,” he said. “But the computer proof may be just part of the game.” He also pointed out that proof generation and proof checking are quite different. Georges Gonthier, for example, used computers to formally verify the proof of the four

  2. #2 peak

    we must pluck.’⏎Source: John Keats, 14 Feb – 3 May 1819, Selected Letters, ed. Robert Gittings (Oxford World Classics, 2002/200

  3. #3 peak

    had been trying to crack for 2,000 years.⏎"If we can determine what that distance is, we can scale the solar system," explained Chris Giersch, education and outreach manager at NASA Langley Research Center in Hampton.⏎

  4. #4 peak

    increases the rate of plant production and has been used for the elimination of diseases in certain plants.⏎Tissue Culture History⏎The first person to attempt cell culture was G. Haberlandt in 1902. He was unsuccessful, but

  5. #5 peak

    Johannes Vogel, Beat Grenacher, Robert A. Jacobs, Margarete Arras, and Max Gassmann. Acute and chronic elevation of erythropoietin in the

  6. #6 peak

    - The initial patchiness of this lesion with, until a very late<|endoftext|>Editor’s note: The following is extracted from Our Soldiers, by W. H. G. Kingston (published 1863). All spelling in the original.⏎

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 place of a later material to the place of an earlier material in the material place place later material. It’s just that material to material, material to material, material to material, material to material, material to material, material to material

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

    , the first of Europe's Automated Transfer Vehicles, was launched from the ESA spaceport in French Guiana early March 9. It linked up with the space station almost a month

  2. #2 peak

    California's Guaymas Basin in the⏎"We have a huge problem on our hands," says

  3. #3 peak

    ean Drama to each Unit are provided in the list of UG-CBCS Central University & State University Syllabus eg. Dibrugarh University and Guwahati University so that you can easily browse through different College and University Guide

  4. #4 peak

    0742006000100003.⏎This article compares three eighteenth century evaluations about jesuitic missions among the Guaranis - Relação abbreviada da republica, que os religios

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. L23:1489 excites
  2. L19:916 suppresses
  3. L20:350 weakly suppresses
  4. L20:236 weakly suppresses
  5. L22:1412 weakly excites
  6. L21:2864 weakly excites
  7. L19:1004 weakly suppresses
  8. L22:530 weakly suppresses
  9. L2:2029 weakly excites
  10. L18:202 weakly suppresses unclear: list commas and mixed tokens
  11. L19:860 weakly excites
  12. L13:58 weakly suppresses
  13. L19:688 weakly excites
  14. L5:46 weakly excites
  15. L17:1061 weakly suppresses
  16. L18:404 weakly excites 'get', 'go' and 'got' after a subject
  17. L21:2826 weakly excites
  18. L13:1856 weakly suppresses
  19. L22:384 weakly excites
  20. L23:1067 weakly suppresses
  21. L3:1423 weakly suppresses
  22. L16:1572 weakly excites
  23. L15:199 weakly excites
  24. L20:486 weakly excites

Feeds later layers

  1. L27:661 excites
  2. L27:315 weakly excites
  3. L25:1641 weakly excites
  4. L26:463 weakly excites digits inside years and numbers
  5. L25:2912 weakly excites
  6. L27:1255 weakly excites
  7. L27:1647 weakly suppresses
  8. L25:2628 weakly suppresses
  9. L25:44 weakly excites
  10. L26:2530 weakly excites
  11. L27:1084 weakly excites
  12. L25:776 weakly excites
  13. L25:368 weakly excites
  14. L25:1864 weakly suppresses
  15. L25:564 weakly suppresses
  16. L26:602 weakly excites
  17. L27:2312 weakly suppresses
  18. L26:1655 weakly suppresses
  19. L25:1991 weakly excites
  20. L26:2799 weakly suppresses
  21. L26:2957 weakly suppresses
  22. L25:1285 weakly suppresses
  23. L26:451 weakly suppresses
  24. L25:73 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 -67.25 and -64.971 values between -19.35 and -17.071 values between -14.79 and -12.511 values between -12.51 and -10.233 values between -10.23 and -7.953 values between -7.95 and -5.674 values between -5.67 and -3.3917 values between -3.39 and -1.113863 values between -1.11 and 1.1810 values between 1.18 and 3.461 values between 3.46 and 5.741 values between 8.02 and 10.301 values between 12.58 and 14.861 values between 21.70 and 23.98offnegativepositiveoften

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. "
  2. '
  3. "%
  4. “
  5. thất
  6. “
  7. so
  8. 「
  9. lege
  10. .dom

Suppresses

  1. OCI
  2. olib
  3. redund
  4. serter
  5. OV
  6. oci
  7. uries
  8. Oswald
  9. 喹
  10. suite