Neuron Explorer

Qwen3-0.6B

'st' starting a word fragment

A neuron in layer 27 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

The neuron lights up on the letters 'st' when they start the second half of a word that gets split up, like 'sto' in stoma or 'stav' in staves. When it was turned up, the model's replies drifted into odd, repetitive phrasing rather than anything clearly about those words.

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

    banding is specially designed to limit food consumption in severely obese patients. When placed around the patient's stomach, the saline-filled band forms a small gastric pouch and stoma. Inflation of the band can be adjusted to control the amount of food

  2. #2 peak

    first step should be to become familiar with the music notes.⏎In the Western system of music notes, written music is arranged on two sets of five lines, called staves. The upper stave is called the treble stave; the lower

  3. #3 peak

    take your chances in the no-holds-barred class, or on the roof.⏎Passengers were still pushing their way on board and trying to find somewhere to stow their endless boxes and bags of luggage when a bell clanged and we were

  4. #4 peak

    chunks of dirt and remove rocks. The axe-style blade on the head of a cutter mattock is ideal tool for chopping through root-laden soil near tree stumps.⏎Forks are particularly useful for turning soil, digging up root crops,

  5. #5 peak

    may rise by 3,000 percent by 2029 with better regulations.⏎The ever increasing demand for carbon credits could be the straw that stirs the milkshake for climate innovation in Mongolia and possibly resolves the air pollution problem

  6. #6 peak

    for domestic rubbish. This is because the biomass in domestic rubbish can only be gasified at low temperatures, but this in turn creates tars and waxes that stifle the process.⏎However, it turns out that these tars and waxes

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: maybe even digress even digress digress digdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdigdig

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

    a thin, flimsy plastic but you could, and still can, buy proper-lensed 3D blue and red glasses) use what is called anaglyph technology. This has since been superseded by Passive and Active 3D

  2. #2 peak

    wall next to the art, a statement from the artist about their intent for the piece.⏎When Marcel Duchamp submitted an upside-down porcelain urinal signed “R.Mutt”<|endoftext|>The Galileo Museum in Florence is an international reference on Galileo

  3. #3 peak

    Re became Tefnut's father. According to one myth, Tefnut became angry with her father Re and ran off to Nubia. The god Thoth went to find her and conjoled her into returning to Egypt.⏎

  4. #4 peak

    00-205. See general information about how to correct material in RePEc.⏎For technical questions regarding this item, or to correct its authors, title, abstract, bibliographic or download information, contact: (Michael McNulty

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. L21:237 suppresses
  2. L23:206 suppresses
  3. L20:669 suppresses
  4. L24:105 weakly suppresses
  5. L18:151 weakly suppresses
  6. L1:3019 weakly suppresses
  7. L16:1129 weakly excites
  8. L18:1077 weakly excites
  9. L15:434 weakly excites
  10. L20:52 weakly suppresses
  11. L18:775 weakly suppresses
  12. L20:2642 weakly excites
  13. L21:475 weakly excites
  14. L18:2453 weakly suppresses
  15. L2:289 weakly suppresses
  16. L5:2370 weakly suppresses "from" after change or transfer verbs
  17. L21:2904 weakly excites
  18. L20:151 weakly excites
  19. L22:589 weakly excites
  20. L9:1544 weakly excites
  21. L22:3027 weakly excites
  22. L17:1948 weakly excites
  23. L4:2778 weakly suppresses
  24. L12:2943 weakly suppresses

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.

2 values between -13.56 and -12.701 values between -10.98 and -10.121 values between -10.12 and -9.261 values between -8.40 and -7.541 values between -6.68 and -5.821 values between -5.82 and -4.961 values between -4.96 and -4.101 values between -3.24 and -2.384 values between -2.38 and -1.5242 values between -1.52 and -0.663474 values between -0.66 and 0.20349 values between 0.20 and 1.0610 values between 1.06 and 1.924 values between 1.92 and 2.781 values between 2.78 and 3.641 values between 3.64 and 4.503 values between 4.50 and 5.362 values between 5.36 and 6.221 values between 6.22 and 7.081 values between 7.08 and 7.941 values between 7.94 and 8.791 values between 8.79 and 9.651 values between 9.65 and 10.511 values between 12.23 and 13.091 values between 13.95 and 14.811 values between 17.39 and 18.251 values between 19.97 and 20.83offnegativepositiveoften

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. üler
  2. lients
  3. ows
  4. owed
  5. 男士
  6. ected
  7. ipse
  8. -win
  9. ieri
  10. Best

Suppresses

  1. dig
  2. уд
  3. よう
  4. cepts
  5. 二代
  6. code
  7. rens
  8. //**⏎
  9. 侗
  10. odia