Neuron Explorer

Qwen3-0.6B

unclear: inside numbers and names

A neuron in layer 17 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 pieces of a word, name or number that are split in the middle, like the 'al' in a place name, the '3' in a model code or the 'ic' in a longer word. Its meaning is not clear, and turning it up only produced repeated words or long strings of zeros.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    not get drunk on wine but is instead filled with the Spirit; a noble man does not love the world or anything in the world. Proverbs 31:8-9 -- a noble man stands up for what is right and shows concern

  2. #2 peak

    was capable of playing 27 chords, and early models required a rubber plectrum to play, though later models featured an updated strum plate. The OM27 was a commercial failure, so Suzuki released the OM36 and OM

  3. #3 peak

    general rule, trough loading of 45-percent can be used for light, free-flowing and non-abrasive bulk materials. Trough loadings of 15 and 30-percent are typically used for denser, sluggish and

  4. #4 peak

    lava flows; the flows surround the 322--meter high stratovolcano (Jabal is translated as "mountain" in Arabic). Jabal Abyad, in the center

  5. #5 peak

    upright lilies and used by many for forcing. In fact, Peter Schenk bred it with forcing and the cut flower trade in mind. However, L. 'Siberia' and L. 'Jan de Graaf', two other upright

  6. #6 peak

    or distress, especially if you’re unable to respond to someone else’s situation accordingly. Empathic anger is a response to a person’s pain. Empathic distress is feeling that pain, which can later

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 the fact that I have been thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that

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

    aurus at seeNotice is not used in the progressive. You say: I notice you’re wearing glasses. ✗Don’t say: I’m noticing you’re wearing glasses.2 → be noticed/get (somebody) noticed3 → somebody can

  2. #2 peak

    lottery ticket cannot be rationally accounted for by decision models based on expected value maximization, because the expected value of a lottery ticket is always zero (or close to it). However, the non-monetary benefits obtained from playing the lottery can outweigh

  3. #3 peak

    we had vegetables in late September. And each year, for the past seven years, we managed to extend our season for an additional month into the winter. To accomplish this, we focused on three key elements that helped us prolong our winter

  4. #4 peak

    Colo 80401⏎Final Grant Report⏎Physics Department, Particulate Science Laboratory, Colorado School of Mines, Golden, Colorado⏎Colorado School of Mines, Golden, Colorado<|endoftext|>It is almost Easter and inevitably this meme will be

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. L16:1991 strongly suppresses
  2. L16:1512 excites
  3. L16:942 excites
  4. L16:2567 excites
  5. L16:2657 excites
  6. L16:872 suppresses
  7. L16:1151 excites
  8. L16:1613 suppresses
  9. L16:324 excites
  10. L16:779 suppresses
  11. L15:2915 excites
  12. L14:1180 excites
  13. L15:262 excites
  14. L16:1975 excites
  15. L16:2364 suppresses
  16. L16:2086 excites
  17. L15:2659 excites
  18. L16:2044 suppresses
  19. L15:1400 excites
  20. L15:968 excites
  21. L14:1817 suppresses
  22. L14:2388 suppresses
  23. L16:2827 excites
  24. L14:1613 suppresses

Feeds later layers

  1. L19:1462 strongly excites
  2. L18:1119 strongly excites
  3. L18:202 suppresses unclear: list commas and mixed tokens
  4. L18:812 suppresses
  5. L19:1137 excites
  6. L18:2522 excites
  7. L18:1275 excites
  8. L18:154 excites
  9. L19:1737 excites
  10. L19:1214 excites
  11. L19:656 suppresses
  12. L19:687 excites
  13. L18:494 excites
  14. L18:1174 excites
  15. L18:2923 excites
  16. L19:2281 suppresses
  17. L18:2113 excites
  18. L19:2199 excites
  19. L20:668 excites
  20. L20:294 suppresses
  21. L18:686 suppresses
  22. L19:1665 weakly suppresses
  23. L18:183 weakly suppresses
  24. L18:817 weakly 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.

1 values between -9.63 and -9.261 values between -8.53 and -8.162 values between -7.79 and -7.421 values between -7.05 and -6.682 values between -6.32 and -5.951 values between -5.95 and -5.583 values between -5.58 and -5.212 values between -5.21 and -4.843 values between -4.84 and -4.475 values between -4.47 and -4.1010 values between -4.10 and -3.7410 values between -3.74 and -3.3714 values between -3.37 and -3.0021 values between -3.00 and -2.6325 values between -2.63 and -2.2641 values between -2.26 and -1.8962 values between -1.89 and -1.5395 values between -1.53 and -1.16206 values between -1.16 and -0.79389 values between -0.79 and -0.421131 values between -0.42 and -0.051580 values between -0.05 and 0.32184 values between 0.32 and 0.6965 values between 0.69 and 1.0623 values between 1.06 and 1.4216 values between 1.42 and 1.796 values between 1.79 and 2.163 values between 2.16 and 2.531 values between 2.53 and 2.901 values between 2.90 and 3.273 values between 3.27 and 3.641 values between 4.74 and 5.11offnegativepositiveoften

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. ony
  2. rid
  3. idal
  4. asca
  5. bracelets
  6. urg
  7. chie
  8. ница
  9. Mirage
  10. 万台

Suppresses

  1. フェ
  2. 感
  3. 小时前
  4. �认
  5. 操纵
  6. омер
  7. omer
  8. ificación
  9. (fe
  10. 等活动