Neuron Explorer

Qwen3-0.6B

the letter p starting a word piece

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

This neuron responds to the letter p at the start of a word piece, such as the p in pheasant, pire, pooks or piggery. Negative values appear on p after a slash or hyphen, as in pulses and paralysis-style splits. Turning it up made the model repeat words like 'validity validity', which does not clearly connect to the letter.

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

    (pheasant) is literally 'mountain bird', and ashibiki is a pillow-word for mountain, which is itself the first half of the word for pheasant.<|endoftext|>North American freshwater fishes are going extinct at an alarming rate compared

  2. #2 peak

    normal circumstances I would have resigned immediately.⏎to continue as normal se poursuivre normalement⏎→ Life in Israel will continue as normal.⏎worse than normal pire que d'habitude⏎→ London's traffic was worse than normal during

  3. #3 peak

    a muddy hill. Below us an orange glow floats like fog above the border checkpoint at San Ysidro, California. “En nombre del cielo, les pido posada,” sings one group. “Pues no puede andar mi

  4. #4 peak

    gay,⏎I met a pretty dairy-maid all in the new mown hay.⏎* 4 *⏎As I was ridding out one day I saw some pooks of hay;⏎Is this not a very pretty place for boys and maids

  5. #5 peak

    x."⏎But it is reasonably certain that Roma quadrata is also used in the sense of mundus or augural centre of the city-templum, and p421therefore we should be obliged to assume three different meanings of the

  6. #6 peak

    health. Brushing after meals, using a dentist-approved mouthwash and mandating routine visits to a Sydney CBD dentist for can help you avoid this and keep your pearly whites and gums healthy.⏎3. Having one or more oral piercings

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:retalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalalal

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

    These studies principally showed that meat and dairy are among the largest contributors to GHGE; whereas high intake of fruits, vegetables (F&V), and legumes/pulses/nuts consumption is associated with the lowest GHGE

  2. #2 peak

    important to look at these perceptions in context, however. High climate variability is very much part of life in the case study areas. Indeed, pastoral and agro-pastoral systems exhibit important features that make them well suited to climate uncertainty and variability

  3. #3 peak

    .⏎The prosthetic hand has five ‘fingers’ embedded with segments of fibre that mimic the hand bones and are connected to a 3D-printed ‘palm’ shaped like a human hand.⏎In place of mounted electrical motors, the

  4. #4 peak

    systems, including continuous web.⏎How does Plasma Compare to Corona?Corona treatment is an electrical process that uses ionized air to increase the surface tension of non-porous substrates. Normally, corona-treating systems operate at an electrical voltage

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. L20:322 suppresses
  2. L18:103 suppresses
  3. L21:346 suppresses
  4. L23:1099 excites
  5. L22:979 excites
  6. L17:214 excites
  7. L19:557 excites
  8. L10:237 excites
  9. L6:875 suppresses
  10. L18:196 suppresses
  11. L0:497 excites
  12. L2:1554 excites
  13. L5:2009 excites
  14. L16:248 excites
  15. L23:206 weakly suppresses
  16. L11:29 weakly excites
  17. L7:300 weakly excites
  18. L16:353 weakly suppresses
  19. L15:1272 weakly suppresses
  20. L19:165 weakly excites
  21. L8:14 weakly excites
  22. L14:72 weakly excites
  23. L26:1564 weakly excites
  24. L16:829 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 -10.38 and -9.761 values between -6.66 and -6.042 values between -2.32 and -1.708 values between -1.70 and -1.08133 values between -1.08 and -0.463482 values between -0.46 and 0.16258 values between 0.16 and 0.7815 values between 0.78 and 1.401 values between 1.40 and 2.021 values between 2.02 and 2.642 values between 2.64 and 3.261 values between 3.87 and 4.491 values between 5.11 and 5.731 values between 8.83 and 9.451 values between 13.79 and 14.41offnegativepositiveoften

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. futures
  2. (valid
  3. validity
  4. tribution
  5. 他又
  6. valid
  7. valid
  8. VALID
  9. Valid
  10. ventional

Suppresses

  1. rem
  2. ret
  3. reste
  4. rel
  5. ри
  6. rep
  7. dfs
  8. ei
  9. apers
  10. olo