Neuron Explorer

Qwen3-0.6B

mid-word fragments, often 'intric'

A neuron in layer 17 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featured, so it has steering outputs.

What this neuron does

This neuron fires on the first piece of a long word that the model has split into smaller chunks, most often the start of 'intricately' or 'intricacies'. When it was turned up, the model fell into repeated nonsense syllables, so the steering results are broken.

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

    Your garden is an orchestra! Healthy gardens play to the eyes like a fine-tuned orchestra plays to a conductor’s ears. Not unlike a musical composition, the intricately woven notes of a successful garden involve a high level of behind-the-scene

  2. #2 peak

    property transfers with the sale. Typically, courts will favor tenants and buyers in landlord-tenant disputes and seller-buyer disputes.⏎Written Agreement: A written agreement supersedes any standard interpretation of real

  3. #3 peak

    balance between mind, body, and spirit. Ayurveda is known to be the science of healing. Ayurveda takes it that the web of life is intricately interwoven and interconnected. Ayurvedic medicine is considered to be one

  4. #4 peak

    , there is no real harm in using either variation.⏎The advantage in using one formula over another is left to the analyst to decide or may be determined by the intricacies of the analysis being done.⏎How to Interpret the Fixed Assets Turnover Ratio

  5. #5 peak

    ). Use a whole-house, integrated air-filtration system with a HEPA (high-efficiency particulate air) filter. Use only low-toxic adhesives and plastic-free grout when installing tile. Make sure the kitchen range hood

  6. #6 peak

    in future programs.⏎Pulse of the Planet is presented with support provided by the National Science Foundation. I’m Jim Metzner.⏎Around the world, mangroves are being destroyed to make room

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, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one, one,

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 authors speculated that:⏎It was suggested that these responses occurred in order to preserve a similar rate of change in systemic responses (i.e. RPE, ⩒O2 and HR) and NMF development, ultimately resulting in same TT

  2. #2 peak

    a 4-km cycling time trial (TT)". The results were not what I expected:⏎TT performance was not different amongst pacing profiles.⏎RPE, ⩒O2 and HR increased similarly throughout the TT regardless the pacing strategy.⏎Therefore

  3. #3 peak

    understand – לְהָבִין. But whereas in להבין, the root is plugged into the active-causative הִפְעִיל verb form, in ביון, the root is plugged

  4. #4 peak

    sores in the animals’ mouth and the feet, as well as issues with their reproductive systems. The virus does not affect humans and it belongs to the Picornaviridae family. Currently, there are seven known strains, of which

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. L6:293 strongly suppresses
  2. L7:9 strongly suppresses
  3. L11:42 strongly suppresses
  4. L5:58 strongly excites
  5. L10:43 excites "terms" in "in terms of"
  6. L6:1846 excites
  7. L4:1045 excites
  8. L12:2995 excites
  9. L6:1151 excites
  10. L5:142 excites
  11. L6:1074 suppresses
  12. L6:2400 suppresses
  13. L16:1130 suppresses
  14. L7:210 suppresses
  15. L4:146 suppresses
  16. L8:393 suppresses
  17. L10:105 suppresses
  18. L13:1512 suppresses
  19. L6:184 excites
  20. L8:62 suppresses
  21. L4:1781 suppresses
  22. L10:810 suppresses
  23. L3:2921 excites
  24. L10:1872 excites

Feeds later layers

  1. L18:170 strongly excites
  2. L19:614 strongly excites
  3. L19:541 strongly excites
  4. L21:189 strongly excites document break token mid-sentence
  5. L20:131 excites
  6. L18:489 excites
  7. L19:813 suppresses
  8. L21:1904 excites
  9. L22:164 excites
  10. L24:385 excites
  11. L19:121 suppresses
  12. L25:1175 excites
  13. L18:91 excites
  14. L19:1167 excites
  15. L20:1843 excites
  16. L22:45 suppresses word before a number, range or quantity
  17. L19:1143 excites
  18. L20:390 excites
  19. L20:545 excites
  20. L21:237 excites
  21. L19:696 excites
  22. L25:206 excites
  23. L18:228 excites
  24. L23:209 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 -4.18 and -3.721 values between -3.72 and -3.251 values between -3.25 and -2.7921 values between -1.86 and -1.39340 values between -1.39 and -0.931728 values between -0.93 and -0.461534 values between -0.46 and 0.0052 values between 0.00 and 0.4726 values between 0.47 and 0.9324 values between 0.93 and 1.4023 values between 1.40 and 1.8618 values between 1.86 and 2.339 values between 2.33 and 2.7913 values between 2.79 and 3.2611 values between 3.26 and 3.7214 values between 3.72 and 4.199 values between 4.19 and 4.658 values between 4.65 and 5.122 values between 5.12 and 5.586 values between 5.58 and 6.054 values between 6.05 and 6.517 values between 6.51 and 6.986 values between 6.98 and 7.446 values between 7.44 and 7.914 values between 7.91 and 8.372 values between 8.37 and 8.844 values between 8.84 and 9.306 values between 9.30 and 9.775 values between 9.77 and 10.236 values between 10.23 and 10.705 values between 10.70 and 11.161 values between 11.16 and 11.632 values between 11.63 and 12.092 values between 12.09 and 12.563 values between 12.56 and 13.021 values between 13.02 and 13.482 values between 13.48 and 13.951 values between 13.95 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. esModule
  2. onus
  3. ques
  4. 싱
  5. plevel
  6. ewed
  7. edException
  8. �示
  9. odied
  10. efs

Suppresses

  1. 留
  2. 外
  3. ,
  4. </
  5. -
  6. ”
  7. +
  8. j
  9. <j
  10. >