Neuron Explorer

Qwen3-0.6B

'tr' word-start 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

This neuron responds to the syllable 'tr' when it starts the second half of a split word, like 'tr|achea' or 'tr|unks'. Turning it up made the model repeat the word 'operations', and turning it down gave rambling text.

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

    originating in the pharynx and serving to amplify sound (vocal sacs); (3) sacs that branch off from the widened posterior end of the trachea (in male striped seals) or from the esophagus (in

  2. #2 peak

    humerus. The posterior fat pad is not seen because it lies in the olecranon fossa of the distal humerus, depressed superficially by the triceps tendon and anconeus muscle. Traumatic effusions will lift the posterior

  3. #3 peak

    mother, measurement of two hormones in the mother’s blood and the scan findings of nuchal translucency thickness, nasal bone , and blood flow through the tricuspid valve of the fetal heart and ductus venosus and fetal abnormalities

  4. #4 peak

    the famous general Kato Kiyomasa, 1561-1611) grips the reins as his horse rears in fright, trampling men beneath its hooves. At lower left, another warrior turns his head

  5. #5 peak

    not prevent lanternflies from moving around in a landscape and returning.⏎When the nymphs first hatch, they will walk up the trunks of trees to feed on the softer, new growth of the plant. People

  6. #6 peak

    white satin wraparounds, with head-dresses of coconut palm fronds and marigolds, danced in the temple complex to sacred tunes played by gamelan troupes.⏎Young men dragged a buffalo three times around the temple, with the

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: if I wanted to sendererererererererererererererererererererererererererererererererererererererererererer

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

    history of Eastern Europe and adjacent territories. The main and most numerous group is the Russian Cossacks (казаки) of the Don, Kuban, Terek and Ural rivers and the Ukrainian Cossacks (козаки

  2. #2 peak

    bridement to remove loose edges⏎- Autologous chondrocyte implantation to create new cartilage and replace what is damaged⏎- Mosaicplasty to take healthy cartilage and use it to replace damaged cartilage⏎You

  3. #3 peak

    environment properly, the consequences for people, the environment and the Northern Powerhouse economy can be disastrous."⏎Fires on Rivington Moor and Winter Hill, near Chorley, were declared as major incidents as flames spread over dry ground during the

  4. #4 peak

    underlying infection(s) but the immune dysfunction as well. Treatment can include antibiotics, immune modulating therapies, such as Intravenous immunoglobulin (IVIG), plasmapheresis and steroids. Anti-inflammatories, anti-histamines

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. L26:1676 strongly suppresses
  2. L24:1269 strongly excites
  3. L21:2401 suppresses
  4. L24:2197 excites
  5. L22:2019 excites
  6. L26:1564 excites
  7. L24:2619 suppresses
  8. L22:2340 excites
  9. L26:1703 excites
  10. L24:385 excites
  11. L26:123 excites
  12. L21:781 weakly suppresses
  13. L24:8 weakly suppresses
  14. L8:16 weakly suppresses
  15. L20:1097 weakly excites
  16. L20:1138 weakly suppresses
  17. L17:2710 weakly suppresses
  18. L18:569 weakly excites
  19. L25:1048 weakly excites
  20. L20:1814 weakly excites
  21. L8:14 weakly excites
  22. L6:601 weakly suppresses
  23. L23:2111 weakly excites
  24. L8:4 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 -37.19 and -35.232 values between -11.73 and -9.772 values between -9.77 and -7.814 values between -7.81 and -5.853 values between -5.85 and -3.8911 values between -3.89 and -1.933218 values between -1.93 and 0.03627 values between 0.03 and 1.9813 values between 1.98 and 3.9415 values between 3.94 and 5.902 values between 5.90 and 7.861 values between 7.86 and 9.821 values between 9.82 and 11.781 values between 15.70 and 17.652 values between 17.65 and 19.613 values between 19.61 and 21.571 values between 25.49 and 27.451 values between 39.20 and 41.16offnegativepositiveoften

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. operations
  2. operation
  3. Operation
  4. operations
  5. Operations
  6. operation
  7. Operation
  8. 操作
  9. (operation
  10. Operations

Suppresses

  1. bruar
  2. issing
  3. eria
  4. 氪
  5. eri
  6. ansa
  7. er
  8. po
  9. ◄
  10. 眼里