Neuron Explorer

Qwen3-0.6B

words before water, lava or magnetic fields

A neuron in layer 22 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 the word that comes just before a fluid or flow, such as 'flows of lava', 'magnetic field', or 'rinse with water'. When it was turned up, the model wrote more about water, clear liquid and currents. When it was turned down, the text moved toward talk about roles and positions.

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

    .⏎Mount Etna volcano on the island of Sicily, Italy, started erupting again on Wednesday, sending flames and sparks into the air, and flows of lava down its slopes. The air space of the nearby

  2. #2 peak

    ⏎"This is an erupting volcano, showing the process of releasing lava and smoke. Volcanoes erupt because heat within the Earth melts rocks and creates a substance called magma. Magma rises and pushes through the Earths surface and the volcano erupt

  3. #3 peak

    that eventually resulted from this discovery is driven by the Lorentz force, which is the force on an electric charge as it moves through a magnetic field. The development of traditional technologies such as fuel cells and the electric

  4. #4 peak

    ⏎Generation of Alternating Voltages and Currents⏎An alternating voltage may be generated⏎(1) by rotating a coil at constant angular velocity in a uniform magnetic field as shown in Fig. (a).⏎(2) by rotating a magnetic

  5. #5 peak

    The presence of charged ions means that a plasma is highly electrically conductive and responds strongly to magnetic and electric fields. Its behavior is most comparable to that of a gas, as

  6. #6 peak

    . As these products are not registered as FDA-approved, accidental ingestion may prove dangerous. If you are without an appropriate wash, the USDA suggests rinsing your produce with cold tap water and cleaning sturdy produce with a brush for further contaminant removal.⏎

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 role of a role in a business. A role is a position that has a certain job role, or a role in the life of a business. A role in the role, in the role, in the role, in the role

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

    be applied, such as process water for breweries, bottled water plants and pharmaceutical facilities.⏎UV-C light, often referred to as Germicidal UV, exists between 200 and 280nm and is therefore invisible to the naked

  2. #2 peak

    engine control unit will be able to control how much fuel gets injected into the engine. As you may know, there needs to be just the right amount of air and fuel mixed together in the internal combustion chamber of an engine.⏎Therefore, if the

  3. #3 peak

    these products are still a must-own for any outdoor enthusiast.⏎There are two types of ultraviolet rays that can cause harm: long-wave UVA and short-wave UVB.⏎- UVA Rays: These cause the skin to prematurely age and

  4. #4 peak

    uum 3 x 108⏎Air 3 x 108⏎Water 2.258 x 108⏎Glass 2 x 108⏎The change in speed causes the direction of light

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. L10:2446 suppresses
  2. L20:1660 suppresses
  3. L15:2975 excites
  4. L20:2079 suppresses
  5. L11:1369 weakly suppresses
  6. L21:1634 weakly suppresses
  7. L10:2754 weakly suppresses
  8. L21:1325 weakly suppresses
  9. L21:2970 weakly excites
  10. L10:3067 weakly excites
  11. L21:2361 weakly excites
  12. L8:2392 weakly excites
  13. L10:1079 weakly excites
  14. L21:669 weakly excites
  15. L13:2978 weakly suppresses
  16. L21:240 weakly excites
  17. L12:1065 weakly suppresses
  18. L20:2611 weakly suppresses
  19. L10:1496 weakly suppresses
  20. L16:2464 weakly excites
  21. L9:1807 weakly suppresses unclear: mixed, health lists and 'weight'
  22. L17:2626 weakly suppresses
  23. L7:1439 weakly suppresses
  24. L3:1850 weakly suppresses

Feeds later layers

  1. L27:2614 weakly excites
  2. L26:2499 weakly excites
  3. L23:1675 weakly excites
  4. L26:1869 weakly excites
  5. L27:1417 weakly excites
  6. L25:447 weakly excites
  7. L24:1802 weakly excites
  8. L23:1220 weakly excites
  9. L27:2466 weakly excites
  10. L24:2135 weakly excites
  11. L27:1641 weakly excites
  12. L24:1698 weakly excites
  13. L24:1562 weakly excites
  14. L25:1365 weakly excites
  15. L24:621 weakly excites
  16. L27:469 weakly excites
  17. L24:2820 weakly excites
  18. L23:967 weakly excites
  19. L24:2137 weakly suppresses
  20. L24:576 weakly excites prepositions in fixed phrases
  21. L23:704 weakly suppresses
  22. L25:1964 weakly excites
  23. L25:2051 weakly excites
  24. L23:1887 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 -8.15 and -7.391 values between -7.39 and -6.642 values between -6.64 and -5.883 values between -5.88 and -5.124 values between -5.12 and -4.375 values between -4.37 and -3.619 values between -3.61 and -2.8622 values between -2.86 and -2.1038 values between -2.10 and -1.34253 values between -1.34 and -0.592703 values between -0.59 and 0.17613 values between 0.17 and 0.92100 values between 0.92 and 1.6845 values between 1.68 and 2.4429 values between 2.44 and 3.1920 values between 3.19 and 3.9515 values between 3.95 and 4.7012 values between 4.70 and 5.467 values between 5.46 and 6.221 values between 6.22 and 6.973 values between 6.97 and 7.732 values between 7.73 and 8.493 values between 8.49 and 9.242 values between 9.24 and 10.003 values between 10.00 and 10.751 values between 10.75 and 11.511 values between 11.51 and 12.273 values between 12.27 and 13.021 values between 13.02 and 13.781 values between 13.78 and 14.532 values between 16.05 and 16.801 values between 16.80 and 17.561 values between 18.31 and 19.071 values between 21.34 and 22.09offnegativepositiveoften

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. 水流
  2. flows
  3. 水
  4. water
  5. 流水
  6. 水量
  7. flows
  8. streams
  9. water
  10. liqu

Suppresses

  1. enda
  2. SetName
  3. TabControl
  4. 身份证
  5. 号码
  6. anny
  7. ecycle
  8. ích
  9. ROLE
  10. номер