Neuron Explorer

Qwen3-0.6B

'less' before a comparative or negative effect

A neuron in layer 15 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 fires on the word 'less' when it means something is reduced, like 'makes it less prone to drought' or 'less acidic'. When it was turned up, the model's writing drifted into loose, rambling text and did not clearly change topic.

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

    At 144 degrees, the proteins in eggs—which look like tightly coiled springs—begin to unfold and link together, making the dough firmer and less squishy.⏎- At 310 degrees, the so-called Ma

  2. #2 peak

    cause people to become dehydrated more easily. Some over-the-counter allergy drugs reduce sweating, an important cooling mechanism. Some mental health medications make a person feel less thirsty, meaning he or<|endoftext|>Women who had been diagnosed with breast cancer need to

  3. #3 peak

    raw milk. Proponents of raw milk argue that food in it’s most natural form is always more nutritious, and that the heating process makes pasteurized milk less nutritious. This may be true, but since Americans get a wide variety of nutrients

  4. #4 peak

    nutrient depletion and to increase organic matter, which makes it less prone to drought, flooding and water-logging. The project has helped reduce pressure

  5. #5 peak

    "aged". This means that after they are picked, the beans are dried and then kept from three to eight years. This "aging" gives the coffee a less acidic taste. Coffee that has been aged is often mixed or "blended"

  6. #6 peak

    a speculative context. One important advantage is that such a setting can help dampen emotional responses and enable a more rational discussion. The speculative context helps make the discussion less threatening to some who might react with greater hostility to discussions focused on the actual world

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 a new problem that is more difficult, and I am not more confused about myself and myself in the same place of my own room, I am not more confused about myself in the same room in the same room, I am not

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

    generate various water shapes. At present, there are great differences in the quality of domestically produced nozzles. Compared with similar products abroad, the biggest problem is not only the appearance but also the design quality.⏎In actual use

  2. #2 peak

    skald or scholar. The dwarves reported to the Æsir that Kvasir had choked on his own shrewdness, since there was none so wise there as to be able to question his wisdom.⏎Kvasir also

  3. #3 peak

    3, he became one of the first to author the word aight in published text. He’s even glossed it “(Alright)” for everyone who hadn’t caught on yet.⏎We learn from the OED entry that aight represents a

  4. #4 peak

    work.⏎How to Get Back in the Routine⏎So you’re sold on keeping your 8:00 p.m. standing date — but your kid, not so much. To dull resistance:⏎1. Do it<|endoftext|>Our Fifth Grade students

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. L14:2441 strongly excites
  2. L10:111 strongly suppresses
  3. L13:383 strongly suppresses
  4. L9:442 strongly suppresses
  5. L9:54 strongly excites
  6. L4:9 excites
  7. L11:1095 suppresses
  8. L8:1395 excites
  9. L12:1976 suppresses
  10. L7:1819 suppresses
  11. L11:274 suppresses
  12. L14:2518 excites
  13. L14:1859 suppresses
  14. L10:823 suppresses
  15. L7:32 excites
  16. L13:642 excites
  17. L9:1472 suppresses
  18. L13:2809 excites
  19. L8:140 excites
  20. L14:211 suppresses
  21. L6:339 excites
  22. L11:658 excites
  23. L14:1632 suppresses
  24. L5:6 suppresses

Feeds later layers

  1. L16:1221 strongly excites
  2. L16:1692 strongly excites
  3. L16:445 strongly excites
  4. L16:447 strongly excites
  5. L16:728 strongly excites
  6. L16:697 strongly excites
  7. L16:1828 strongly excites
  8. L16:804 strongly excites
  9. L16:1997 strongly excites
  10. L16:2763 strongly excites
  11. L16:2898 strongly excites
  12. L16:2754 strongly excites
  13. L16:1850 excites verbs of reducing or harming
  14. L16:497 excites
  15. L16:279 excites
  16. L18:1792 excites
  17. L19:1301 excites
  18. L16:2907 suppresses
  19. L17:443 excites
  20. L17:206 excites
  21. L16:2561 excites
  22. L16:295 excites
  23. L19:2909 excites
  24. L16:964 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.

2 values between -5.04 and -4.653 values between -4.65 and -4.261 values between -4.26 and -3.864 values between -3.86 and -3.472 values between -3.47 and -3.071 values between -3.07 and -2.684 values between -2.68 and -2.292 values between -2.29 and -1.896 values between -1.89 and -1.507 values between -1.50 and -1.1046 values between -1.10 and -0.71936 values between -0.71 and -0.312539 values between -0.31 and 0.08198 values between 0.08 and 0.4755 values between 0.47 and 0.8725 values between 0.87 and 1.2621 values between 1.26 and 1.668 values between 1.66 and 2.0511 values between 2.05 and 2.4411 values between 2.44 and 2.844 values between 2.84 and 3.237 values between 3.23 and 3.635 values between 4.02 and 4.413 values between 4.41 and 4.812 values between 4.81 and 5.202 values between 5.20 and 5.601 values between 5.60 and 5.991 values between 7.57 and 7.961 values between 10.33 and 10.72offnegativepositiveoften

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. ent
  2. 任何
  3. anymore
  4. 有任何
  5. 打扰
  6. ▸
  7. icie
  8. any
  9. entar
  10. ative

Suppresses

  1. drivers
  2. awah
  3. @s
  4. 新一代
  5. zac
  6. never
  7. ikip
  8. sublic
  9. 岁以下
  10. igans