Neuron Explorer

Qwen3-0.6B

'should' in advice and claims

A neuron in layer 17 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 strongly on the word "should", especially in phrases like "should not be confused" or "should be diagnosed". Turning it up made the text more likely to shift into "should" statements about what ought to be done, but only in one of the four prompts.

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

    the Polish Cossacks (Kozacy) and the Tatar Cossacks (Nağaybäklär). The name “Cossack” should not be confused with Kazakhs (native to Kazakhstan, a country in Central

  2. #2 peak

    of the major components of the forward price index methodology. Variable cost index also known as VAT, is often confused with the expression “cost of good sale” which should not be confused as the same thing. The main difference is that the first expression

  3. #3 peak

    facial swelling) and respiratory (e.g. wheezing) issues. In severe cases, anaphylaxis can occur.⏎Cow’s milk allergy should be diagnosed by a registered clinical expert who will assess the detailed clinical history in combination

  4. #4 peak

    that the influenza A virus can be transmitted to mammalian species such as tigers and domestic cats and dogs through ingestion of infected birds. Another group of animals that should also be considered as potential hosts for the influenza A virus is the crocodilians

  5. #5 peak

    Siberian Yeti implanted by the Russians as a spy, I would suggest that was satire. Both the source and the content suggest that it is not something which should be believed.⏎On the other hand, when a Fox news anchor claimed he

  6. #6 peak

    reference to the darker, streaked heartwood found in some Ash trees, which tends to resemble the wood of Olive trees in the Olea genus. And it should come as little surprise that Olive Ash can be a dead ringer for actual Olive

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 want to have the same number of 100000011110011110011100001000100010010

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

    grams of protein. It also gives details per 100g: 3.2g of protein. Aha! I could - if I needed to, figure out how much a portion weighs using proportion! It’s time for the

  2. #2 peak

    habits like too much taking of processed food and large quantities of fried food ,which can harm your liver in the long run.⏎- So, the liver requires a proper amount of vitamins, minerals, and antioxidants to improve its functions. The role of

  3. #3 peak

    the disabled, the unemployed and similar groups of people are to be taken care of in a welfare state. It has the duty of eliminating economic inequalities to ensure a reasonable and equitable standard of living, to all citizens of the state.⏎Education, health

  4. #4 peak

    is going in. Education is serious, many other countries are way ahead of us and we are falling behind. We have rising educational costs with colleges, which require effective solutions to ensure that we have educated adults entering the work force, without the funds

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. L4:2226 suppresses
  2. L7:1569 excites
  3. L4:2008 excites 'to' after 'have got' or 'will have' (obligation)
  4. L16:2349 suppresses
  5. L15:886 excites
  6. L5:371 excites
  7. L10:2166 excites
  8. L9:640 suppresses
  9. L5:2538 excites
  10. L5:387 excites
  11. L13:1272 excites
  12. L10:1125 suppresses
  13. L13:909 excites 'to' before a verb after a noun
  14. L11:2224 excites
  15. L14:1309 excites
  16. L7:1056 excites
  17. L3:3037 suppresses
  18. L10:2233 suppresses
  19. L10:562 excites
  20. L16:2304 suppresses
  21. L14:2077 weakly excites
  22. L5:1917 weakly excites 'to' after 'what' or 'where' (what to do)
  23. L9:1848 weakly suppresses
  24. L6:2057 weakly suppresses

Feeds later layers

  1. L18:522 excites
  2. L18:1162 excites
  3. L19:1709 suppresses
  4. L19:1294 excites
  5. L21:1906 excites
  6. L21:2990 suppresses
  7. L18:2568 excites
  8. L19:1824 excites
  9. L20:2395 excites
  10. L27:1123 suppresses
  11. L19:2282 weakly excites
  12. L18:1481 weakly suppresses
  13. L18:2733 weakly excites
  14. L18:700 weakly suppresses
  15. L24:764 weakly suppresses
  16. L21:2252 weakly excites
  17. L20:2134 weakly suppresses
  18. L20:2858 weakly excites
  19. L22:1487 weakly excites
  20. L19:2433 weakly excites
  21. L18:717 weakly excites
  22. L21:1575 weakly excites
  23. L19:1850 weakly excites
  24. L22:2001 weakly suppresses

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.

179 values between -1.82 and -0.893509 values between -0.89 and 0.05151 values between 0.05 and 0.9823 values between 0.98 and 1.9111 values between 1.91 and 2.845 values between 2.84 and 3.773 values between 3.77 and 4.705 values between 4.70 and 5.634 values between 5.63 and 6.562 values between 6.56 and 7.493 values between 8.42 and 9.351 values between 9.35 and 10.283 values between 11.21 and 12.142 values between 12.14 and 13.071 values between 14.93 and 15.861 values between 18.66 and 19.591 values between 22.38 and 23.311 values between 23.31 and 24.241 values between 29.82 and 30.751 values between 30.75 and 31.681 values between 34.48 and 35.41offnegativepositiveoften

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. tein
  2. 考虑
  3. 结晶
  4. ndo
  5. be
  6. �性
  7. ��
  8. beware
  9. 恙
  10. .clip

Suppresses

  1. 需要
  2. 需
  3. orer
  4. 你需要
  5. :::
  6. bor
  7. filtr
  8. needs
  9. 都需要
  10. oris