Neuron Explorer

Qwen3-0.6B

'that' after nouns of belief or fact

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

The neuron lights up on the word 'that' when it introduces a clause after a noun or verb about a belief, fact, or sign, like 'a clue that this is happening' or 'a belief that underlies'. The opposite side fires on verbs like 'say' or 'indicates' that come just before an expected 'that'. Turning it up did not give a clear theme in the writing.

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

    hormones which release stored sugar from the liver into the blood. The blood sugar may remain high for up to 3 days after a low blood sugar. A clue that this is happening is seeing 1 to 3 days of high blood sugars with

  2. #2 peak

    a trendy restaurant. Suddenly, my simple domestic pleasures paled in comparison with the things I could be doing."⏎Grohol said that fear is "a legitimate sense that some people have. But I think it's sort of a false dichotomy.

  3. #3 peak

    sealing components implemented that prevents all ice dams and ice formations.<|endoftext|>Machu Picchu, Peru⏎Everything about Machu Picchu makes you marvel that it ever came to exist. The lost city of the Incas is built on

  4. #4 peak

    very anxious that perhaps he had committed some error because of which his Lord had become angry with him and had forsaken him. Thereupon he was given the consolation that revelation had not been stopped because of some displeasure but this was necessitated by

  5. #5 peak

    guerilla war they really had no hope of winning but did.⏎Part of the reason they won…other than sheer suicidal bravery and determination…was another belief that underlies the two involving women discussed here: Resurrection.⏎I came to identify

  6. #6 peak

    all Akali dal factions including their popular preacher- Bhindranwale. One instance led to another and there was burgeoning mistrust amongst the average Sikh mind that their community was not trusted by the Indian government.⏎During this course, the Ak

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: that of a 3-digit number, that of a 2-digit number, that of a 1-digit number. That also, of a 3-digit number, that of a 2-digit number, and that of a 1

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

    lack of academic resources, or poor mentorship and guidance – the reason for all is somehow rooted in the COVID-19 pandemic. Of course, you might say that

  2. #2 peak

    start the next paragraph, list, heading or whatever.⏎Strictly speaking a page should start <HTML> and should end </HTML>, but the HTML specification also says that clients should perform correctly without them, and so many people omit them. Similarly

  3. #3 peak

    ?⏎Based on our data, we think this problem is relevant for Professor Finnegan's class at WSU.⏎What textbook is this problem found in?⏎Our data indicates that this problem or a close variation was asked in Chemistry: An Atoms First

  4. #4 peak

    is not a nation. Indeed,⏎Constitutional experts say the motion carries no legal weight and would be unlikely to give Quebec separatists a

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. L11:1774 excites
  2. L16:1482 excites
  3. L6:2313 excites
  4. L5:3034 suppresses
  5. L5:1807 suppresses
  6. L8:2146 excites "as" in "it seems as if" phrases
  7. L12:1380 excites
  8. L10:251 suppresses
  9. L3:1652 excites
  10. L8:369 suppresses 'of' after 'learned' or 'because'
  11. L13:1427 excites
  12. L9:2247 weakly suppresses
  13. L4:2399 weakly suppresses
  14. L6:1105 weakly suppresses 'that' after a finding or surprise verb
  15. L16:2913 weakly excites
  16. L14:1876 weakly excites
  17. L14:1128 weakly excites
  18. L7:73 weakly excites
  19. L16:2461 weakly excites
  20. L14:2460 weakly suppresses
  21. L3:2989 weakly excites
  22. L16:2204 weakly excites
  23. L7:1368 weakly excites
  24. L7:2656 weakly excites

Feeds later layers

  1. L18:1389 strongly suppresses
  2. L19:1177 suppresses
  3. L23:914 suppresses
  4. L26:586 suppresses
  5. L18:1617 suppresses
  6. L20:850 suppresses
  7. L20:1043 suppresses
  8. L22:631 suppresses
  9. L18:1519 suppresses
  10. L20:2237 excites
  11. L21:639 suppresses
  12. L21:918 suppresses
  13. L20:1959 suppresses
  14. L20:982 excites
  15. L27:616 suppresses
  16. L18:3010 suppresses 'to' before 'be' after considered
  17. L22:1186 suppresses
  18. L26:86 excites
  19. L27:717 weakly suppresses
  20. L18:229 weakly excites
  21. L23:298 weakly suppresses
  22. L25:234 weakly suppresses
  23. L24:306 weakly excites
  24. L21:191 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.

2 values between -31.31 and -30.441 values between -22.59 and -21.721 values between -21.72 and -20.841 values between -20.84 and -19.971 values between -19.10 and -18.232 values between -15.61 and -14.741 values between -14.74 and -13.862 values between -13.86 and -12.991 values between -12.12 and -11.252 values between -11.25 and -10.371 values between -9.50 and -8.632 values between -8.63 and -7.763 values between -7.76 and -6.881 values between -6.88 and -6.016 values between -6.01 and -5.144 values between -5.14 and -4.276 values between -4.27 and -3.396 values between -3.39 and -2.5214 values between -2.52 and -1.6537 values between -1.65 and -0.781878 values between -0.78 and 0.101849 values between 0.10 and 0.9776 values between 0.97 and 1.847 values between 1.84 and 2.724 values between 2.72 and 3.59offnegativepositiveoften

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. ar
  2. DownLatch
  3. celed
  4. orne
  5. 决定
  6. رون
  7. allele
  8. _PTR
  9. ond
  10. 小鸟

Suppresses

  1. that
  2. that
  3. →that
  4. bahwa
  5. _that
  6. That
  7. That
  8. 、
  9. rằng
  10. że