Neuron Explorer

Qwen3-0.6B

mid-word pieces in science words

A neuron in layer 7 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 responds weakly to the middle pieces of long scientific words that are split into parts, such as 'poll-in-ators' or 'macro-ph-ages'. Its strongest signals are small, and turning it up or down gave no clear theme.

Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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

    are to highlight an area of research that Google believes to be important and to gain real-world experience with the larger data structures that post-quantum algorithms will likely require.⏎We're indebted to Erdem Alkim, Léo

  2. #2 peak

    on these plants for precious nutrients.⏎These gardens are small and easy to maintain, but they play a vital role in the well-being of the surrounding environment. Pollinators are in trouble due to habitat loss, pesticide use and the increasing threat of

  3. #3 peak

    on crops that are not attractive to pollinators or for crop- production scenarios that minimize or eliminate potential exposure to bees. The registration is very protective of pollinators and includes fewer crops than were allowed under sulfoxaflor’s previous

  4. #4 peak

    against a loss of bee diversity and abundance. Our results will contribute to understanding the mechanisms that determine the role of native biodiversity on crop pollination and provide key information to farmers and conservation managers in the region.<|endoftext|>North Atlantic

  5. #5 peak

    cohabitation – laws that are not universally enforced because they also criminalize other behaviors that are otherwise socially sanctioned. However, some “fundamentalist” polygamists marry off girls before the age of consent or commit fraud in obtaining social and

  6. #6 peak

    now announced that they have found certain unique aspects to the fossil that definitively prove their hunch, and are able to reveal the precise statistics of the giant herbivore. It seems that the beast would have stretched for 20 meters

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 following is a statement that is not to be ignored as it is: 'No code' in the problem statement. Also, the statement is a direct or indirect statement about the given problem. So, the statement is a generalization

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

    as far as information content.⏎Even in the area of accumulating and using knowledge, we have advanced more in the digital age, which encompasses the last sixty years or so, than in all of previous human history. I don't see this as "

  2. #2 peak

    offea) and the name of the drink that is made from this plant. The coffee plant is a bush or tree that can grow up to ten meters (about 32 feet) high, but is usually cut shorter. Coffee plants originally

  3. #3 peak

    60%.⏎—The results support the hypothesis that genetic factors play a major role in the etiology of alcoholism in women. Women should be well represented in the efforts currently under way to elucidate the molecular basis of the genetic susceptibility

  4. #4 peak

    ans from Gulf War-era II were much less likely to be high school dropouts (2 percent) than were nonveterans (14 percent). About 25 percent of recent veterans had a college degree, as did a similar percentage

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. L5:1931 excites
  2. L6:295 excites
  3. L6:1877 excites
  4. L2:461 excites
  5. L6:377 suppresses
  6. L5:399 excites
  7. L6:1363 excites
  8. L5:1044 suppresses
  9. L5:656 excites
  10. L4:2375 suppresses
  11. L5:797 suppresses
  12. L6:289 excites
  13. L4:526 suppresses
  14. L4:361 weakly suppresses
  15. L6:510 weakly suppresses
  16. L5:328 weakly excites
  17. L6:2323 weakly excites
  18. L5:167 weakly excites
  19. L5:1521 weakly suppresses
  20. L5:1960 weakly excites
  21. L5:854 weakly suppresses
  22. L4:1185 weakly excites
  23. L6:2262 weakly suppresses
  24. L6:2250 weakly suppresses

Feeds later layers

  1. L8:228 suppresses
  2. L8:679 excites
  3. L12:2310 suppresses
  4. L13:1542 excites
  5. L10:332 suppresses
  6. L12:626 excites
  7. L14:770 suppresses
  8. L15:536 excites
  9. L9:164 suppresses
  10. L10:641 excites
  11. L8:868 suppresses
  12. L8:519 excites
  13. L8:623 suppresses
  14. L8:276 excites
  15. L11:216 suppresses
  16. L14:1740 excites
  17. L8:1895 suppresses
  18. L8:46 suppresses
  19. L13:616 suppresses
  20. L13:470 excites
  21. L12:503 suppresses
  22. L8:877 excites
  23. L9:1154 suppresses
  24. L13:2826 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.

1 values between -2.42 and -2.341 values between -2.34 and -2.262 values between -2.18 and -2.101 values between -2.02 and -1.942 values between -1.94 and -1.863 values between -1.86 and -1.783 values between -1.78 and -1.703 values between -1.70 and -1.621 values between -1.62 and -1.541 values between -1.54 and -1.4610 values between -1.46 and -1.389 values between -1.38 and -1.3010 values between -1.30 and -1.2215 values between -1.22 and -1.1415 values between -1.14 and -1.0617 values between -1.06 and -0.9821 values between -0.98 and -0.9034 values between -0.90 and -0.8250 values between -0.82 and -0.7434 values between -0.74 and -0.6682 values between -0.66 and -0.5788 values between -0.57 and -0.49100 values between -0.49 and -0.41140 values between -0.41 and -0.33178 values between -0.33 and -0.25232 values between -0.25 and -0.17261 values between -0.17 and -0.09387 values between -0.09 and -0.01520 values between -0.01 and 0.07467 values between 0.07 and 0.15420 values between 0.15 and 0.23357 values between 0.23 and 0.31242 values between 0.31 and 0.39135 values between 0.39 and 0.4749 values between 0.47 and 0.5511 values between 0.55 and 0.632 values between 0.63 and 0.714 values between 0.71 and 0.79offnegativepositiveoften

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. %',⏎
  2. %',
  3. -meta
  4. %",
  5. ###⏎
  6. 自治
  7. arb
  8. %",⏎
  9. much
  10. %D

Suppresses

  1. бит
  2. akra
  3. 请点击
  4. igated
  5. 黑马
  6. готов
  7. ypi
  8. /pi
  9. returnValue
  10. oque