unclear: weak sentence-end punctuation
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 16 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron is weak compared with the others. It sometimes fires on full stops that end a clause or on the start of a list item, but the signal is faint and inconsistent. Steering did not show a clear effect.
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 activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
would be measured in distance units, such as meters. To indicate this measure, the arc is preceded by the lower case letter L (for 'length'). For eaxmaple⏎would be read as "the length of the arc AB
-
]⏎Cite This Essay⏎(2012, 05). The Struggle for Power in the Ussr. StudyMode.com. Retrieved 05, 2012, from http://www.studym
-
the fan blade by hand. If the fan blade does not turn freely, replace the fan motor. Additionally, if the motor is unusually noisy, replace it. Finally, if the motor does not run at all, use a multimeter to test
-
include linear algebra, calculus, or finite mathematics, although there are other possibilities. ⏎Students are expected to develop their understanding of programming beyond the skill level (that
-
you are pregnant or breastfeeding, because it may trigger a miscarriage.1 Yet, most people would not think to look into the safety of drinking a
-
promoting the rule of law, ensuring access to justice, and strengthening institutions. It is the intersection of SDG 11 and SDG 16
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
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.
Here is something I have been thinking about lately: "I need to know it, I need to know it, it is for me." It is, but it is for me, it is, but, it, as well, I, It, as, Well, as,,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
have successfully created an account.⏎A verifications link was sent to your email at . Please check your email and click on the link to confirm your email address and fully activate your iCPALMS account. Please check your spam folder.<|endoftext|>W
-
Planum temporale|⏎|Approximate location of Wernicke's area highlighted in gray|⏎| This article needs additional citations for verification.|⏎Please help improve this article by adding reliable references. Unsourced material may be challenged and removed.
-
subject and verb must agree with one another in their tense. If the subject is in plural form, the verb should also be in plur al form (and vice versa).⏎Which is better or which one is better?⏎As your sentences mean the same
-
Copyright © University of Cambridge. All rights reserved.⏎'The Solid' printed from http://nrich.maths.org/⏎is one of a series of problems designed to develop learners' team⏎working skills. Other tasks in the series can be found
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L11:635 strongly suppresses
- L15:1111 excites
- L9:173 suppresses
- L15:1229 suppresses
- L14:2266 suppresses
- L15:703 weakly suppresses
- L4:2073 weakly suppresses
- L10:196 weakly excites
- L10:1263 weakly excites
- L8:254 weakly excites
- L4:19 weakly suppresses
- L4:1158 weakly excites
- L2:1075 weakly suppresses
- L11:2971 weakly suppresses
- L2:3035 weakly excites
- L10:2455 weakly suppresses
- L8:643 weakly excites
- L1:1742 weakly excites
- L7:657 weakly suppresses
- L8:1912 weakly suppresses
- L15:465 weakly excites
- L2:986 weakly excites
- L6:16 weakly excites
- L1:1394 weakly excites
Feeds later layers
- L18:1261 suppresses
- L17:1558 suppresses
- L17:1752 suppresses
- L17:297 excites
- L17:916 weakly suppresses
- L23:733 weakly suppresses
- L24:2694 weakly suppresses
- L26:2461 weakly excites
- L23:438 weakly suppresses
- L17:1554 weakly suppresses
- L17:555 weakly excites
- L21:1936 weakly excites
- L17:2461 weakly suppresses
- L20:843 weakly suppresses
- L26:1822 weakly excites
- L18:1241 weakly excites
- L19:848 weakly excites
- L18:2502 weakly excites
- L23:1040 weakly suppresses
- L21:442 weakly excites
- L19:727 weakly suppresses
- L25:1320 weakly excites
- L21:2164 weakly suppresses
- L17:247 weakly suppresses
Activation histogramA chart of how often a neuron takes each activation value.
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.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
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 layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- doctype
- rise
- rong
- rv
- روف
- ruk
- ACTER
- thood
- StreamWriter
- rag
Suppresses
- ию
- iness
- ;">
- eworld
- apers
- ined
- →⏎→⏎
- quine
- entral
- icing