end-of-document token and code keywords
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. 25 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly 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 fires on the marker that separates one web page from the next, right where a new unrelated document starts. Turning it up made the model write code words like 'if', 'while' and 'else' over and over, so this is a mixed signal.
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 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.
-
wellandgood.com.<|endoftext|>The Technology Behind Bella Hadid’s Spray on Dress⏎During the 20
-
locate.<|endoftext|>Science Fair Project Encyclopedia⏎The World Factbook⏎The World Factbook is an
-
comment.<|endoftext|>100 Chemistry Questions and Answers 8211; Basic Chemistrynbsp
-
ycetica<|endoftext|>Adding Fractions Level 3⏎The 5-step planIn Adding Fractions
-
Project:<|endoftext|>After years of experience and "combing through" catalogs⏎to find the best
-
_to_eat_gmos_so_why_are_we/<|endoftext|>A PhD can open a lot of job opportunities. Most of those jobs are academic
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 am still a student, I am still a student, I am still a student. I am still a student. I am still a student. I am still a student. I am still a student. I am still a student.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
print "This might take a while..."⏎> 4 for i = 1 to n⏎> 5 for j = 1 to i⏎> 6 print
-
and air conditioning systems, it also means that pollutants that originate indoors are getting trapped inside.⏎“People normally<|endoftext|>Spring Dawn by Meng Haoran 春眠不觉晓 处处闻啼茑 夜来风雨声 花落
-
_arr = [1, 2, 3, 4, 5] i = len(given_arr) - 1 while(i >= 0): print(given_arr[i]) i = i - 1⏎In this program
-
Matlab Basics II - Create a vector with a repeated entry⏎Create sequnce 1 4 9 16 25.........⏎Basics: Divide integers to get integer outputs in all cases⏎Determine value using Look-up
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
- L21:131 excites
- L19:1372 suppresses
- L17:2265 excites
- L18:284 suppresses
- L18:3070 excites
- L10:149 excites
- L21:1988 excites
- L22:633 suppresses
- L19:719 excites
- L17:95 suppresses
- L18:569 weakly excites
- L22:1096 weakly excites
- L20:1635 weakly excites
- L6:1800 weakly excites
- L19:415 weakly suppresses
- L14:67 weakly suppresses
- L22:2020 weakly excites
- L20:2632 weakly suppresses
- L23:2258 weakly excites
- L20:96 weakly suppresses
- L20:890 weakly suppresses
- L15:524 weakly excites
- L6:7 weakly suppresses
- L10:1995 weakly suppresses formal nouns after verbs of answering or committing
Feeds later layers
- L27:110 suppresses
- L27:1150 excites
- L27:791 excites
- L27:2076 excites
- L27:338 weakly excites
- L27:511 weakly excites
- L26:2856 weakly suppresses
- L27:94 weakly suppresses
- L27:201 weakly excites
- L27:274 weakly excites document boundary, code keywords next
- L27:760 weakly excites
- L26:82 weakly suppresses
- L27:463 weakly excites
- L27:146 weakly suppresses
- L27:1908 weakly excites the 'd' that starts 'die' or 'dead'
- L27:2484 weakly excites
- L26:1683 weakly suppresses
- L27:1666 weakly excites
- L27:1286 weakly excites
- L26:118 weakly suppresses
- L26:2540 weakly excites
- L27:1335 weakly excites
- L26:884 weakly suppresses
- L26:2462 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
- else
- else
- try
- try
- while
- raise
- assert
- while
- break
- return
Suppresses
- ankind
- imated
- inent
- rending
- ừa
- lemen
- ennessee
- étant
- /browse
- lotte