first piece of a split word
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. 26 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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
It reacts to the opening piece of a longer word that is broken into parts, like 'Suk' in Sukhothai, 'Yuk' in Yukon or 'techn' in technologically. When it was turned down, the model kept coming back to words like 'mind' and 'eye', and when turned up it repeated 'interesting'.
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.
-
populations throughout the region often referred to as the 'deep south'.⏎The old ethnic Malay and Muslim Sultanate of Patani made tributes to the ancient kingdoms of Sukhoth
-
9), and the Piemonte-Toscana region of northern/central Italy (images only, February-March, 2009).⏎In Fort Yukon, Alaska, salmon is a mainstay of the diet. Yet, among
-
them, and this comes from close observation of the plants plus a large application of common sense. Just take one step at a time. Master the fundamentals before you succumb to the excitement of special techniques or the challenge of difficult subjects
-
new methods to make their classes more effective, active, technologic and funny. Games and technologic tools (such as; web tools digital
-
land⏎(1808–1867) Hannah was born and raised in a Quaker family in Sherwood, NY and is best recognized for her work as an abolitionist. The
-
the universe.⏎She has just completed 47 interviews back to back (Len Irish, who took our cover photo of Collins, says she was one of the nicest people he ever photographed – “no fuss”), there’s a television crew waiting
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: thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought thought
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
American democracy has changed A LOT over the centuries. Some people find this fact a little disconcerting, since we often take comfort in the notion that amidst the rapid social
-
behavorialist schools of psychology that reigned in his day, seeking a richer and more complex view of the mind than either provided. In a long, incisive essay, Brian Boyd celebrates the psychological acuity of Nabokov's
-
well as many other groups, highlights areas where some progress has been made, and lays out the paths toward significant solutions.⏎Some shocking statistics⏎These statistics provide some enlightening examples of what the report found — and the discrepancies are dramatic.⏎- African
-
. This makes it much harder for MRSA to develop a resistance. Ninja Polymers are nontoxic<|endoftext|>Watching an athlete choke can be one of the most difficult, and at the same time, engaging moments to witness. Like watching the
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
- L20:2747 excites
- L23:1241 excites
- L8:1049 suppresses
- L9:150 excites
- L2:2154 weakly suppresses
- L16:379 weakly suppresses
- L16:34 weakly suppresses
- L21:2954 weakly excites
- L5:609 weakly suppresses
- L1:14 weakly suppresses
- L23:270 weakly suppresses
- L25:2812 weakly suppresses
- L14:1631 weakly excites
- L19:134 weakly excites
- L15:2331 weakly suppresses
- L22:2088 weakly excites
- L7:3038 weakly excites
- L24:1299 weakly excites the word 'value' in the sense of worth
- L6:2675 weakly excites
- L25:2754 weakly suppresses
- L2:1918 weakly suppresses
- L20:541 weakly suppresses
- L9:2789 weakly excites
- L24:428 weakly suppresses
Feeds later layers
- L27:1297 weakly suppresses
- L27:313 weakly suppresses
- L27:2988 weakly suppresses capital M starting a word
- L27:1436 weakly excites
- L27:1447 weakly excites
- L27:1090 weakly suppresses
- L27:2117 weakly excites
- L27:1099 weakly suppresses
- L27:202 weakly suppresses
- L27:361 weakly suppresses
- L27:486 weakly excites
- L27:415 weakly suppresses
- L27:1912 weakly suppresses
- L27:809 weakly suppresses
- L27:859 weakly excites unclear: token just before a k sound
- L27:2925 weakly excites
- L27:1538 weakly excites
- L27:1214 weakly excites
- L27:2753 weakly suppresses
- L27:2226 weakly suppresses
- L27:1832 weakly suppresses
- L27:888 weakly excites
- L27:2652 weakly suppresses
- L27:2825 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
- 有趣
- eworthy
- utz
- (inter
- خاص
- interesting
- 有趣的
- interes
- Interesting
- df
Suppresses
- mind
- 眼
- 心
- Eye
- eye
- thought
- ⼼
- Eye
- 心灵
- heart