the 'w' in shipwreck
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. 27 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 mostly fires on the letters 'w' in the word shipwreck, where the word is split into pieces. It goes strongly negative on a 'w' that starts other words, such as wax, wick and warts. Turning it up made the model drift into repeated words like 'product' and 'prediction'.
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.
-
/proclivities, etc. Psychometric profiles of civilian<|endoftext|>Recently we were called out to investigate a remnant shipwreck that had materialized on the shoreline near Georgetown, Prince Edward Island. The
-
alike. As for us, we were delighted and slightly surprised at the surge of interest. While we, too, were excited by the chance to investigate a shipwreck, for us it presented a unique archaeological site that required systematic documentation in order
-
m) wide. Among all of the 37 shipwrecks discovered at Yenikapi, there were a total of six Byz
-
found at a single location," said study author Cemal Pulak, of the Institute of Nautical Archaeology at Texas A&M University.⏎The eight shipwrecks highlighted in the new report, which date from the seventh to 1
-
, and even being raised from the sea.⏎Students could conduct a Rapid Research topic where groups look into 17 famous shipwrecks:⏎- The Mary Rose, 1545⏎- The
-
1821, through military action, foundering, grounding, fire and explosion, deliberate sinking to create artificial reefs, and other causes. Such shipwrecks are the major focus of the Florida Navy Legacy Shipwreck Project.⏎
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: since my own ownaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesaclesacles
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
phonon charge-density wave coupling.<|endoftext|>Soy Wax vs. Paraffin Wax:⏎Candles can be made with different types of wax. The most common waxes used are paraffin and soy. Paraffin wax is a refined
-
exhibit. He has more prominent “warts” on his face, and his tusks are more apparent. Spartacus is still a pretty young guy, so those warts and tusks will get bigger as he gets older. He is also
-
-floating wick. Instead, it used a flame-enclosing burner and a wick bent into a cylinder shape, which provided the fire with just enough air.
-
on that migration and hence every child of Adam (as), weeps when he opens his eyes in this world. If this is true, then this reason for wailing certainly supports our point of view that in the grief of a departure, w
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
- L22:2646 excites
- L20:417 excites
- L18:533 suppresses
- L23:400 excites
- L21:463 suppresses
- L0:3043 suppresses
- L23:206 suppresses
- L7:421 excites
- L18:365 excites
- L18:569 excites
- L16:67 weakly suppresses
- L2:2142 weakly suppresses
- L8:13 weakly suppresses
- L19:1015 weakly excites
- L26:1564 weakly excites
- L8:7 weakly excites
- L22:1511 weakly suppresses
- L6:13 weakly excites
- L13:810 weakly suppresses
- L19:196 weakly excites
- L7:287 weakly excites
- L8:4 weakly excites
- L22:1550 weakly suppresses
- L6:16 weakly excites
Feeds later layers
None: this is the last layer.
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
- age
- -product
- yz
- product
- 味
- -products
- edd
- orn
- 效力
- ep
Suppresses
- ähnlich
- utation
- oose
- まして
- 쟁
- acles
- oteca
- 饵
- endwhile
- ница