'of' after 'property' or 'feature'
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. 1 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 word 'of' when it comes right after nouns like 'property', 'feature' or 'characteristic', as in 'a property of the system'. Turning it up did not produce a clear change in behaviour that matches this idea.
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.
-
The compiler takes care of that.⏎> Another example is to imagine a Tree in which the leaves contain⏎> "objects." If I want to change a property of one of those leaf objects,⏎You can't in general, the thing with a
-
expressions⏎Integer Practice Problems - Addition⏎Review and Practice Problems for adding positive and negative numbers⏎<|endoftext|>In physics, a conservation law states that a particular measurable property of an isolated physical system does not change as the system evolves over time. Exact conservation
-
System (CTS):⏎Each programming language designer has to specify a very important property of it, which is called type system. The type system indicates which data type are
-
on their own behalf and through their own stores. The dealer can independently set the final price of the goods or use the manufacturer’s recommended price.⏎An important feature of the dealer is that he takes the last place in the commodity exchange
-
of a pyruvate kinase splice-variant shows therapeutic promise⏎Cold Spring Harbor, N.Y.⏎– Cancer cells grow fast. That’s an essential characteristic of what makes them cancer cells. They’ve crashed through all the cell-cycle checkpoints and
-
cells, or cells sensitive to different wavelength ranges, the organism would theoretically be able to perceive colour or at the very least colour differences. Polarisation, another property of EM radiation, can be detected by some organisms, with insects and cephalop
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: the state of a country that is a country. So, the country is a country. So, the country is a country. So, the country is a country. So, the country is a country. So, the country is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Legion of the Seventh Seal and the First Shogun. As to whether it was a foreign invasion or a civil war it is not known outside of Hachiman, for Hachiman has been isolated for many years, the only contact
-
azzucato. Copyright © 2018. Available from PublicAffairs, an imprint of Perseus Books, LLC, a subsidiary of Hachette Book Group, Inc.⏎The January 2010 edition of The
-
to the south-west Khulian name for the God of Battle Rage, Oiden. Perhaps some ancient link can be inferred.⏎The Shogunate of Hachiman is just over five hundred years old. A huge battle took place around
-
ogun is not known for sure. The first Shogun was still a good 800 years after the Splitting of Irritaria and Hachiman was probably ruled by an Emperor/ress before the Shoguns took complete
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
- L0:546 weakly suppresses
- L0:886 weakly suppresses
- L0:1228 weakly excites
- L0:979 weakly excites
- L0:2820 weakly excites
- L0:178 weakly suppresses
- L0:2606 weakly excites
- L0:2621 weakly suppresses
- L0:2113 weakly excites
- L0:2781 weakly excites
- L0:1370 weakly excites
- L0:1301 weakly suppresses
- L0:1224 weakly suppresses
- L0:1745 weakly suppresses quantity word after large, high or low
- L0:84 weakly suppresses
- L0:283 weakly excites
- L0:2975 weakly excites
- L0:1877 weakly excites
- L0:614 weakly suppresses
- L0:2333 weakly suppresses
- L0:1543 weakly suppresses
- L0:2456 weakly excites
- L0:1630 weakly suppresses
- L0:2207 weakly suppresses
Feeds later layers
- L10:1385 weakly suppresses
- L7:845 weakly suppresses
- L6:1201 weakly excites
- L14:962 weakly suppresses
- L14:980 weakly excites
- L7:543 weakly excites
- L2:1732 weakly suppresses
- L25:1380 weakly suppresses
- L2:674 weakly suppresses
- L24:991 weakly suppresses
- L6:3032 weakly suppresses
- L15:1879 weakly excites
- L2:813 weakly excites
- L13:2873 weakly suppresses
- L11:1320 weakly suppresses
- L21:1579 weakly excites
- L9:1198 weakly suppresses
- L2:1755 weakly excites
- L9:1228 weakly excites
- L9:956 weakly excites
- L17:946 weakly suppresses
- L10:902 weakly excites
- L13:2099 weakly suppresses
- L13:1097 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
- emain
- 勋
- ess
- едь
- applications
- предус
- \data
- verständ
- ronic
- McDon
Suppresses
- lake
- +-+-+-+-+-+-+-+-
- овал
- owski
- sposób
- opi
- 办法
- 之心
- hma
- ubit