'the' inside a mid-sentence phrase
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. 20 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 'the' when it comes in the middle of a sentence and a noun phrase is about to follow, such as 'as the ship' or 'without the owner'. The pattern is loose, and the steering outputs don't clearly show a clear theme.
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.
-
so by including a specific date or a triggering event on which the trust is to terminate. By way of illustration, the Settlor might include a term requiring the trust to terminate upon the marriage of the trust’s only beneficiary or upon the youngest
-
from Southhampton, England on April 10, 1912.⏎This is one of the last photographs of Titanic taken from the shore as the ship departed towards New York City.⏎This photo shows some of the workers from the
-
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 laws include conservation of energy, conservation of linear
-
a minimum, they should be cognizant that it is malicious software that collects personal information, usually with the intent of sending that information to third parties, without the owner's knowledge or consent. If anyone would like to learn the historical, technical
-
who they claim to be. Typical ways of doing this include asking for:⏎- Knowledge – Information only the user would know. Potentially a personal question about a favorite book or similar⏎
-
, to be much more economically efficient.<|endoftext|>Sepsis is a life-threatening illness caused by your body’s response to an infection. It develops when the chemicals the immune system releases into the bloodstream to fight an infection cause inflammation throughout the entire body
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: A *this* *meat* the *this* *the* *the* *the* *the* *the* *the* ** ** *the* ** ** ** * * * **
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.<|endoftext|>DualWingGenerator is a power generation with the wing-beating principle. Inspired by birds this flapping generator could make modern wind turbines a thing of the past⏎German company Festo has unveiled DualWingGenerator to generate electricity,
-
This image was based upon the head of the great statue of Helios, better known as the Colossus of Rhodes, one of the Seven Wonders of the Ancient World. The Colossus memorialized the war with Demetrios Pol
-
reveal the 10 objects they have chosen to tell a history of Manchester and its place in the world.⏎The objects are part of the wider A History of the World project, a unique partnership between the BBC, the British Museum and 3
-
help find out more about all the servicemen whose images make up the online exhibition but about whom very little else is known.⏎The collection of photographs - Faces of the First World War - is part of the IWM's preparations to mark the cent
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
- L18:861 strongly suppresses
- L19:2283 strongly suppresses
- L19:1023 excites
- L19:644 excites
- L17:372 suppresses
- L19:695 suppresses
- L18:1498 excites
- L17:30 suppresses
- L18:1359 excites
- L17:507 suppresses 'of' after a noun
- L16:1320 suppresses
- L16:419 excites
- L12:435 suppresses
- L19:2894 excites
- L16:792 excites "the" inside a prepositional phrase
- L8:636 suppresses
- L18:1513 suppresses
- L18:1041 suppresses
- L16:1851 suppresses
- L9:24 excites "the" before a named organisation or title
- L19:466 suppresses
- L19:1305 suppresses unclear: verbs after a noun or 'that'
- L18:1909 suppresses
- L17:56 suppresses
Feeds later layers
- L25:525 excites
- L21:843 excites
- L21:713 suppresses
- L21:1119 excites
- L23:24 suppresses
- L21:309 suppresses
- L21:696 excites
- L21:2279 suppresses
- L23:436 excites
- L21:920 excites
- L25:284 suppresses
- L27:335 excites
- L27:592 weakly excites
- L24:2040 weakly excites
- L27:1038 weakly excites
- L23:79 weakly suppresses
- L21:1227 weakly excites
- L27:533 weakly excites
- L22:1318 weakly suppresses
- L22:303 weakly suppresses
- L27:687 weakly excites
- L24:2454 weakly excites
- L27:449 weakly excites
- L24:268 weakly excites
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
- server
- 墟
- udad
- website
- tern
- device
- pherd
- 公司在
- company
- .server
Suppresses
- тики
- features
- Messaging
- (features
- amu
- gets
- lots
- /features
- тик
- ockets