single letters starting names
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. 24 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 fires on a lone letter, often a capital S or I, that starts a proper name or word, such as 'Ithaca', 'Sardinia' or 'Anglo-Saxons'. Its negative side fires on capital S in names too. Steering shows no clear theme, only some repeated 'first' wording.
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.
-
of different penalties, including jail time, probation, fines and compensation to victims.⏎<|endoftext|>The second Principle of Natural Justice is Nemo Iudex In Causa Sua which means that a person is not allowed
-
Not always successful, and fraught with problems, but relevant enough for an intellectual tradition to be discovered. The latter can be defended, but the former demands us to jettison its nationalistic undertones, and implicitly acknowledges the complexities of nationalism that
-
newer populist members of his party, prompting Liberal Party elites to<|endoftext|>By Francesca De Chatel⏎"There is none amongst the believers who plants a tree, or sows a seed, and then a bird, or a person, or an animal
-
the present.⏎By asking why the Easter Rising occurred, Ward is able to review the history of Anglo-Irish relations, from the time of Norman settlement to World War I, as well
-
, pre-Christianity, were all based on Earth and planetary cycles, most had a spring goddess. The spring goddess in Scandinavian was "Ostra," for Anglo-Saxons it was "Eostre," and the current region of Germany she
-
of the church.⏎Thus started the great persecution of 235. Both Pontian and Hippolytus were arrested and exiled to the island of Sardinia to work in the mines. This was considered a death sentence. Pont
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 fact that a single individual with a single bit of information has the power to design an entire design to something he must design something. But this design must be something that he design to something he must design something, which must be something that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
use? This column by Ramachandra Guha in The Hindu traces the phrase back to a seminal 1955 essay by the sociologist M N Srinivas titled 'The Social System of a Mysore Village', quoting
-
molecular test is being used as the gold standard method for detecting SARS-CoV-2, its complexity limits its use at the point of need
-
Gender Impacts of Mobile Money’ written by US-based economists Tavneet Suri and William Jack, and published in the prestigious journal Science. The critique,
-
pediatric hematologists and those who practice in tertiary care centers.⏎Gene therapy offers promise as a curative therapy for SCD but is not yet a reality. Stem cell transplantation has successfully cured many severely
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
- L22:666 suppresses
- L22:1434 suppresses
- L20:435 suppresses
- L21:1326 excites capital 'S' starting a name
- L19:1057 weakly suppresses
- L17:786 weakly suppresses
- L21:1492 weakly excites
- L23:678 weakly excites
- L22:1084 weakly excites
- L1:341 weakly suppresses
- L13:2503 weakly suppresses
- L1:5 weakly suppresses
- L4:1802 weakly suppresses
- L17:367 weakly suppresses
- L16:84 weakly suppresses
- L21:38 weakly excites
- L1:2922 weakly excites
- L1:2695 weakly excites
- L14:775 weakly suppresses
- L22:151 weakly suppresses
- L20:1964 weakly suppresses
- L9:37 weakly suppresses
- L20:669 weakly suppresses
- L17:1992 weakly excites
Feeds later layers
- L26:941 weakly excites
- L25:150 weakly suppresses
- L26:1342 weakly suppresses
- L27:1915 weakly suppresses
- L25:1398 weakly suppresses
- L26:579 weakly excites
- L26:1655 weakly excites
- L26:232 weakly excites
- L25:2303 weakly suppresses
- L25:1749 weakly excites
- L25:1402 weakly excites
- L25:2467 weakly excites
- L27:1701 weakly suppresses
- L27:633 weakly excites
- L25:1978 weakly suppresses
- L25:1949 weakly suppresses
- L25:1727 weakly excites
- L26:790 weakly suppresses
- L26:729 weakly excites
- L27:1091 weakly excites
- L27:1076 weakly excites
- L27:2293 weakly suppresses
- L27:2432 weakly excites
- L26:2747 weakly suppresses first letters fl, w and ew
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
- \Annotation
- rors
- olina
- ACKET
- oline
- uela
- queryInterface
- ól
- cout
- oram
Suppresses
- ivers
- 备
- enco
- ivering
- tribution
- 体
- świadczenie
- pedig
- 证据
- 开关