How "attraction in Paris" lights up composite neurons without entity units
lucia_quirke · x · 2026-10-18
luciaquirke offers a concrete mechanism in an interpretability debate: the prompt "What is an attraction in Paris?" could activate the neuron for (France, Paris, Eiffel, Country) and the neuron for (attraction, Eiffel, Golden Gate Bridge), with their boosts hitting the "Eiff" token hardest — in this ontology, entities don't exist as such.
The context: livgorton questions how entity-level information can be read in downstream circuits without dedicated entity representations, given monosemanticity and no superposition.
Related event: Monosemanticity debate: Does neuron count cap what a model can know?(8 posts)→