Neural networks effectively represent words as points in vector space
Categorization
- corroborated: 2 independent passages support this claim
- XREF: connects to: Training word vectors solely to predict context words yields powerful meaning representations; The second channel — because embeddings miss the best links
Nearest-neighbor retrieval systematically finds the obvious connections and systematically misses the valuable ones. Two claims that contradict each other are often not lexically similar. Cross-domain analogies — the single highest-value connection type — are by definition far apart in embedding space. A system built on cosine similarity alone will quietly hand you a graph of things you already knew were related.
Two additions fix this:
Entity channel. Extract named entities and key concepts per claim; index them; generate candidates on shared entities regardless of embedding distance. Catches Merzenich turning up in a neuroscience book and a pedagogy book, where the claims share no vocabulary.
Weekly deep pass (deep-link). Sunday 03:00. Sample claim pairs from distant clusters
and ask Opus 4.8 for structural analogy — same shape of argument, different domain. Low
volume (a few hundred pairs), high value, and it’s the only part of the system that can find
a connection neither channel was looking for. Budget a few dollars a week.
Contradictions get a lower threshold than corroborations. Asymmetric on purpose: a false positive contradiction costs you five seconds to dismiss; a missed one costs you a wrong belief you keep citing.; Machine pattern recognition mirrors Gestalt grouping via feature vectors and similarity metrics
Library Connections
- Training word vectors solely to predict context words yields powerful meaning representations
#exemplifies(0.95) - CLAUDE
#complicates(0.90) — The second channel — because embeddings miss the best links
Nearest-neighbor retrieval systematically finds the *obviou…