Preprint describes million-scale LLM-native knowledge base

A new preprint describes GPTKB 2.0, a method for building a disambiguated knowledge base directly from large language models. The authors report a materialized database with more than 1 million disambiguated entities and 38.4 million triples, using on-the-fly disambiguation of entities, relations and classes. They characterize trade-offs among accuracy, scale and cost, and say the method and associated resource are available online. The work has not been independently validated.