I have long wanted to aggregate all of human knowledge into a form that is not merely accessible, but directly usable: organized as completely and coherently as the knowledge itself permits, searchable from any direction, and able to evolve as our understanding changes. Google made the world’s information findable; Wikipedia organized an extraordinary portion of it; language models made much of it conversational. None has produced a living, explicit model of what humanity knows - one that can represent not only facts and concepts, but evidence, disagreement, argument, causality, uncertainty, and the changing relationships among them. The same need recurs at smaller scales throughout my work: every serious research, writing, or discovery project must somehow turn everything it learns into a structure that humans and machines can inspect, challenge, search, and build upon. I believe automated graph construction offers a path toward that larger ambition.
The difficulty is that knowledge graphs are generally built for worlds whose structure has already been decided. A domain is bounded, an ontology specifies what may exist within it, and information-extraction systems search source material for entities and relationships that fit the prescribed schema. This works when the questions are known and the relevant structure can be designed in advance. It cannot scale to all human knowledge - or even to the changing needs of a single ambitious research project. The same body of material may need to support several fundamentally different representations:
These are not smaller views of the same graph. They recognize different objects, answer different questions, and obey different rules.
AutoSchemaKG, Extract-Define-Canonicalize, and related systems have made important advances toward constructing graphs without a predefined ontology, and much of my thinking builds upon their work. But these systems generally retain what has become the standard premise of cross-document knowledge-graph construction: use language models to extract raw subject–predicate–object triples from each source, then canonicalize the resulting entities and relations across the corpus. Language models are remarkably capable of producing those triples, and the triples themselves remain necessary. But treating them as the primary extraction artifact is insufficient for truly open-ended schema induction and high-fidelity entity resolution. By the time a source has been reduced to triples, the extractor has already committed to which mentions denote entities, which references belong together, which distinctions can be ignored, and how complex relationships should be divided into binary edges. The context and competing evidence needed to reconsider those decisions may already be gone. The question is not whether triples belong in the graph, but when they should be produced: as the beginning of construction, or as one downstream projection of a richer evidentiary process.
OKGC - Open Knowledge Graph Construction - takes the latter approach. It begins by extracting a richer, source-faithful body of evidence: addressable mentions, possible identities and explicit distinctions, semantic and value signals, coreference candidates, complex relational frames, assertion context, provenance, and uncertainty. Open-ended entity resolution consumes this evidence directly, allowing observations across many sources to accumulate before the system determines what refers to the same thing, what must remain separate, and which relational structures recur. Schemas can then be induced over resolved entities and relationships rather than over a fragmented collection of prematurely constructed triples. Binary graph edges remain available where useful, but they are produced after resolution rather than serving as its lossy input. Every interpretation remains connected to its evidence and available for revision as the corpus grows.
A universal semantic graph is one possible product of that engine, but not its endpoint. Caro requires several interoperable graphs to make a research corpus intelligible: semantic, event, claim, evidence, argument, discourse, narrative, and composition structures that share anchors without collapsing into one another. Kensho requires graph-based evolutionary substrates whose components can be recombined, mutated, searched, and evaluated. Other applications will require representations not yet anticipated. OKGC is therefore enabling infrastructure for all three areas of my research: a representational architecture through which understanding becomes explicit, and a construction system through which those representations become available for systematic discovery. The underlying wager is that knowledge can become not merely findable or conversational, but structurally usable: a living substrate that humans and machines can inspect, contest, explore, and continuously improve.