Episteme
The goal of Episteme is to build, maintain, and serve an open repository of claims overseen by an LLM-based bureaucracy, where contributions from humans, research agents, and other parts of civilization's “full epistemic stack” get appropriately evaluated and propagated.
It starts with creating a database of claims from a large corpus of internet discourse, academic papers, and the like. These are deduplicated, canonicalized, and organized in a wiki. For each claim in the claim graph, an LLM admin maintains the claim page according to principles of good claim administration laid out in a constitution in the system prompt. Contributions are welcome via an open protocol, and are handled according to rules and procedures defined in the Constitution and further developed through a form of precedent and common law.
The project builds on the Forethought design sketches for collective epistemics, and is developed in the open on GitHub. What follows is a discussion of technical solutions to the main subproblems:
- Decomposing documents into claims
- Splitting vs. lumping similar claims
- Canonical forms of claims
- Structuring the claim graph (Wiki vs. DAG vs. DCG)
- Defining edges between claims
- Determining claim importance
- Matching instances of claims to claims in the graph
- Traversing the claim graph
- Propagating evidence and updates through related claims
- Appropriately weighting evidence, especially from public contributors
- Maintaining neutrality while avoiding nihilism