An AI system was given 45,047 Enron emails and wrote 214 pages covering the companies, regulators and banks the energy firm dealt with. Every fact on these pages cites the specific email it came from. When the correspondence shows a fact stopped being true, the page expires that fact on the relevant date. No person edited any page or fact.
The project produced 49,401 facts, each linked to its source email. A held-out audit by a second AI found 93% of sampled facts fully supported by their email, and none of 500 sampled facts wrong at their core. The most-discussed organizations in the corpus include the Federal Energy Regulatory Commission (FERC), General Electric, Pacific Gas and Electric, and Dynegy, a Houston rival that agreed to acquire Enron in November 2001 before terminating the merger.
Each entity page includes a date slider. Moving it shows what was known at that point in time, with temporary facts expiring as the timeline advances. Other highlighted counterparties include El Paso Corporation, the California Public Utilities Commission, Duke Energy, the California ISO, Dow Jones and Reliant Energy.
The builders argue that facts in a company brain are sentences with dates and sources, not fixed fields such as headcount or revenue. Those structured fields belong in a system of record, while this kind of knowledge layer holds information that changes too often or does not fit into fields. The team grouped entities into broad classes, retaining the pipeline's narrower types as tags, and chose which entity kinds merited pages before building anything.
Source: Hacker News · Summarized by HeadlinesBriefing