Generative Time Capsule

Bernac, Kay and Street's inversion of archival practice: transmit a civilization's cognitive function rather than its records. A neural network already stores a "superhistory" of its training data as a lossy compression that remembers patterns and forgets noise — but unlike a document it can be run, so "the approximated function is more complete than data because it has the capacity to generalize to new conditions." Discovery becomes interaction rather than reading. Scored against three desiderata (completeness, durability, accessibility) on which the Rosetta Stone and a frontier chatbot each fail one; instantiated speculatively as Chronoseed, weights encoded in DNA inside multilayered kernels distributed in an anomalously regular, self-signposting pattern. Its unsolved blocker is the interface: nothing in the kernel can run the model.