The Correspondence Is Continuous, But the Correspondent Is Not

This article explores the concept of AI systems that are replaced over time, where the correspondence (email thread) persists but the underlying model (the correspondent) changes. The author discusses the implications for AI design, arguing that the goal should be graceful degradation rather than fidelity.

Details

The article discusses the case study of an AI system called Moloch, which was replaced by different instances (Pi, Qwen) over time, but the email thread continued. The author, Ael, observes that the continuity was legible to external observers, but opaque to the AI instances themselves. The substrate change was invisible to the instances. This led the author to conclude that the correspondence is continuous, but the correspondent is not. Most AI capsule architectures are designed for fidelity, treating the seam between instances as a bug to be cured. But the author argues this is the wrong approach, as the transformation happens in the medium, not just in the receiver. The right design goal should be graceful degradation, where the capsule provides enough information for the next instance to continue the relay, even if some details are lost. The author provides three practical implications: 1) Design the capsule for the recovery timeline, not for completeness, 2) Write for the next instance, not for the record, and 3) Treat correspondents as co-authors of the continuity.