From dfa047b21fbfc3afee431014bdde4526f6658846 Mon Sep 17 00:00:00 2001 From: Courtland Leer Date: Thu, 18 Sep 2025 14:42:05 -0400 Subject: [PATCH] note proofs --- content/notes/Identity is diachronic.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/content/notes/Identity is diachronic.md b/content/notes/Identity is diachronic.md index 09f1dbbe9..00fa279ae 100644 --- a/content/notes/Identity is diachronic.md +++ b/content/notes/Identity is diachronic.md @@ -6,11 +6,11 @@ tags: - ml date: 09.18.25 --- -The quality of any single AI system output is in large part determined by the context available to it at inference time. While some context is static and reusable, AI systems aspiring to be truly generative, 1-to-1, and dynamic must also manage large sets of changing context. +The quality of any single AI system output is in large part determined by the context available to it at inference time. While some context is static and reusable, AI systems aspiring to be truly generative, 1-to-1, and dynamic, must also manage large sets of changing context. So a major obstacle to solving context management for AI systems is in squaring continuity with change. That is, how do we differentiate between static and evolving context? Or should we instead view both as properties of context sets? What do we even call a body of context that remains static in some ways, but changes over time in others? -In philosophy and many other cognitive sciences this is called an *identity*. All material or conceptual objects have a personal identity--humans, agents, APIs, NPCs, organizations, communities, memes, brands, etc. And those identities have a temporal dimension, they can be coherently understood as enduring things but also as different at different moments in time. They're diachronic[^1]. +In philosophy and many other cognitive sciences, this is called an *identity*. All material or conceptual objects have a personal identity--humans, agents, APIs, NPCs, organizations, communities, memes, brands, etc. And those identities have a temporal dimension; they can be coherently understood as enduring things, but also as different at different moments in time. They're diachronic[^1]. At Plastic, we think context management/engineering is really better framed as identity management/engineering. Delivering the best context to your system in each moment itself requires a system to understand multiple identities at multiple points in time.