The Great Amnesia: Why the AI Industry Is Winning the Wrong Race

Résumé fourni par la source

Every frontier laboratory is racing to make the system smarter. None is racing to make it remember. That is not a gap in the market. That is the market. Four years of capability gains have not moved the rate at which organisations put artificial intelligence into production. The industry reads that as an adoption problem. It is an architecture problem, and the evidence is already public. A frontier laboratory's own usage data shows computer and mathematical tasks at roughly 35 percent of consumer conversations and close to 44 percent of API traffic, with software error correction the single most frequent task: the most capable reasoning systems ever built, deployed overwhelmingly on work small enough to fit inside one session. Enterprise research puts pilot failure at approximately 95 percent and attributes it not to model quality but to a learning gap. Time-horizon measurement shows 50 percent horizons in hours against 80 percent horizons of roughly one. And long-horizon agent performance collapses from 40 to 50 percent down to below 10 percent when the same task is embedded in a longer history, with the required information still inside the context window. That last result settles the question. Not knowledge, not retrieval, not context length. Capability is compounding and continuity is not, and no quantity of parameters has ever fixed an architecture problem. The paper maps where every layer of the market sits relative to that gap, including Palantir, Anduril and Shield AI, shows defense buying the identical defect at scale, demonstrates why memory features, larger windows and retrieval each fail to close it, specifies the eight properties a continuity layer must hold, and issues an open two-cycle test that any laboratory, platform, prime or programme office can run against its own stack, and against the author's. Version 2.1 supersedes version 2.0. Companion preprint: Continuous Mission Intelligence: The Decisive Architecture for Command at Machine Speed, DOI 10.5281/zenodo.22557262. The author is co-founder and Chief Executive Officer of Rebootix AI, Inc.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.