Escrever uma avaliação PREreview

The rapid integration of large language models into commerce, productivity, and daily decision-making is creating a new class of AI intermediary—one that not only recommends but transacts, delegates, and acts on behalf of users. This concentration of influence raises a structural question: who governs the representation of what a user wants? We introduce Personal Intelligence as a research and design programme whose primary objective is preference continuity under user control. We propose a core construct, the Personal Preference Substrate (PPS): a user-governed, evolving, portable representation of preferences, constraints, context, and provenance. We argue that PPS is naturally expressible as a multi-relational graph — a taste graph—because preferences are compositional, contextual, and evidence-dependent. The paper develops seven interlocking contributions: (1) a motivation grounded in platform incentive decay, the agentic-commerce zeitgeist, and the epistemic fragility of current AI systems; (2) a formal definition of PPS with axioms for user governance, provenance, temporal awareness, epistemic separation, and boundary-conditional disclosure; (3) a hybrid learning architecture combining passive behavioural signals with active micro-queries and signal-gated refresh; (4) a boundary-layer framework addressing egress control, ingress validation, incentive insulation, bounded delegation, and adversarial resilience; (5) a privacy-preserving collective improvement model spanning federated learning, secure aggregation, and differential privacy; (6) a portability model distinguishing representation, state, and capability portability with scoped identity; and (7) an evaluation framework measuring preference fidelity, autonomy preservation, epistemic quality, and regret reduction rather than engagement alone. We ground the framework in three pluripotent use cases—commerce, tool mediation, and epistemic filtering—and conclude with a research agenda identifying open problems in graph-based preference representation, on-device inference, portable identity, and incentive-aligned collective learning.