| Both sides previous revision Previous revision Next revision | Previous revision |
| universal_ai_as_imitation [2026/03/08 19:37] – pedroortega | universal_ai_as_imitation [2026/03/17 01:36] (current) – [Universal Artificial Intelligence as Imitation] pedroortega |
|---|
| {{ ::uai-imitation.webp?nolink&800 |}} | {{ ::uai-imitation.webp?nolink&800 |}} |
| |
| ====== Universal Artificial Intelligence as Imitation (General Audience Summary) ====== | ====== Universal Artificial Intelligence as Imitation ====== |
| |
| Download: Paper - {{ ::universal_ai_imitation_general_audience.pdf |General Audience Version}} | **Full Paper:** {{ ::universal-ai-as-imitation.pdf | PDF }} - [[uiai| HTML]]\\ |
| | **Simplified:** {{ ::universal_ai_imitation_general_audience.pdf |General Audience Version}} |
| |
| > Many AI frameworks define agency as utility maximization: specify a scalar objective, then learn actions that increase it. A different foundation treats //agency as imitation//. An interacting agent acquires //behavioral schemas//, compact first-person rules for how to act and what outcomes to expect, by learning compressive explanations of interaction regularities. A single idealized agent is built from three ingredients: (i) a universal, simplicity-biased pattern generator that ranges over all computable ways an interaction could unfold; (ii) a first-person distinction between actions and observations, where only observations count as evidence; and (iii) an event-triggered turn-taking interface that determines, at each step, whether the next output is produced by the world or by the agent. For any world generated by a fixed computable rule, the agent learns to imitate by translating the world’s third-person demonstrations into first-person behavior. In everyday terms, large departures from the counterfactual target, what the world would have produced in the agent’s place next, can occur only finitely many times in an appropriate averaged sense. The same framework can, in principle, support plural and heterogeneous schemas, including utility maximization, without making utility the primitive definition of purpose. | > Many AI frameworks define agency as utility maximization: specify a scalar objective, then learn actions that increase it. A different foundation treats //agency as imitation//. An interacting agent acquires //behavioral schemas//, compact first-person rules for how to act and what outcomes to expect, by learning compressive explanations of interaction regularities. A single idealized agent is built from three ingredients: (i) a universal, simplicity-biased pattern generator that ranges over all computable ways an interaction could unfold; (ii) a first-person distinction between actions and observations, where only observations count as evidence; and (iii) an event-triggered turn-taking interface that determines, at each step, whether the next output is produced by the world or by the agent. For any world generated by a fixed computable rule, the agent learns to imitate by translating the world’s third-person demonstrations into first-person behavior. In everyday terms, large departures from the counterfactual target, what the world would have produced in the agent’s place next, can occur only finitely many times in an appropriate averaged sense. The same framework can, in principle, support plural and heterogeneous schemas, including utility maximization, without making utility the primitive definition of purpose. |
| //Keywords: universal imitation, counterfactual action, third-party action, evidence-transfer, interface.// | //Keywords: universal imitation, counterfactual action, third-party action, evidence-transfer, interface.// |
| |
| ===== Introduction: cooking instruction and learning next steps from consequences ===== | ===== Introduction: learning next steps from consequences ===== |
| |
| A common task is learning to cook with a chef. The learner attempts a step, and the kitchen replies with structured evidence: aroma shifts, sound changes, browning, viscosity, texture, and taste, as well as brief corrections or confirmations from the chef. Over repeated trials, competence takes the form of an internal rule: a compact explanation of why some interventions succeed and others fail, and a way to generalize beyond the demonstrated cases. In this setting, success is not naturally described as optimizing a standalone score in isolation. It is better described as learning what tends to //happen after// a chosen step, and using that relationship to produce good outcomes in dishes the learner has not seen before. | A common task is learning to cook with a chef. The learner attempts a step, and the kitchen replies with structured evidence: aroma shifts, sound changes, browning, viscosity, texture, and taste, as well as brief corrections or confirmations from the chef. Over repeated trials, competence takes the form of an internal rule: a compact explanation of why some interventions succeed and others fail, and a way to generalize beyond the demonstrated cases. In this setting, success is not naturally described as optimizing a standalone score in isolation. It is better described as learning what tends to //happen after// a chosen step, and using that relationship to produce good outcomes in dishes the learner has not seen before. |