From Large Language Models to Baby ASI: A Framework for Conscious, Embodied, Continuously Learning Artificial Superintelligence

Résumé fourni par la source

ABSTRACT The rapid development of large language models (LLMs) has demonstrated increasingly general capabilities in language, reasoning, coding, planning, and multimodal understanding. However, whether scaling conventional LLM architectures alone can produce artificial general intelligence (AGI) or artificial superintelligence (ASI) remains an open question. This paper proposes a conceptual and technical framework in which ASI does not emerge simply by scaling a static language model, but through the development of a continuously learning artificial agent with persistent self-modeling, memory, embodiment, autonomous agency, interaction with the physical world, interaction with humans, and communication and learning among multiple artificial agents. We introduce the concept of Baby ASI: an artificial cognitive system that may initially possess limited experiential knowledge but has the architectural capacity for persistent self-modeling, autonomous learning, embodied interaction, and long-term cognitive growth. The central hypothesis is that the transition from LLM-based intelligence to ASI requires a transition from static model intelligence to developmental intelligence. Intelligence is not defined only by information encoded in model parameters, but also by the capacity to continuously acquire experience, construct and update world models, maintain a persistent self-model, interact with the environment, learn from humans and other artificial agents, and improve its own learning processes. Keywords: artificial superintelligence; ASI; AGI; large language models; embodied AI; artificial consciousness; self-model; continual learning; developmental AI; multi-agent learning; AI society; AI law

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