What will the future of artificial intelligence (AI) entail? How can we get a complete overview of the evolving AI landscape? The research paper “Designing Ecosystems of Intelligence from First Principles” by Friston et al. (2024) outlines a visionary vision for the field of artificial intelligence (AI) in the next decade and beyond. This vision focuses on the development of a cyber-physical ecosystem comprising both natural and synthetic elements that collectively contribute to what is called “shared intelligence”. This concept emphasizes the integral role of humans in these ecosystems. The paper highlights a specific approach to AI known as “active inference,” which is seen as a physics-based approach to understanding and designing intelligent agents. This approach shares fundamental principles with quantum, classical, and statistical mechanics.
Active inference applies to AI design, suggesting that next-generation AI systems should be equipped with explicit beliefs about the world, including a specific perspective under a generative model. This contrasts with traditional AI approaches such as reinforcement learning, which focus primarily on choosing an action to maximize rewards. In active reasoning, exploration and curiosity are seen as equally central to intelligence, driving actions that are expected to reduce uncertainty.
The multi-scale active inference architecture is another crucial aspect. It recognizes different time scales in model training and selection, working in similar ways on nested time scales to maximize model evidence. In this context, intelligence is inherently perspectival, involving an active engagement with the world from a specific set of beliefs.
Communication in these intelligent systems is also a key topic. The paper argues that intelligence at any scale requires a shared generative model and common ground, which can be achieved through various methods such as ensemble learning, mixtures of experts, and Bayesian model averaging. An important aspect of active inference in this context is the selection of messages or viewpoints that provide the greatest expected information retrieval.
Finally, the paper addresses ethical considerations, emphasizing the importance of valuing and preserving individuality in the development of large-scale collective intelligence systems. This approach contrasts with models such as eusocial insects, where individuals are largely substitutable. The authors advocate a cyber-physical network of emergent intelligence that respects the individuality of all participants, human or otherwise.
In summary, the white paper by Friston et al. presents a visionary approach to AI development centered around active inference and the creation of intelligent ecosystems that include and respect the individuality of both human and non-human agents. This approach implies a significant paradigm shift in the way AI is conceptualized and developed, with implications for the future of technology and society.
Image source: Shutterstock