The presentation outlines the evolving reference architecture needed to support the next generation of edge AI, moving beyond simple data processing to complex edge reasoning. It explores the convergence of novel heterogeneous solutions that optimise the hardware-software-AI-data quad required to run GenAI, agentic AI, Small Language Models (SLMs) and Vision Language Models (VLMs) efficiently at the edge. By defining a holistic edge AI technology stack that integrates trustworthy edge AI, software platforms and datasets, with hardware-aware optimisation, it is possible to lay the groundwork for mesh intelligence, in which edge systems possess the agency to sense, reason, plan, learn, memorise and execute tasks autonomously. These fundamental elements support the orchestration of the interactions within a collaborative mesh of autonomous AI-defined entities and components. The presentation opens the floor for discussion to provide a roadmap for shifting edge AI systems to dynamic, GenAI-enabled architectures capable of supporting the high computational demands of real-time agentic reasoning, with applications in AI-Defined Vehicles and robotics.
Peter Friess (Tue,) studied this question.