Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore the evolution from conversational AI to autonomous agentic systems that plan, select tools, and execute actions in dynamic physical environments through this 25-minute conference talk. Discover how small language models (1-10B parameters) often outperform larger counterparts at the edge by delivering superior accuracy for specialized tasks while dramatically reducing latency, bandwidth requirements, and operational costs.
Learn about the technical advances driving small language model efficiency, including Mixture-of-Experts routing, KV-cache optimization, attention mechanisms, quantization techniques, and knowledge distillation methods. Understand why task-specific models tuned for mathematics, computer vision, or reasoning consistently outperform generalist models despite having significantly fewer parameters, and how this specialization advantage compounds in real-world applications where cameras, robots, and drones continuously stream data requiring immediate decision-making.
Examine a comprehensive layered architecture for edge-based AI agents, featuring an application control layer that manages goals, planning, constraints, guardrails, and tool selection. Investigate the model service layer that hosts multiple purpose-built models optimized for different hardware targets, supported by a metastore for performance tracking and insights. Analyze the data management layer responsible for transforming raw sensor streams into clean, actionable inputs, while orchestration coordinates capabilities across diverse devices through distributed memory sharing and intelligent load balancing.
Compare different tool implementation strategies, from lightweight single-frame vision processing that minimizes network traffic to multi-frame, high-resolution pipelines that create logarithmic increases in data transfers and inference requirements. Understand how heterogeneous hardware deployment becomes advantageous when paired with intelligent scheduling systems that move beyond random placement to capability-aware and load-aware distribution strategies.
Review practical deployment scenarios across warehouses, agricultural operations, drone systems, and clinical monitoring environments. Gain insight into current system readiness, including existing building blocks and critical gaps in robust guardrails, predictable iteration limits, and resource-aware orchestration that separate demonstration systems from production-ready deployments.
