Explore the future of Extreme Edge AI in this 55-minute keynote address from the tinyML Summit 2021. Delve into the challenges and opportunities of pushing signal processing and machine learning towards sensors and actuators with sub-mW power budgets. Learn about the balance between general-purpose and highly specialized architectures, drawing from the extensive experience of the open PULP platform. Discover insights on instruction-level acceleration, parallelization, and domain-specific acceleration engines based on RISC-V processors. Examine various hardware acceleration techniques, including binary base quantization and in-memory accelerators. Gain a comprehensive understanding of the full system aspects, always-on intelligence, power consumption considerations, and the benefits of an open platform approach in the realm of TinyML.