From data to silicon
Summary
From data to silicon argues for extreme specialization of hardware by learning AI models directly into silicon, proposing that cheap custom ASICs could be widely produced for specific ML tasks. The piece showcases a small, autonomous-end-to-end design: a single-cycle Connect-4 chip built on a 130 nm process with ~14k gates, emphasizing energy efficiency and a path toward automated hardware design for AI workloads.