The Download: Boosting AI's memory, and data centers' unhappy neighbors (www.technologyreview.com)

Chinese startup DeepSeek unveiled an optical character recognition (OCR) model that researchers say introduces a new way of storing and retrieving information inside AI systems—essentially changing how the model “remembers” extracted text from images. On the surface it’s an OCR tool (used in scanner apps, photo translation and accessibility), but its main innovation is an altered memory/processing pipeline that could make retrieval of past inputs more efficient. If validated at scale, this approach could cut the inference compute needed for tasks that require long-term context, lowering latency and the energy footprint of memory-heavy applications. That shift matters as the industry grapples with the infrastructure side of AI growth. The MIT Tech Review’s AI Hype Index flags mounting tensions between massive data-center power demand and local communities—grids are straining, outages are more salient, and policymakers are eyeing big moves (including new nuclear projects) to feed the AI boom. For practitioners this reinforces two priorities: designing models and memory systems that reduce runtime compute and energy, and engaging with infrastructure realities (siting, grid capacity, and policy) that will shape where and how large-scale AI can sustainably operate.