Tokyo's Light-Based AI Slashes Compute Costs

🔥 What happened University of Tokyo researchers, with NTT, built a physics-based AI that processes information as light through a 200-meter fiber loop instead of transistors. Their SS-DFA training plus single-map architecture skips most intermediate measurements, cutting computation, storage, and control overhead. 💡 Why it matters The stripped-down optical system matched or beat far more complex neural networks in tests while using a fraction of the compute. Light carries signals fast and in parallel with little heat — potentially slashing both power and cooling costs in future AI data centers. ⚡ Our take If this scales to large language models, the GPU monopoly starts to wobble. Today it's a lab curiosity — but it's the first credible strike against the bigger-hotter-pricier dogma.