Friday, July 17, 2026

1. Online Neural Space Time Memory: Real-Time Memory Mechanism for Dynamic Novel View Synthesis
   T2arXiv AI/CL/LG（论文）

   For online novel view synthesis from multi-view streaming videos, this paper proposes a framework that decouples memory update frequency from application, using periodic updates and per-frame memory application with cross-view attention for deformation, plus auxiliary Memory Loss and caching to stabilize long contexts, achieving real-time state-of-the-art performance on dynamic human scenes.

2. Firefox Compiled to WebAssembly, Runs Inside a Browser
   T2Simon Willison 博客（Atom）

   The Puter project has successfully compiled Firefox's Gecko engine to WebAssembly, allowing it to run as a full browser within another browser. The project leveraged AI-assisted programming using Claude Opus and Fable tokens, costing approximately $25,000 via a Claude Max subscription. All network traffic is proxied over WebSocket using the Wisp protocol, with end-to-end encryption support. This breakthrough showcases the immense potential of combining AI with WebAssembly, opening new possibilities for browser nesting and remote operations.

—

1. OpenAI Introduces AI Age Scorecard: A New Way to Measure ROI
   T1OpenAI：News（RSS）

   OpenAI CFO Sarah Friar proposes a practical AI scorecard that measures ROI through useful work, cost per successful task, dependability, and return on compute. This framework aims to help organizations objectively evaluate the real value of AI deployments, avoiding a sole focus on technical metrics while ignoring business benefits.

1. Data-Driven Block Replacement Scheduling: Multi-Armed Bandit Based Nonparametric Cost Optimization
   T2arXiv AI/CL/LG（论文）

   This paper addresses the block replacement scheduling problem for N independent identical machines with unknown lifetime distributions. By formulating the interval selection as a stochastic multi-armed bandit, the authors propose Hoeffding- and Bernstein-based lower-confidence-bound algorithms achieving O(K log T) regret, matching the Lai–Robbins lower bound. Exploiting the nested observation property unique to block replacement, correlated variants attain O((K-k*) log T) regret and require only O(1) direct pulls of suboptimal arms k < k*. A complementary Kaplan–Meier renewal algorithm nonparametrically estimates the lifetime distribution from censored data, ensuring almost-sure policy consistency. Numerical experiments confirm theoretical ordering and reveal structural cost gaps between optimal block and age-dependent replacement.

1. To Ease Data Center Water Stress, Hyperscalers Should Buy Golf Courses, Convert to Birdwatching Parks
   T2Simon Willison 博客（Atom）

   Addressing the enormous water consumption of AI data centers, blogger Simon Willison proposes a counterintuitive solution: hyperscalers should buy and repurpose golf courses. For example, Google used 10.9 billion gallons in 2025, while the Coachella Valley's 120 golf courses each consume ~750,000 gallons per day. Acquiring 40 of those (one-third) would offset Google's daily water use. The plan turns golf courses into public birdwatching parks and funds former members to adopt the sustainable hobby, balancing environmentalism and water conservation.