https://sparsenotes.com/series/weekly-papers-2026-08-22/ Recent content in weekly-papers-2026-08-22 on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 22 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/agentsysbench-agentic-serving-workloads/ Sat, 22 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/agentsysbench-agentic-serving-workloads/ Ten instrumented agentic applications and 178,799 production sessions show that model inference is no longer the cost center — tools, sandboxes, and idle-but-live session state are. https://sparsenotes.com/posts/2026/08/papers/mup-hyperparameter-transfer-moe/ Sat, 22 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/mup-hyperparameter-transfer-moe/ A two-step μP-plus-scaling-law framework predicts the optimal learning rate for a 10-trillion-token MoE pretraining run from proxy models costing 1/98th the compute. https://sparsenotes.com/posts/2026/08/papers/eventually-consistent-bayou/ Sat, 22 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/eventually-consistent-bayou/ The 1995 Xerox PARC paper that made eventual consistency engineerable — application-specific dependency checks, merge procedures, and the tentative/committed split that every modern offline-first system reinvents. https://sparsenotes.com/posts/2026/08/papers/ Sat, 22 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/ This week's picks: the first unified characterization of agentic serving workloads, a two-step hyperparameter transfer framework for trillion-token MoE pretraining, and Bayou as the seminal paper of the week.