https://sparsenotes.com/series/weekly-papers-2026-08-15/ Recent content in weekly-papers-2026-08-15 on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 15 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/codd-relational-model-1970/ Sat, 15 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/codd-relational-model-1970/ Seminal paper of the week: the eight-page CACM paper that replaced pointer-chasing with sets, invented data independence, and made the query optimizer possible. https://sparsenotes.com/posts/2026/08/papers/synthetic-persona-pretraining-alignment-token-zero/ Sat, 15 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/synthetic-persona-pretraining-alignment-token-zero/ Annotating 10% of the pretraining corpus with first-person value reflections installs the assistant persona from step zero — and the alignment advantage over post-training-only grows with compute. https://sparsenotes.com/posts/2026/08/papers/ Sat, 15 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/ One seminal paper plus two top picks from the week: Codd's relational model at 56, alignment installed from token zero of pretraining, and the case that the Linux page cache should own your MoE expert tier. https://sparsenotes.com/posts/2026/08/papers/kernel-managed-expert-cache-moe-inference/ Sat, 15 Aug 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/08/papers/kernel-managed-expert-cache-moe-inference/ A measurement study says the Linux page cache is a legitimate expert tier for trillion-parameter MoE serving — the hand-tuned user-space cache the whole field builds is worth 1.09×, and the scary reclaim knee is a benchmark artifact.