https://sparsenotes.com/series/weekly-papers-2026-10-31/ Recent content in weekly-papers-2026-10-31 on Sparse Notes https://sparsenotes.com/images/og-default.png https://sparsenotes.com/images/og-default.png Hugo -- gohugo.io Sat, 31 Oct 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/10/papers/milner-theory-of-type-polymorphism-1978/ Sat, 31 Oct 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/10/papers/milner-theory-of-type-polymorphism-1978/ Seminal paper of the week: Robin Milner gave ML a type system in which programmers write no type annotations, the compiler infers the most general type for every expression, and well-typed programs provably cannot go wrong. Algorithm W and let-polymorphism are still the core of OCaml, Haskell and F#, and the template for inference in Rust and Swift. https://sparsenotes.com/posts/2026/10/papers/metadb-file-system-structure-distributed-database/ Sat, 31 Oct 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/10/papers/metadb-file-system-structure-distributed-database/ Baidu's metadata database for its exabyte-scale object and file storage. Sharding on directory boundaries, asynchronous aggregation of hot parent-directory attributes, and in-memory MVCC with locality-aware compaction give 2.2–4.1× the throughput of their previous TiDB-style database on half the machines, across 1,000+ nodes and tens of trillions of records. https://sparsenotes.com/posts/2026/10/papers/platonic-representation-theorem-jepa/ Sat, 31 Oct 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/10/papers/platonic-representation-theorem-jepa/ Balestriero and LeCun prove that JEPAs trained separately on different modalities learn the same representation up to one rotation, provided their positive pairs are two observations of the same scene. Next-token prediction turns out to be a JEPA with bad views, and a 109M-parameter text JEPA aligns with vision encoders better than 72B language models. https://sparsenotes.com/posts/2026/10/papers/ Sat, 31 Oct 2026 00:00:00 +0000 https://sparsenotes.com/posts/2026/10/papers/ This week's picks: Balestriero and LeCun's proof that JEPAs with the right positive pairs learn the same representation across modalities; Baidu's MetaDB, the metadata database under its exabyte-scale storage; and Milner's 1978 type polymorphism paper as the seminal pick.