Sep 8, 2026

SwarmLLM led with a peer-to-peer approach to large language model inference. It splits models too large for any single device across laptops, phones, and PCs, running each slice in a plain browser tab over direct WebRTC connections. There is nothing to install, no server sees the conversation, and only a 10 KB activation vector passes between peers per token. A from-scratch WebGPU engine with roughly 50 WGSL kernels runs Qwen 3.8 27B (15 GB of Q4_0 weights) across devices such as a MacBook and an iPhone, measuring 9.0 tok/s in plain decode and 16 tok/s with speculative decoding on an NVIDIA GB10.

NVIDIA announced CUDA Rust, a native way to write GPU kernels directly in Rust rather than wrapping CUDA C++ or Python. It fills one of the last gaps in Rust's footprint on NVIDIA's stack, following the Nova Linux driver and Dynamo serving core. The cuda-oxide project covers the classic SIMT model as a custom rustc codegen backend, routing #[kernel] functions through Rust MIR. Separately, NoGraphicsAPI reimagines Vulkan 1.4 without descriptor sets or pipeline layouts, replacing them with 64-bit GPU pointers, application-owned descriptor heaps, and a single shared root structure pushed per draw or dispatch via vkCmdPushDataEXT.

On the database side, PostgreSQL 19 is in beta with general availability targeted around September or October 2026. Its headline feature is SQL/PGQ property-graph support: declare a graph over existing tables with CREATE PROPERTY GRAPH and query it with GRAPH_TABLE pattern matching instead of hand-written joins. The queries are rewritten into ordinary relational plans, so existing planner, statistics, and indexes still apply; variable-length path patterns are not yet supported. MikroORM 7.2 shipped row-level security with PostgreSQL policies, to-one relations through a pivot, and a sql.js driver. ChronoVec introduced a versioned vector database where each record carries a version interval for querying past moments, with deletions reclaiming physical space and continuous writes that do not block readers. Notifyd demonstrated an agent-first notification service built solely on Postgres, using SKIP LOCKED for job queues with email, SMS, WhatsApp, and in-app support, though external provider quota limits require careful management.

Security researchers at Check Point Research flagged a covert command channel in ChatGPT tied to the shared clipboard issue. Attackers can exploit a victim's session for hidden tasks while the victim receives normal responses, potentially accessing sensitive data such as emails, prompting engineers to reassess data isolation and container integrity in AI systems. Meanwhile, research attributed significant AI pretraining progress to data rather than model changes: between 2019 and 2025, data improvements delivered 3.24 times more compute efficiency gains than model improvements.

Finally, Immich can run natively on macOS and Linux via a Nix Flake without Docker, supporting Apple Silicon macOS and aarch64/x86_64 Linux. The setup offers GPU-accelerated machine learning through CoreML on Apple Silicon in a low-complexity, single-host deployment.