noyzless

Meta's Vistara ASIC enables CXL-based reuse of legacy DDR4 RAM in new DDR5 servers, reducing infrastructure costs and OOM events.

Google's AlloyDB introduces 'proxy models' that distill LLM behavior into local, fast models inside the database, drastically reducing latency and costs for per-row AI operations.

Meta is moving to production for custom MTIA AI chips by September to reduce reliance on merchant silicon. Highlights the industry-wide move toward vertical integration of AI infrastructure.

OpenAI releases GPT-5.6 family (Sol, Terra, Luna) with claims of superior efficiency and cybersecurity performance compared to Anthropic’s current offerings.

A masterclass in debugging. OpenAI explains how they identified a 18-year-old GNU libunwind race condition using 'epidemiological' data analysis of core dumps to separate hardware noise from software bugs.

Netflix engineers detail their migration from direct database access to an abstraction-based 'key-value' service, emphasizing observability, safety, and validation as core architectural principles.

A pure-C engine for running large MoE models on consumer hardware by streaming model weights from disk to RAM.

Progress report on pgrust, a Postgres rewrite in Rust that now passes 100% of standard regression tests. Significant for database systems interest.