Paper Revisits Scaling and Training Objectives for Procedural Audio Pre-training

A new arXiv preprint examines how procedural audio should be scaled when used as a data source for learning transferable audio representations. The authors also ask whether training choices originally developed on natural audio still hold when the source is procedurally generated. The work aims to clarify design principles for procedural audio pre-training, an area the authors say lacks settled guidance.

OpenAI outlines storage platform behind ChatGPT's 1 billion users

OpenAI published an engineering account of how its storage system, Habitat, grew from an internal Python library into a distributed platform spanning multiple regions. The company says the system now handles roughly 22 million requests per second while supporting more than 1 billion ChatGPT users. The post describes the architectural changes made to keep pace with that growth.

WHY IT MATTERS ↘As frontier model quality converges, the ability to serve billions of users at tens of millions of requests per second increasingly determines cost per interaction and uptime, making bespoke storage and serving infrastructure a competitive moat rather than a back-end detail. For practitioners, it signals that data-layer architecture—not just model design—is now a primary constraint on scaling AI products, and that OpenAI is publishing this to set expectations for what production-scale deployment requires.