AI & ML Research 24 Hours

Last updated: September 16, 2026, 2:02 PM ET

AI & ML Research

OpenAI is testing ad formats that embed sponsored results directly inside AI-generated responses, a significant shift for a company whose core product has so far been ad-free.

Infrastructure & Memory

A detailed VRAM budget formula for LLM serving shows why inference servers exhaust GPU memory long before compute becomes the constraint, with three optimization strategies mapped to specific bottlenecks. The N squared pizza problem uses an unexpectedly practical analogy to illustrate quadratic memory blowups in 3D scan matching pipelines, a pattern that recurs across ML workloads.

Search & Training

Google researchers propose Retrieve-for-Train, a method that accelerates complex AI search by restructuring how retrieval interacts with training, targeting inference bottlenecks in large-scale systems.

Materials & Industry

The AI boom is increasingly a materials challenge, as pushing computing into new thermal and density regimes demands advances in substrates, packaging, and interconnect that traditional semiconductor roadmaps were not designed to supply. MIT Technology Review's daily digest covers the trillion-dollar gamble underlying current AI capex and OpenAI's reported push to acquire large-scale biology datasets for model training.