engblogs

DeepMind and A24 announce a pioneering research partnership to blend cutting-edge research with film-production workflows, enabling artists to shape future tools and expand storytelling possibilities through hands-on collaboration and feedback.

MIT leaders call for curiosity-driven basic research and a human-centric, ethically grounded AI-enabled curriculum, backed by federal funding, to sustain America's innovation and education for the next 250 years.

MIT in the media makes the case that curiosity-driven research, human-centric AI, and foundational STEM education, supported by stable federal funding, are essential to sustaining America's innovation pipeline for the next 250 years.

Meta’s AI storage blueprint at scale outlines a unified BLOB-storage architecture on Tectonic that maximizes GPU utilization and research velocity via unified metadata, regional deployment, zero-overhead data access, and smart caching with prefetch and on-demand hydration.

Meta reveals a scalable, GPU-centric BLOB-storage stack atop Tectonic, featuring unified metadata, regional deployment, and tiered caching with prefetching to maximize GPU utilization and accelerate AI research velocity.

BAIR's 2026 Ph.D. cohort spans robotics, large language models, computer vision, generative modeling, AI safety, and human–AI interaction, highlighting research that advances theory, systems, and real-world impact.

Celebrating BAIR's 2026 PhD graduates whose diverse AI and ML research spans robotics, large language models, computer vision, generative modeling, AI safety, and real-world impact across science and healthcare.

A practical method to accelerate experiments in interdependent marketplaces by using order-level CUPED and pre-treatment predictions to reduce variance below the coarse randomization grain.

GeneBench-Pro is a synthetic, research-level benchmark that challenges AI agents to perform judgment-heavy, ambiguity-rich analyses across genomics, quantitative biology, and translational medicine, capturing the iterative, decision-critical process of real computational biology research to drive reproducible insights.

Inside Genebench-Pro: a technical tour of the benchmark with prompts, datasets, and 10 case studies illustrating how genomics prompts drive evidence-based clinical insights.

Core dump epidemiology: a data-driven investigation that reveals two distinct crash populations in Rockset—an 18-year-old race in GNU libunwind and a hardware-induced misalignment—with instrumentation and a switch to libgcc unwinder delivering the fix.

GeneBench-Pro is a research-level benchmark that measures AI agents' ability to navigate ambiguity, make higher-order judgments, and iteratively analyze synthetic problems in computational biology across genomics, quantitative biology, and translational medicine.