engblogs

Two harness settings—retained reasoning and compaction—in the ARC-AGI-3 benchmark dramatically boosted GPT-5.6 Sol's score by preserving past reasoning across actions and reducing output tokens.

From MIT's early ECG data digitization to PhysioNet's global data-sharing standard, the post explains how de-identified clinical datasets such as MIMIC and expanding repositories of EHR and imaging data are enabling biomedical research and AI.

From MIT's early ECG digitization to a global, open-standard platform, PhysioNet catalyzed biomedical data sharing, enabling AI-driven health insights and collaboration across researchers and institutions.

Tame Dependabot by grouping updates, slowing cadence, and preserving fast security fixes across ecosystems, demonstrated on a Microsoft open source project.

Tame Dependabot by grouping updates, slowing cadence, and keeping security fixes fast, demonstrated with a practical YAML-driven pattern on a Microsoft open-source project.

K-Search, powered by a CUDA-to-MLX translation layer and a native MLX backend, demonstrates automatic transfer of decades of CUDA kernel optimization to Apple Silicon, delivering near-expert attention performance and a 20x prefill speedup on SSM while outlining a generalizable path for cross-hardware kernel portability.

K-Search translates decades of CUDA kernel optimization expertise into efficient Apple Silicon GPUs using a structured translation layer and an MLX backend, delivering near-expert performance and cross-platform kernel portability.

Qualcomm's acquisition of Modular combines an AI-native software platform with energy-efficient compute to accelerate generative and agentic AI from edge to cloud within an open, heterogeneous ecosystem.

Lyria 3.5 arrives in Google Flow Music with enhanced musicality, lyrics, and vocals plus finer creative control, delivering more expressive, natural-sounding song generation.

Explores how agentic AI and coding agents accelerate scientific computing—from genomics to data-rich research pipelines—by reducing engineering toil, enabling faster prototyping, and shifting researchers toward verification and long-term software stewardship.

Agentic AI and coding agents accelerate scientific software development and maintenance in genomics and data-rich life sciences, shifting researchers toward verification, orchestration, and long-term stewardship.

A field report on how agentic AI and coding agents accelerate scientific software development for genomics and other data-rich disciplines, boosting speed and maintainability while highlighting the need for verification, governance, and long-term stewardship.