engblogs

Enhancements in Ray on Google Kubernetes Engine (GKE) provide a more native, efficient, and scalable experience for running AI workloads on Cloud TPUs with improved TPU-aware scheduling, simplified multi-host training, and integrated performance monitoring.

Mercado Libre leverages a multi-layered Spanner architecture, integrated with the Fury platform and advanced analytics, to achieve global consistency, scalability, and agility while enabling AI-driven applications and operational excellence.

Discover how Gemini Enterprise, Code Assist, and CLI integrate AI agents with Google Cloud's HPC to accelerate and enhance scientific research workflows from hypothesis generation to publication.

The Data Engineering Agent in BigQuery preview leverages AI to automate data pipeline creation, maintenance, and troubleshooting, enhancing productivity and simplifying complex data engineering tasks with integration to Dataplex and Gemini.

Analyzing Cloudflare Radar data reveals significant TCP connection anomalies and increased IP reachability in Turkmenistan during mid-2024, supporting reports of firewall testing and large-scale unblocking.

Embedding Atlas provides a scalable, low-friction, and interactive web-based tool for visualizing large-scale embeddings with advanced clustering and automated labeling features.

LinEAS introduces an end-to-end linear activation steering method using a global distributional loss to robustly control generative models with minimal unpaired data, outperforming existing activation-based interventions in toxicity mitigation and concept inclusion across modalities.

Policy maps provide an interactive framework for AI practitioners to design, navigate, and enforce customizable policies that manage the complex behavior space of large language models.

Switchboard, MD leverages Amazon Nova Sonic's real-time, accurate, and cost-effective transcription integrated with serverless AWS architecture to automate clinical call center workflows and enhance patient care.

Introducing Transitive RL, a novel divide-and-conquer paradigm for off-policy reinforcement learning that overcomes the scalability and error accumulation issues of temporal difference learning in long-horizon tasks by recursively splitting trajectories and using expectile regression to select subgoals.

An in-depth analysis of an HTTP/2 interop issue caused by improper handling of response bodies in Go clients, resulting in ENHANCE_YOUR_CALM errors due to PING flood mitigations, and practical solutions to avoid it.

An in-depth exploration of BGP zombies caused by stuck more-specific prefix withdrawals and excessive path hunting delays, their impact on internet routing, and strategies to mitigate these routing issues.