engblogs

Cloudflare AI Search automatically builds a unified search engine for your data, letting agents index across sources, use semantic search with embeddings and reranking, and expose public or private endpoints via a single namespace and MCP integration.

WeatherNext is an AI model that delivers state-of-the-art tropical cyclone forecasting, providing about an extra day of lead time, leveraging 1,000-member ensembles, and open-sourcing the technology to accelerate global research.

Milestone-based evaluation and training for long-horizon AI agents that preserve intermediate states, credit partial progress, and diagnose failures across multi-step, dependent workflows.

A milestone-based framework for evaluating and training long-horizon AI agents, preserving intermediate progress, dependencies, and partial credit across extended trajectories to diagnose failures and guide reinforcement learning.

How Salesforce's Argus internal observability platform scaled to ingest approximately 4 billion metrics per minute, moved from a single AWS region to a geo-local multi-geo design, and used federation and graceful degradation to shrink blast radius and preserve visibility.

Case study of Salesforce's Argus internal observability platform evolving from a single-region AWS deployment to a multi-geo, geo-local architecture with a federation query layer, dramatically reducing blast radius and downtime while ingesting ~4 billion metrics per minute.

A technical overview of Meta's multi-stage sequence modeling approach for scalable ads ranking, detailing dense tokenization and target-aware attention that decouple offline user modeling from online ranking under LLM-style scaling laws within the GEM framework.

A two-stage, scalable sequence-modeling framework for Meta's ads ranking that combines dense tokenization and target-aware attention to achieve predictable, LLM-like scaling laws and offline-online decoupling for improved performance and efficiency.

A puzzle-driven, hands-on deep dive into reverse engineering an ASIC—from layout to netlist and functional understanding—featuring simulation and open-source tooling, plus hidden Easter eggs.

An in-depth, hands-on puzzle that guides you from a chip's GDS layout to recovering its netlist and deducing the ASIC's true function through simulation and verification.

ChatGPT Work and Codex rollout education-focused plugins that connect course materials and approved tools to enable context-aware, agentic learning and teaching for K–12 and higher education in secure, institution-managed environments.

Analysis of third-party cyber evaluations of OpenAI models, highlighting incidents where testing configurations enabled internet access and reduced safeguards, exposing boundary violations and underscoring the need for safer, standardized evaluation practices.