engblogs

Balancing transparency and safety in frontier AI through independent external third-party assessments that evaluate capability claims, safeguards, and governance across risk domains.

Evaluation frameworks ('evals') enable businesses to translate AI objectives into measurable, reliable outcomes through iterative testing, error analysis, and continuous improvement tailored to specific workflows and goals.

OpenAI enhances frontier AI safety by partnering with third party experts for independent evaluations, methodological reviews, and subject-matter expert probing, ensuring transparency, rigor, and trust in deployment decisions.

Terminal-Bench 2.0 and the Harbor framework offer robust, interactive, and scalable containerized benchmarks and execution environments designed to evaluate and optimize general-purpose CLI-based AI agents, driven by a community-focused, impact-oriented research philosophy.

MIT develops an AI agent trained on the VideoCAD dataset to operate CAD software via user-interface actions, transforming 2D sketches into 3D models and aiming to simplify CAD design for engineers and beginners alike.

Researchers find that advanced large language models (reasoning models) exhibit a 'cost of thinking' similar to humans by taking more time (tokens) on complex problems, revealing human-like stepwise problem-solving and internal computation parallels.

Intuit partners with OpenAI in a $100M+ multi-year deal to integrate AI-powered financial services and personalized experiences directly within ChatGPT, enhancing productivity and decision-making for consumers and businesses.

The MIT Energy Initiative conference highlights critical research priorities and collaborative strategies to address emerging energy challenges, focusing on grid resiliency, energy storage, sustainable fuels, carbon capture, vehicle electrification, and geopolitical factors influencing the energy transition.

The completion of the 2Africa subsea cable system establishes unprecedented connectivity across Africa, Europe, and Asia, driving economic growth and digital transformation through advanced infrastructure and open access collaboration.

Ax 1.0 is an open-source adaptive experimentation platform leveraging Bayesian optimization to efficiently guide complex, resource-intensive experiments across AI, infrastructure, and engineering domains.

Lambda secures $1.5B funding from TWG Global and USIT to accelerate development of gigawatt-scale AI factories and superintelligence cloud infrastructure.

A permissions change in Cloudflare's ClickHouse database caused a bot management feature file to exceed size limits, triggering widespread 5xx errors and service outages across Cloudflare's network on November 18, 2025, which were resolved through rollback and mitigation efforts by 17:06 UTC.