engblogs

MIT names David Autor as head of the Department of Economics, positioning the department to harness AI-enabled research on labor-market impacts of technological change and globalization, including job polarization and inequality.

Masked IRL uses large language models to clarify ambiguous robot prompts and distill essential details from demonstrations into safe, efficient motion planning with fewer examples.

David Autor is named head of MIT's Department of Economics, positioning the department to advance labor economics research and AI-enhanced teaching amid globalization, technological change, and rising inequality.

Codex-powered agentic AI is transforming knowledge work from short, single interactions to long-horizon, cross-department tasks, driving rapid adoption across developers and non-developers and reshaping the future of work.

Explores how agentic AI and Codex transform knowledge work from short, single interactions to long-horizon, multi-agent task orchestration across OpenAI and other organizations, driving broad non-developer adoption and a new productivity paradigm for the future of work.

Codex-enabled agents are transforming work by moving from single chatbot interactions to long-horizon, cross-functional task orchestration across departments, reshaping productivity with agentic AI and automated workflows.

An in-depth look at Continual Learning Bench, analyzing stateful vs. stateless evaluation, gain-based metrics, and the shift toward context management and parametric learning for truly adaptive AI.

A deep dive into continual learning benchmarks and measurement—showing how Continual Learning Bench quantifies true online learning across sequences, contrasting context-management with parametric approaches and highlighting the gain metric.

Profiling and a five-step optimization stack (quantized communications, balanced sharding, bandwidth-aware embedding reshaping, and 2D parallelism) enable near-linear multi-node training scalability for Pinterest's embedding-heavy foundation models.

A data-driven playbook for achieving near-linear multi-node training scalability of embedding-heavy foundation models, combining EFA-enabled networking, quantized communications, balanced sharding, bandwidth-aware embedding reshaping, and topology-aware 2D parallelism.

Murakkab automates the design and dynamic optimization of agentic workflows to maximize speed while minimizing energy usage and cost on cloud platforms.

A concise look at how curiosity-driven science, supported by steady public investment and MIT-led collaboration, powers American innovation, safety, and prosperity.