engblogs

Case study of avatarin leveraging GPT-Realtime to deploy a 24/7 multilingual, voice-enabled retail agent for Yamada Denki, integrating retrieval-augmented generation, multimodal (voice/text/visual) support, and proactive conversation design to extend expert sales coverage beyond hours.

GPT-5.6 advances price-performance by delivering Luna, Terra, and Sol with improved efficiency and new Fast mode, reducing costs while boosting throughput for enterprise AI workloads.

GenRec shows how an LLM-backed ranking model trained on Netflix-specific data and verbalized contexts, with a catalog-aware scoring head and reward-weighted objectives, can outperform traditional production rankers while reducing labeled data and serving costs.

GenRec shows how an LLM-backed, catalog-aware ranking model verbalizes user histories and context into text, post-trains a Netflix-adapted foundation LLM for ranking, and delivers cost-efficient, prefill-only recommendations with strong online and offline gains.

GenRec demonstrates an LLM-backed ranking approach that verbalizes user history and item context, post-trains a Netflix-adapted foundation LLM with a catalog-aware head, and uses reward-weighted objectives and context engineering to deliver cost-efficient, long-term aligned recommendations at Netflix.

MIT students and postdocs connect research to policy on Capitol Hill through Congressional Visit Days, illustrating how federal funding, science communication, and policymaker engagement shape science-driven policy.

Daniela Rus is honored with the Bavarian Minister-President's High-Tech Prize for pioneering real-world robotics and AI—from self-organizing robot collectives and soft robotics to autonomous mobility and brain-inspired, efficient neural architectures.

MIT's Congressional Visit Days reveal how scientists translate research into policy on Capitol Hill through training, direct engagement with congressional staff, and advocacy for sustained federal funding.

MIT CSAIL director Daniela Rus wins Bavaria's 2026 High-Tech Prize for pioneering robotics, soft robotics, autonomous systems, and brain-inspired artificial intelligence that enable real-world autonomous machines across transportation, agriculture, medicine, and environmental monitoring.

Salesforce describes how the TPIL team's agentic engineering enablement approach scales thousands of engineers by building a shared language and a four-level Proficiency Level Framework that guides behavior change from AI-assisted to AI-native productivity.

Salesforce scaled agentic engineering across thousands of software engineers by building a four-level Proficiency Level Framework and a shared language for human–AI collaboration, anchored in a learning culture that enables AI-native practices at scale.

Orchestrating 100k battles of untrusted agent code across four hot-loaded GPUs with a single reloadable Caddy chokepoint, runtime-resizable concurrency, and strict containment to keep benchmarks on schedule.