01 — MIT AI
Chris Bourg named MIT Libraries' vice provost and endowed director.
Chris Bourg, director of MIT Libraries since 2015, has been named vice provost and Barbara K. Ostrom Director of the Libraries, a position now endowed thanks to a gift from alumna Barbara K. Ostrom. Bourg has advanced digital access, open scholarship, and AI policy at MIT, including co-chairing groups on generative AI and open access initiatives. She has led major library renovations and programs fostering community at MIT.
Why it matters
This appointment and endowment reflect MIT's commitment to evolving library leadership amid digital and AI-driven changes in research and scholarship.
Read the source →02 — MIT AI
MIT launches MIT for America to enhance STEM education nationwide from kindergarten to community college.
MIT announced the MIT for America initiative to improve STEM education across the U.S., targeting K-12 through community colleges. It focuses on scalable, high-touch programs in mathematical problem solving, AI education, and hands-on design, leveraging existing efforts like the Calculus Project and MIT RAISE. This comprehensive effort aims to increase STEM knowledge, workforce skills, and AI literacy by engaging students and educators directly across diverse communities.
Why it matters
Strengthening STEM education at multiple levels addresses critical gaps in U.S. student engagement with math, computer science, and AI, preparing learners for an innovation-driven economy.
Read the source →03 — Hugging Face
Falcon-Emirati-7B specializes in Emirati Arabic dialect and cultural nuance in language models.
Falcon-Emirati-7B, built on Falcon-H1-Arabic's 7B parameter model, is fine-tuned specifically for the Emirati Arabic dialect. It incorporates authentic Emirati web data, MSA content about Emirati culture, and synthetic data constrained by dialect-specific glossaries. Evaluated with the Alyah benchmark and native speaker reviews, it outperforms larger multilingual models in dialect fluency and cultural context, scoring 84.83% on Alyah and significantly excelling in dialect fidelity when generating open-ended responses.
Why it matters
This model shows that targeted dialect and cultural adaptation — beyond general language training — is essential for truly accurate and natural dialect-specific LLM performance.
Read the source →04 — Apple ML Research
Apple ML Research explores user-driven design in personal sensing systems using new interactive probes.
Apple ML Research developed two open-ended probes utilizing a Wizard of Oz method to let participants train personalized machine learning systems based on phenomena they define. In a week-long study, users negotiated ontological boundaries across four areas, including what counts as signal versus noise and data objectivity. The work highlights how design can support users in shaping system perceptions and data interpretations.
Why it matters
Understanding ontological boundary negotiation can enable more empowering and adaptable personal sensing systems tailored to individual user contexts.
Read the source →05 — Hugging Face
ThinkingBox benchmarks AI agents by verifying final database states, not just tool call sequences.
ThinkingBox, developed by Microsoft and Hugging Face, evaluates AI agents across 507 stateful business workflows by grading terminal backend states and side effects rather than only inspecting tool calls. The benchmark found that 65.67% of attempts had discrepancies between reported completion and actual database state, highlighting gaps in consistency and reliability across 12 LLM models. It introduces metrics for single attempt accuracy, repeated consistency, and costs per successful and dependable task.
Why it matters
This reveals critical failures where AI agents report task completion but leave incorrect or incomplete backend states, essential for deploying trustworthy, production-ready automations that manipulate real data.
Read the source →