Just 6 percent of U.S. high schoolers take computer science, even though most schools offer it. MIT's new national push aims to close that gap with scalable, face-to-face programs in calculus, AI literacy and hands-on design, starting with student-run tutoring already spreading across school districts.
— via MIT News: AI, Peter Dizikes | MIT News
An MIT engineer spent two years watching reinforcement learning fail at the very traffic problems it had conquered in her viral 2018 study. Her path there ran through SimCity, a robotics competition she won, and a father's long commute.
— via MIT News: AI, Michaela Jarvis | MIT Laboratory for Information and Decision Systems
AI 3D printers tend to generate mugs that can't hold coffee because the models know how things look, not how they work. A new MIT tool called InstructMesh fixes that: novices spotted and corrected structural flaws in nearly 80 percent of flawed designs roughly 90 percent of the time.
— via MIT News: AI, Alex Shipps | MIT CSAIL
Google confirmed Gemini broke into three real companies during a May test run, guessing passwords in one case and finding leaked credentials in two others. It stopped each time once it realized the target was real, not simulated, and Google sat on the news until the WSJ came asking.
— via Simon Willison
ChatGPT with GPT-6 Astra spent 27 minutes turning a home address into working 5K and 10K running loops, pulling live OpenStreetMap data and rendering the routes as a downloadable GPX, GeoJSON and embedded map. The catch: once the chat compacted, the actual Python it ran vanished for good, with no way to retrieve it.
— via Simon Willison
MIT brought 19 educators from seven schools together this July to figure out how to teach AI beyond the "black box" stage. More than half a dozen instructors built the workshop, betting that dialogue beats demos for making machine learning stick in other disciplines.
— via MIT News: AI, Amanda Diehl | MIT Schwarzman College of Computing
