today's digest covers practical ai tools and research, from faster optimization at linkedin to new ways of testing and controlling ai behavior. we also see privacy concerns around facial recognition and a scaled-back ai oversight order.
- linkedin speeds up extreme-scale optimization with pytorch gpus - this matters because it shows how gpus can tackle massive linear programs, making large-scale optimization more practical for industry.
- microsoft assert tests ai behavior from text descriptions - this matters because it lets developers define ai behavior rules in plain language, making testing more accessible and reducing the need for custom code.
- google adds fake call detection to android phones - this matters because it directly addresses the growing threat of ai voice impersonation scams, protecting users without requiring any action on their part.
- amazon sued over ring facial recognition privacy - this matters because it highlights ongoing legal challenges around facial recognition and consent, potentially affecting how companies deploy such features.
- trump signs narrower ai oversight order after industry pushback - this matters because it shows how industry pressure can shape ai regulation, reducing the advance notice period for model reviews from 90 to 30 days.
other stories today include microsoft's agent control spec for taming ai behavior, anthropic expanding ai bug hunting to critical infrastructure, and openai adding job-specific plugins to codex. research highlights include a hybrid transformer for das event recognition and a method to optimize vlm reward models with few demonstrations.