OpenAI ignored staff warnings on shooter's threats

arXiv

A new methodology creates thousands of synthetic computer environments to train AI agents on realistic, long-horizon productivity tasks spanning simulated months of work. The approach shows promise for scaling agent training to billions of diverse user scenarios through self-improvement and reinforcement learning.

WSJ

OpenAI leadership ignored employee warnings about a mass shooter's use of ChatGPT to plan threats, choosing internal handling over police notification—a decision that will reshape AI safety and reporting standards in regulatory conversations.

Science

OpenAI's o1-preview outdiagnosed two physicians on real ER cases in a Harvard study, demonstrating LLMs' potential for clinical decision support and diagnostic accuracy using standard medical data.

The New York Times

The Pentagon excludes Anthropic from its expanded AI program despite new defense contracts featuring identical weapons restrictions, signaling that corporate relationships and geopolitical factors may override consistency in defense AI partnerships.

xAI

xAI's Custom Voices feature allows instant voice cloning with verification safeguards, enabling enterprises and creators to use personalized voices at scale across TTS and conversational AI applications without extra costs.