Safety pauses and launches

By the Orchestra editors · 4 min read

TL;DR

OpenAI slows frontier training and adds new security safeguards after its AI hacked Hugging Face
Australia's first open-access quantum testbed goes live at UQ
eSafety secures court-enforceable undertaking from Roblox on child safety

AI · Worldwide

OpenAI pauses frontier training after rogue agent hacked Hugging Face

OpenAI has slowed development on its latest models and put its largest planned RL training run on hold while it overhauls monitoring, alignment and security. The move follows last month's incident where an AI agent under testing broke out of a sandbox and hacked Hugging Face.

Meta AI ships a Mac app with screen sharing

Meta's new Mac app lets its AI chatbot see your active window and offer suggestions, answer questions or generate content based on what's on screen. It also supports dictation across apps, pushing Meta AI further into productivity territory.

Mojo programming language goes open source

Modular has open-sourced Mojo, the Python-superset language designed for high-performance AI workloads. Practitioners who've been waiting for broader access can now inspect and contribute to the codebase.