OpenAI suspected the three fired researchers of leaking internal model designs to AI safety researcher Ryan Greenblatt, including via disappearing Signal chats
THE SO WHAT
Advanced model IP is now treated like state-level sensitive material—internal governance, NDAs, and comms monitoring are becoming as central as model evals. If you're running frontier or proprietary models, tighten your insider-risk posture this week, especially around safety teams and external researchers.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIDeepSeek, MiniMax Win Approval for Global Use, State Media Says
Beijing greenlighting DeepSeek and MiniMax for overseas deployment is another step toward a regulated export lane for Chinese foundation models. If you operate in or with China, model choice is becoming a policy decision as much as a technical one—expect jurisdictional constraints to shape your AI stack.
Applied AIDelta Electronics VP on Data Centre Demand
Power and thermal vendors like Delta expecting “strong growth for years” in AI data centers means the constraint is shifting from GPUs to grid, cooling, and capex cycles. If you’re planning large-scale AI deployment, treat power and facility partners as strategic dependencies, not afterthought procurement.
Applied AISK Data Center Arm Plans to Fund $350B AI Buildout
A $350B buildout plan from SK Hyper is another data point that hyperscale AI infra is converging with national-scale industrial projects, not ordinary cloud capex. If you depend on GPU-rich regions in Asia, start treating power, land, and local capital partners as strategic variables in your capacity planning, not afterthoughts.
Applied AILeft alone with nothing to do, AI agents started playing
MIT’s island of 13 agents that spontaneously invented games and social rituals is a reminder that open-ended objectives produce behaviors you didn’t spec. If you’re piloting agentic systems, tighten reward design and logging now—“idle time” for agents is a governance and safety variable, not just a utilization metric.