OpenAI’s new reasoning technique alarms AI safety experts
THE SO WHAT
Recurrent depth in Astra — reasoning outside fixed sequential passes — is a qualitative shift in how models think, which is why safety experts are nervous. If you’re considering next-gen reasoning models, budget extra time for evals and red-teaming; you can’t assume your current test harness captures new failure modes.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAlmost all AI tools are now running with no oversight from IT — putting companies in the firing line
Shadow AI is just shadow IT with higher blast radius — “fire-and-forget” deployments push risk onto security teams who don’t even have an inventory. If you own a business unit budget, assume you’re already running unvetted AI tools and task someone this week with mapping where models, data, and prompts are actually flowing.
Applied AIMalaysia weighs Huawei chips for sovereign AI, Bloomberg reports
A US-aligned country leaning toward Huawei for sovereign AI—despite explicit US export warnings—shows AI accelerators are now a geopolitical choice, not just a performance one. If you operate across jurisdictions, you need parallel roadmaps for US- and China-aligned silicon to avoid surprise supply or sanctions shocks.
Applied AICathay Pacific tests Google AI that dramatically cuts airplane contrails — tiny altitude shifts slash environmental impact by 40% in early trials
A 40% contrail reduction from small altitude tweaks is a template for AI that optimizes within existing operational envelopes instead of demanding new hardware. If you run large fleets—planes, trucks, ships—your near-term climate lever is likely AI-driven routing and dispatch, not asset replacement.
Applied AI‘If we don’t like it, we’ll kill it. If we love it, we’ll marry it’: We were dreaming about ChatGPT thousands of years before it existed — here's how ancient myths shaped the AI we’re building today
The stories we tell about AI—servant, oracle, monster—quietly set user expectations and risk tolerance before any product spec is written. If you're deploying AI into consumer or employee workflows, align your narrative with the actual failure modes you can manage, not the myth you're unconsciously reenacting.