Open-weight AI caught the frontier on capability. On safety, it didn’t
THE SO WHAT
Open-weight models like GLM-5.2 narrowing the capability gap while lagging on safety means the risk surface is decoupling from the big labs’ governance. If you’re adopting open weights, you own evals, guardrails, and abuse controls end to end—treat them as a security product, not a free upgrade.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIDevelopers in Africa are increasingly choosing Chinese open-source AI models over US models, saying they are downloadable, easier to customize, and much cheaper
African teams building tools like Sunflower on Chinese open-source models because they’re cheaper and easier to customize is a concrete shift in AI soft power. If you’re a US or European model provider, local language support, offline access, and permissive licensing are no longer nice-to-haves in emerging markets — they’re the entry ticket.
Applied AIA University Used AI to Police Its Entrance Exam. Now 58,000 Students Need a Do-Over
An AI proctoring failure that forces 58,000 retakes is a governance story, not a model story — risk moved from cheating to institutional legitimacy overnight. Any high-stakes AI deployment — exams, hiring, credit — now needs a rollback plan, appeal process, and human override baked in before launch.
8 working parents share how they use AI to give them more time with their kids
When working parents use ChatGPT and Claude for scheduling and admin, AI stops being a novelty and becomes invisible infrastructure in the household. For consumer and productivity products, the bar is now “buys back an hour of life” — not “adds another app to manage.”
Applied AIAnthropic is hiring an AI chip design team
Model labs hiring in-house chip teams is vertical integration — co-designing Claude and silicon to squeeze more performance per watt and per dollar. If you’re a heavy AI user, expect a future where “which model?” and “which hardware?” are a single architectural decision, not two separate vendor choices.