Anthropic says its AI safety scores depend heavily on how they’re measured
THE SO WHAT
A 60% vs 94% safety score swing between API and app for the same model shows how fragile headline benchmarks are—evaluation design is now a first-class product decision. If you’re buying or deploying models in sensitive domains, demand eval details by channel and context, not just a single “safety score.”
READ THE SOURCE
MORE FROM THE WIRE
FAA SMART AI system vendor and Microsoft both say ground stop isn't their fault
When a national ground stop happens, every AI system anywhere near the stack becomes a political and reputational risk surface. If you sell critical-infrastructure AI, design for forensic transparency and clear blast-radius boundaries so you can quickly prove non-involvement when things break.
Applied AIAmazon has a plan to index the world’s online shopping data
A “World Product Graph” means Amazon wants to own the canonical product ontology for the internet—whether or not it sells the item. If you’re in ecommerce, assume your catalog is being modeled as a feature in Amazon’s discovery engine and decide now whether to lean into that surface or invest in alternative demand channels.
Applied AIColette Kress Will Face Scrutiny at Nvidia
Interim leadership at Nvidia shifts the conversation from vision to capital allocation — Colette Kress will now be the public face of how $10B+ AI capex cycles get financed and prioritized. For operators, expect more explicit guidance on margins, supply commitments, and customer mix as she leans into the finance narrative.
Applied AINvidia’s Board Failed Jensen Huang: Randall Peterson
Board-governance debates around Nvidia are a reminder that AI infrastructure risk is now board-level, not just product-level. If you’re building on a single vendor’s roadmap, assume leadership volatility and board pressure as part of your dependency model.