The 'Godfather of AI' backs a new watchdog plan to track OpenAI and Anthropic's AI risks from the inside
THE SO WHAT
External evaluators with inside-the-lab access, backed by names like Geoffrey Hinton, move AI safety from blog posts to quasi-regulatory oversight. Enterprise buyers should expect ‘independent evaluation’ badges to become part of vendor due diligence, especially for high-stakes deployments.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic has quietly built a biology lab, Reuters reports
A frontier lab standing up a wet lab and explicitly saying it’s not for drug discovery points to one thing: they want ground-truth data on biological risk, not just better bio tools. For anyone building AI in sensitive domains, expect the bar for in-house domain expertise, experimental validation, and safety signaling to rise fast.
Applied AIHuawei says AI will burn 100,000 times more tokens by 2035
A forecast of 100,000x more tokens by 2035 is less about precision and more about direction—Huawei is telling operators to plan for orders-of-magnitude more inference, not incremental growth. That should push CIOs to stress-test network, storage, and power plans against worst-case usage curves, not last year’s dashboards.
Applied AIOpenAI’s highest robotics salary is $500,000, and it is not for building robots
A $380,000–$500,000 salary for a robotics-focused ML engineer building data and training pipelines makes it clear: the leverage is in foundation models and infra, not the metal. Robotics startups should budget accordingly—top-tier model and data talent will cost frontier-lab money or demand equity that reflects it.
Applied AIHuawei’s chairman says Chinese AI can’t yet see the risks it should fear
Huawei’s chair arguing that Chinese AI developers haven’t yet ‘felt’ the same AI risks as US labs highlights a divergence in lived experience driving safety culture. For multinationals, that means governance frameworks and red lines can’t just be copy-pasted across regions—you’ll need localized risk narratives and controls.