The Advanced Research and Invention Agency (ARIA), a British R&D unit, has revealed plans for "quantitative safety guarantees" in AI. ARIA aims to develop standards similar to those in nuclear power and aviation, providing probabilistic guarantees that no harm will occur with machine learning actions. The agency will invest £59 million in the project and plans to demonstrate a scalable proof-of-concept in a specific domain, such as electricity grid balancing or supply chain management. The ultimate goal is to secure high-stakes AI applications, like critical infrastructure improvement and clinical trial optimization. ARIA's approach involves a "gatekeeper" AI that ensures other AI agents operate within defined guardrails. Program director David Dalrymple, who co-invented cryptocurrency Filecoin, has combined commercial and academic concepts to create this approach.

Deep interdisciplinary collaboration will be crucial, making ARIA similar to DARPA, the US Pentagon's research unit. ARIA's project has drawn inspiration from DARPA's HACMS program, which aimed to create unhackable software using formal verification. The concept also aligns with the call for "quantitative safety guarantees" made by computer scientist Yoshua Bengio. The success of this project could potentially address questions about ARIA's budget and funding. ARIA has received £800 million over five years, a fraction of other government research bodies, and aims to reduce AI training costs significantly with another recent program.