World running out of time for AI safety: AI expert David Dalrymple

David Dalrymple, a programme director at the UK's Advanced Research and Invention Agency (ARIA), has warned that the window to prepare for the safety risks of advanced AI is rapidly closing. Speaking to The Guardian, Dalrymple emphasized that AI development is moving at a pace that prevents society from assuming these systems are reliable. He noted that we face being outcompeted in domains essential for maintaining control over civilization, leading to a potential destabilization of global security and economies.

The expert cautioned that within five years, most economically valuable tasks will likely be performed by machines with higher quality and lower cost than humans. Dalrymple specifically projected that by late 2026, AI systems will be capable of automating an entire day of research and development work. This milestone would trigger a "further acceleration of capabilities," as AI begins to self-improve in the fields of mathematics and computer science, creating a recursive feedback loop of intelligence growth.

ARIA, though publicly funded, operates independently and focuses on safeguarding critical infrastructure like energy networks from AI-related failures. Dalrymple described human civilization as "sleepwalking" into this high-risk transition, driven by intense economic pressure that prioritizes speed over safety science. He advised that the most realistic current path is to focus on technical work to control and mitigate the downsides, as a complete guarantee of reliability may not materialize in time.