Leading AI Expert Warns World May Lack Time to Prepare for AI Safety Risks
David Dalrymple, an AI safety expert and programme director at ARIA, has warned that the world may not have sufficient time to prepare for the safety risks posed by advanced AI.
He highlighted a significant gap in understanding between the public sector and AI companies regarding the power of imminent technological breakthroughs.
Dalrymple predicts that within the next five years, most economically valuable tasks will be performed by machines at a higher quality and lower cost than humans.
He emphasized that governments should not assume advanced AI systems are reliable, and efforts should instead focus on controlling and mitigating potential downsides.
Dalrymple is actively developing safeguards for AI applications in critical infrastructure, including energy networks.
The AI Security Institute (AISI) reports that AI capabilities are rapidly improving, with performance doubling approximately every eight months.
Their findings indicate that leading models now successfully complete apprentice-level tasks 50% of the time, a significant increase from about 10% last year, and can autonomously handle tasks that typically take humans over an hour.
Self-replication tests using cutting-edge models have shown success rates exceeding 60%, although real-world self-replication outcomes remain unlikely to succeed.
