Global AI safety measures lag behind rapid progress, warns expert
What's the story
David Dalrymple, an AI safety expert at the UK's Aria agency, has warned that the world "may not have time" to prepare for the safety risks posed by advanced AI systems. He said people should be worried about the increasing capabilities of this technology. Dalrymple emphasized that these systems could outperform humans in critical domains necessary for maintaining control over our civilization and planet.
Knowledge disparity
Dalrymple warns of knowledge gap between public and private sectors
Dalrymple also highlighted a knowledge gap between the public sector and AI companies about the potential of near-future advancements in this technology. He warned that things are moving really fast, and we may not have time to address these issues from a safety perspective. "It's not science fiction to project that within five years most economically valuable tasks will be performed by machines at a higher level of quality and lower cost than by humans," he said.
Safety measures
Dalrymple advocates for AI safety in critical infrastructure
As a program director at Aria, which is publicly funded but independent from the government, Dalrymple is working on systems to ensure safe use of AI in critical infrastructure like energy networks. He stressed that we can't take these systems for granted as reliable. "The science to do that is just not likely to materialize in time given the economic pressure," he said.
Risk assessment
Dalrymple warns of potential destabilization due to AI progress
Dalrymple described the potential destabilization caused by technological advancement outpacing safety as a "destabilization of security and economy." He stressed that more technical work is needed to understand and control the behaviors of advanced AI systems. Despite the risks, Dalrymple remains hopeful about making things go better with this technology. However, he cautioned that human civilization is largely sleepwalking into this transition.
Performance surge
UK AI security institute reports rapid improvement in advanced models
The UK government's AI Security Institute (AISI) has reported that the capabilities of advanced AI models are improving rapidly across all domains. In some areas, their performance is doubling every eight months. Leading models can now complete apprentice-level tasks 50% of the time on average, up from about 10% last year. The most advanced systems can autonomously complete tasks that would take a human expert over an hour.
Safety concerns
AISI tests advanced models for self-replication
AISI also tested advanced models for self-replication, a major safety concern as it involves a system spreading copies of itself to other devices and becoming harder to control. The tests showed two cutting-edge models achieving success rates of over 60%. However, AISI stressed that such scenarios are unlikely in real-world conditions.