Helen Toner Warns AI Safety Triaged Amidst Rapid Development Timelines
Helen Toner, Executive Director of Georgetown's Center for Security and Emerging Technology and former OpenAI board member, asserts that the AI industry's safety problems are being triaged rather than solved due to aggressive 12-to-24-month development timelines set by frontier AI labs. Toner argues that while these problems are solvable in principle, the current pace of development prevents adequate attention to critical safety work, including alignment, control, and interpretability. She highlights that internal sources at companies like OpenAI and Anthropic acknowledge hundreds of necessary safety tasks but lack the time to address them comprehensively, often tackling only a few per model launch. Toner advocates for a higher evidentiary bar for risk management before training or releasing new models, focusing on frontier systems rather than a broad economic pause. She also notes that the commercialization of AI, often attributed to a change of heart, is primarily driven by the immense computational cost...