In this talk, Helen Toner discusses the concept of 'jaggedness' in AI development, highlighting that while AI models are improving, they also exhibit significant weaknesses in certain areas. She presents examples of AI performance...
About:
Helen Toner is an AI policy expert at CSET, interested in AI systems and national security, sharing speculative ideas on her Substack.
Website:
Specializations:
Interests:
Incoming Links:
Outgoing Links:
Subscribe to RSS:
The post discusses the implications of personalized AI systems, drawing parallels with past issues in social media and targeted advertising. Miranda Bogen, an expert in AI governance, highlights the risks associated with AI's incr...
The post is a transcript of a talk given by Helen Toner at the Technical Innovations for AI Policy Conference, discussing the contradictory narratives surrounding AI advancements. Toner highlights three key debates: the potential ...
OpenAI announced a partnership with the UAE to build large-scale AI data centers, claiming it is rooted in democratic values. However, the UAE has a poor record on civil liberties and political rights. The deal is seen as a signif...
The text discusses the AI safety community and its approach to AI safety ideas, critiquing the community's stasist tendencies and advocating for a dynamist vision for safe superhuman AI. The author draws on Virginia Postrel's book...
The post discusses the importance of verifiable rewards, generalization, and scaling reasoning training in determining the progress of AI in the next couple of years. It explores the questions of where else AI can automatically gr...
The text discusses the risks of AI misuse and the idea of nonproliferation as a solution. The author argues that nonproliferation is not the right approach and suggests a resilience-focused alternative. The post explains the diffe...
The text discusses the concept of AI alignment and suggests using the term 'steerability' instead. It traces the history of the term 'alignment' and how it has evolved to refer to different aspects of controlling AI systems. The p...
The text discusses the rapid advancement of AI and the changing timelines for the development of human-level AI. It contrasts the previous debates about the timeline for advanced AI with the current situation, where the short time...