Vol. 2 No. 1 (2025): Large Language Models I

The issue theme is large language models (LLMs) - their capabilities, limitations, benchmarks, methods to ensure their safety and value alignment with humans, and governance. Mercer et al. and Dario Amodei offer perspective on DeepSeek. Dan Hendrycks et al. take a very hard look at the geopolitics of AGI ('a coherent superintelligence strategy'), and Rehman et al. from RAND Corp. critique their proposal. Qureshi presents a framework for analyzing timelines predicting the advent of AGI. Humanity's Last Exam is a vast and open-ended compilation of 2700 questions 'at the frontier of human knowledge' to test AGI knowledge and reasoning capability. The Road to Artificial Superintelligence is a timely survey of safety & alighment methods.

Future of Life Institute's Uuk et al. scanned the literature for practical, effective, broadly-applicable mitigations for AGI risk; their top three:  safety incident reports and security information sharing, third-party pre-deployment model audits, and pre-deployment risk assessments. Yoshua Bengio & team describe a computationally-efficient Bayesian ML program designed to assess risk probabilities of an agent's actions at runtime. What safety & alignment policies are actually in use at leading AI companies? See Anthropic's Responsible Scaling Policy

Modeling and simulation are essential tools for AGI safety & alignment. Nasim et al. offer non-coders an open-source simulator for opinion dynamics researchers to analyze influence propagation and counter-misinformation strategies in social networks that including LLM agents (see FLI's new proposal to ban models with superhuman persuasion capability). 

Preston Estep contributes to the theory of mind, examines differences between human and artificial intelligence, and looks at how non-human attributes of AI must be taken into account when predicting AGI risks.