JoinARGS - AI Alignment

Tech profiles such as Elon Musk and Geoffrey Hinton have voiced their concern about risks from unaligned AI. Join a talk by Simon Koser on AI alignment, the research field that attempts to ensure AI systems align with human values. The talk will start with an introduction to alignment, exploring the existing ML paradigm and why it might fail.

Learn about possible pitfalls like reward hacking, misgeneralization, reward misspecification, latent knowledge, and instrumental convergence. We'll delve into examples of how human evaluators' intuition can be tricked by algorithms, leading to ineffective feedback.

Discover potential solution agendas such as interpretability, RLHF/RLAIF, "Constitutional AI", adversarial techniques and more. We'll do a deep dive case study on ELK, one approach for AI alignment and see where it fails.

We'll also address other vital considerations: the feasibility of an AI pause, China's role in AI development, computational overhang, and the importance of incentives. Gain actionable advice on AI alignment, and a better grasp of the key research organisations in the field.

This talk provides a good stepping stone for anyone keen on replicating papers, engaging in AI alignment research, and working in the field. Get up to speed on the main alignment research areas and take a step towards safer AI. Lunch will be served to the first 30 people that arrive.