Hill-climbing Alignment using Humans | Sampura Research

​As AI systems grow more capable, the judges we use to train them increasingly fall short. Most of the field responds by replacing humans with LLM judges; we think humans and AIs have complementary strengths, and that human oversight is an uncorrelated layer of defence that AI-only judges can't provide. In this talk we'll share how we're building better judges: a leaderboard of 50+ alignment-relevant datasets, and methods that route, decompose and assist across humans and AIs to hill-climb it. Sampura Research is a London-based nonprofit, founded by Google DeepMind alumni, building Human-AI Complementarity for AI alignment.... more