Systems Reading Group with humans& - ARC-AGI

​If you're excited about ways to define and evaluate human-like intelligence in AI, this session is for you!

​For this session, we will be joined by Jeremy and Niloofar from humans&, a human-centric AI lab with the goal of strengthening relationships by changing how we interact with AI.

​ARC-AGI is a benchmark which measures an AI system's ability to learn and generalize new knowledge, in ways similar to a human. As we build better models and systems, this benchmark serves as an indicator of progress towards this goal.

​In this session, we will dive into the background and principles of this benchmark, and cover promising techniques which made it to the leaderboard, including Jeremy's highest score where he used techniques like multi-agent collaboration with evolutionary test-time compute.