Kanjun Qiu and Josh Albrecht: Generally Intelligent

Avalon Benchmark

The Avalon Environment serves as a cutting-edge benchmark for AI generalization, offering a unique platform for researchers to construct diverse tasks. Kanjun Qiu describes Avalon as a fast 3D simulator, running at 7000 steps per second on a single GPU, designed to challenge AI agents with tasks that mimic real-world problem-solving scenarios 1. Josh Albrecht emphasizes the importance of tuning baselines and integrating various approaches from the literature to enhance performance 2. He notes, "We always want to make sure we're really getting our baselines really good, and then we're excited about self-supervised learning." 2

   

Skill Efficiency

Skill acquisition efficiency is crucial in developing general-purpose AI, focusing on the ability to learn quickly and adapt to new tasks. Josh Albrecht highlights the importance of learning within a lifetime, contrasting it with prolonged training on singular tasks 3. He explains, "The ability to learn within your lifetime and learn very quickly and acquire skills very efficiently is actually, like, a really core thing that I think a lot of our current systems are really missing." 3 Kanjun Qiu adds that their approach is not about designing specific architectures but creating a comprehensive array of tasks to cover various mental capabilities 4.