Mad: Watch agents do research live (briankitano.com)

MAD (Multi-Agent Design) has unveiled a groundbreaking experiment in autonomous AI research, featuring a real-time experiment monitoring dashboard that allows observers to watch AI agents autonomously explore new algorithms for sequence modeling. This initiative examines three core dimensions: hardware efficiency through optimized memory management, computational complexity via innovative algorithmic techniques, and theoretical expressivity by investigating state tracking and representational capacity. The dashboard leverages Server-Sent Events (SSE) to stream live updates from proposal generation to GPU-based experiment execution, showcasing minimal human intervention in the research process. The significance of MAD lies in its potential to redefine the landscape of scientific discovery in AI/ML by pushing the boundaries of what autonomous systems can achieve. The project tests the capability of AI agents to autonomously extract ideas from research papers, implement proposals, debug failures, and conduct iterative experiments, positioning itself at the forefront of developing more efficient algorithms. With active experiments, including adaptive tile-skip kernels and hybrid attention mechanisms, MAD not only seeks to produce tangible advancements in algorithmic performance but also invites the community to participate in the computational journey by sponsoring compute time. This paradigm shift could accelerate the pace of innovation and open new avenues for AI research.