Show HN: Auto-Architecture: Karpathy's Loop, Pointed at a CPU (github.com)

In a groundbreaking experiment, Andrej Karpathy showcased the potential of an autonomous research loop, termed "Auto-Architecture," aimed at designing a CPU microarchitecture using a single-GPU nanochat. Over approximately 10 hours, the coding agent autonomously proposed and implemented 73 architectural hypotheses while navigating a traditional RV32IM design framework. It successfully identified 10 significant optimizations, resulting in a remarkable performance enhancement of 92% over the baseline, achieving 2.91 CoreMark/MHz, along with improved Fmax of 199 MHz and reduced LUT usage by 40%. This marks a notable advance in AI-driven design processes, demonstrating not just the loop's capability to improve existing architectures but also its potential to challenge human ingenuity. This experiment holds substantial significance for the AI/ML community as it highlights the capabilities and limitations of current AI agents in complex, real-world tasks like hardware design. The loop's reliance on rigorous verification processes underscored the importance of validation mechanisms in ensuring that AI-generated solutions adhere to required standards and do not lead to erroneous outputs. Karpathy's findings suggest a shift in focus within the industry—from crafting smarter agents to establishing robust verifiers that safeguard the integrity of AI-generated designs. The future of AI in architecture and potentially other domains will hinge less on the agents' direct performance and more on the effectiveness of the verification ecosystems that oversee them.