Compiler for LLMs, world models, and AGI (arxiv.org)

A new tool called Auto has been unveiled as an "AGI compiler," designed to enhance the efficiency of large language models (LLMs) by converting real-time agent behaviors into verified programs. This innovative system records the actions of LLM agents, identifies deterministic behaviors, and distills them into cognition binaries—optimized WebAssembly artifacts. The significance of this advancement lies in its ability to transform novel experiences into permanent skills at a fraction of the cost, exemplified by a drastic reduction in operational costs from 59 to just 2 micro-dollars per item, all while maintaining high accuracy levels (96.9% parity). Auto introduces a testbed known as AUTO-BENCH, demonstrating that 87.1% of captured agent behaviors are deterministically reproducible. However, it also highlights challenges, such as guard mislabeling and fidelity issues that can impact the accuracy of compiled outputs. This development promises to reshape how LLMs and other agents continuously learn and adapt, making strides toward more reliable artificial general intelligence (AGI) systems. By leveraging the insights from compiled behavior and maintaining strict measures for performance verification, Auto could represent a critical step forward in achieving efficient, scalable AI learning systems.