The Expensive-Loop Problem (www.lesswrong.com)

Recent discussions around large language models (LLMs) highlight a critical distinction between "cheap-loop" and "expensive-loop" tasks. Cheap-loop domains allow for quick and inexpensive feedback, enabling LLMs to excel in tasks that are easily computable and trainable. Conversely, expensive-loop domains—like drug discovery, scientific research, and complex policy-making—require significant resources and time for accurate feedback, posing unique challenges for AI systems. Current models struggle in these environments due to their inability to effectively manage proxies and theories essential for optimizing experiments, often leading to inefficient resource allocation and poor performance. The implications of this divide are profound for the AI/ML community. It suggests that while models may perform well in straightforward, data-rich scenarios, their capabilities diminish significantly when faced with real-world complexities that require critical thinking and strategic planning. To bridge this gap, researchers need to enhance model training with frameworks that support theory building and proxy management, enabling LLMs to navigate the intricate demands of expensive-loop tasks successfully. This paradigm shift could unlock new advancements in AI applications across various high-stakes domains, ultimately refining how we leverage AI in critical fields such as healthcare and scientific discovery.