The Hitchhiker's Guide to System One Models (arxiv.org)

A new comprehensive guide titled "The Hitchhiker's Guide to System One Models" has been released, focusing on innovative decision-making architectures that exceed the limitations of traditional Large Language Models (LLMs). While LLMs excel in reasoning and content generation, they often suffer from slow, costly autoregressive inference methods that hinder real-time applications requiring rapid decision-making. System 1 models, described in the guide, provide a solution by enabling direct mapping from inputs to decisions in a single forward pass, ideal for tasks like routing and classification where sub-second response times are critical. This guide categorizes System 1 decision models into five families and offers extensive details on their architectures, training methodologies, and deployment strategies. A key theme emphasizes the importance of calibration in decision-making, addressing phenomena such as IIA violations and ordinal-scale biases. With in-depth coverage of 18 application domains and a vast open-source ecosystem, the guide serves as a valuable resource for researchers and engineers seeking to enhance AI systems with efficient and calibrated decision-making capabilities. Its hands-on resources, including runnable code and self-assessment questions, position it as a pivotal reference in the evolving landscape of AI.