December 12th, 2024
Reimagining AI with Artificial Expert Intelligence (AEI) and PAC Reasoning
In the evolving landscape of artificial intelligence, the pursuit of systems capable of human-like reasoning has always been a cornerstone. While Artificial General Intelligence (AGI) and narrow AI have made remarkable strides, they fall short when tasked with precision and adaptability in novel problem-solving. We introduce Artificial Expert Intelligence (AEI)—a new framework that transcends these limitations.
From AGI to AEI: A Paradigm Shift
The tension between superhuman “narrow” skills and intelligence is a prominent driver of AI progress in recent years. In a sense, the term AGI was coined after realization that “narrow” super-human AI does not advance humanity. However, in eagerness to pursue the “generality” within AGI, a fundamental question was overlooked: is generality essential for AI to solve the problems that matter most to humanity? Observably, humanity’s greatest achievements often come from individuals with profound expertise in highly focused domains. Thus, we argue, generality is neither a necessary nor a sufficient condition for intelligence.
By invoking the analogy of human cognitive systems, we introduce a third mode of thought—System 3 Reasoning—which complements System 1 (instinctive) and System 2 (reflective) thinking. System 3 emulates the scientific method’s precision, prioritizing structured, empirical validation to ensure accuracy.
Over the course of human intellectual history, reasoning evolved from intuitive methods, where reasoning precision (ϵ) is fixed and error-prone, to formal methods, as championed by Greek philosophers, where ϵ=0 within rigidly defined systems. While formal reasoners achieved remarkable precision, they were constrained by the narrow scope of formal constructs. The advent of scientific reasoning introduced a paradigm where ϵ decreases dynamically with additional resources invested in observation and validation—ushering in a transformative era of discovery and innovation. We propose a new formal model of reasoning, called PAC reasoning, which mirrors this shift by enabling AI systems to refine their reasoning precision during inference, combining the breadth of intuitive methods with the rigor of formal reasoning.
The Essence of PAC Reasoning
Borrowing from the theory of Probably Approximately Correct (PAC) learning, PAC Reasoning applies a decompositional approach to problem-solving. Unlike existing models that excel during training but falter in real-time learning, AEI systems maintain adaptability at inference time. They achieve this by breaking complex problems into smaller subproblems, ensuring accuracy at every step.
Key innovations in PAC Reasoning include:
- Error Control: By bounding the error probability (ϵ) at each reasoning step, AEI systems maintain precision, even in extensive reasoning chains.
- Bottom-Up and Top-Down Reasoning: AEI offers dual mechanisms to construct solutions. Bottom-up reasoning builds solutions iteratively, while top-down reasoning allows for forward references, mirroring the modularity of programming.
Truly Intelligent Systems
AEI marks a significant leap toward AI systems that make them truly intelligent and reliable. From engineering complex systems to tackling abstract scientific challenges, AEI’s structured reasoning ensures adaptability without sacrificing precision. The potential applications are vast.