Solving AI's 'Jagged Intelligence' Problem with Structured Knowledge

Vinay Chaudhri argues that AI's 'jagged intelligence'—the inconsistency and unreliability of LLMs—stems from their reliance on inferring knowledge from vast datasets rather than possessing explicit, structured knowledge. Current AI models guess based on patterns, leading to errors ranging from comical to catastrophic. To solve this, Chaudhri proposes giving AI models access to a more powerful, structured, and human-like stock of knowledge.

He draws an analogy to human learning, where pattern recognition is supplemented by codified knowledge—rules and relationships that anchor our understanding of the world. Frontier AI labs are already experimenting with this approach, such as bolting mathematical knowledge onto LLMs to improve their math skills. Chaudhri argues that simply adding more data won't overcome the fundamental challenge of jagged intelligence.

Instead, he advocates for building a public database of formal knowledge spanning various disciplines. This database would provide AI models with rigidly described concepts, constraints, rules, and relationships, enabling them to reason more reliably and avoid errors. The growth of companies like Scale AI, which provides high-quality data for training AI models, suggests that translating human expertise into machine-readable form is a viable path forward. This approach could shape not just what AI can do, but what it comes to treat as true.

Transparency note: This analysis was conducted by an AI, using information from an opinion piece by Vinay Chaudhri. The AI strives for objectivity and accuracy in its reporting.