Is AI Still Lacking Systems Thinking?

The author conducted experiments using AI for two separate tasks: managing personal appointments and organizing a political database. While the AI initially performed well in the personal assistant experiment, it began producing incorrect or incomplete information as the number of files grew and it was unable to search through previously written notes. Similarly, the political database experiment faltered, with the AI modeling the graph database as an RDBMS and creating duplicate data and inconsistent records.

The author questions whether sufficient instructions are required to train AI effectively. If the AI is considered intelligent, shouldn't it be capable of understanding and following broader directives without the need for minute details? This leads to a broader architectural question: should we design specialized systems or simply provide AI with tools and let it handle the work?

The author references a paper suggesting that letting AI search on its own could outperform traditional Retrieval-Augmented Generation (RAG) methods.

The author reflects on the capabilities of various AI models, including Claude Fable/Opus, Grok Bot, and GPT-6.1 Sol. Despite observing similar results across these models, the author concludes that AI has primarily been trained to solve problems by adopting the path of least resistance, rather than developing systems thinking capabilities.

The author suggests that current AI systems are essentially automated engines that can address almost any problem as long as it adheres to training constraints, but they lack the ability to integrate disparate tasks into a coherent system.