John Haugeland on the failure of micro-worlds (blog.plover.com)

In a recent exploration of John Haugeland’s philosophical insights on artificial intelligence, the author revisits Haugeland's critiques of SHRDLU, a pioneering AI program that managed simple conversations about a simulated blocks world. While Haugeland expressed dissatisfaction with SHRDLU, citing a need for more profound interaction and understanding, he acknowledged its foundational role in AI discourse. The article contrasts SHRDLU with modern large language models (LLMs) like Claude, which demonstrate a sophisticated understanding of the physical world and context, going beyond simple procedural responses. The significance of this development lies in the evolution of AI towards a more integrated understanding of the world, a concept Haugeland championed decades ago. The success of Claude in tackling tasks similar to those of SHRDLU indicates not just a leap in technical capabilities, but also a shift in what can be considered artificial intelligence. This progression raises critical conversations about the definitions of AI and intelligence itself, suggesting that LLMs can meet some of Haugeland's historical expectations, thereby bridging the gap between early AI experiments and contemporary advancements.