World Models: Why the Next AI Revolution Won't Come from ChatGPT

Got story updates? Submit your updates here. ›

The article discusses how current large language models (LLMs) like ChatGPT, despite their impressive abilities, are fundamentally disembodied and lack a coherent understanding of the physical world. It uses the example of an AI navigating Manhattan streets, which would be easily confused if 1% of the streets were closed for construction, to illustrate the 'bag of heuristics' problem where the AI has memorized routes without understanding the underlying street network map.

Why it matters

The author argues that the next major breakthrough in AI will not come from further improvements to LLMs, but rather from developing AI systems that can build 'world models' - coherent representations of the physical and causal structure of the world. This is seen as a critical step towards achieving Artificial General Intelligence (AGI).

The details

Current LLMs like ChatGPT are brilliant at manipulating symbols and language, but are completely blind to physical reality. They have been trained on vast amounts of text data, but lack any embodied understanding of the world. This is exemplified by the fact that they cannot even tell which object is closer - a coffee cup or a keyboard. The author uses the 'Manhattan Street Test' as an illustration - an AI that can navigate Manhattan flawlessly, until 1% of the streets are closed for construction, at which point it becomes completely lost. This demonstrates the 'bag of heuristics' problem, where the AI has memorized routes without understanding the underlying street network map.

• The article was published on January 27, 2026.