In this episode of Bit & Being, we present an analysis of research from Harvard and MIT that questions whether AI models, particularly Large Language Models (LLMs), genuinely understand the underlying "world models" of the data they process, or if they merely predict patterns. The researchers conducted an experiment training a transformer model on planetary orbital trajectories to see if it could not only predict future positions (Kepler-like ability) but also deduce the gravitational forces (Newton-like understanding). The study concluded that AI models, even advanced commercial ones, fail to learn generalizable physical laws and instead develop brittle, non-physical heuristics that are only sufficient for the specific prediction tasks they were trained on. This suggests that LLMs currently lack deep understanding and emergent intelligence, highlighting a significant limitation in their current capabilities.

New tracks tagged #kepler versus newton