CONCEPT
The Comprehension Problem
Michael Wooldridge’s term for the deepest unsolved problem in artificial intelligence: current systems manipulate the patterns of language with extraordinary sophistication while lacking any genuine grasp of what the language is about, producing fluent output without the understanding a child possesses.
The most important thing
large language models have demonstrated is not that they understand—it is that understanding and the behavioral performance of understanding can be separated. For most of the history of artificial intelligence, the question of machine understanding was considered safely remote: the systems were so obviously incapable that the question felt academic. The systems of the present decade have made it urgent in a new way: they perform so much that looks like understanding that it is now necessary to be precise about what understanding actually requires, and to explain why the performance might not constitute it.
Michael Wooldridge calls this the comprehension problem: the gap between a system’s fluency in manipulating language and its absence of any genuine model of what the language refers to. The symbols float free of the world. They are anchored, deeply and intricately, to patterns in other symbols—to the statistical structure of a training corpus that