In a recent opinion in TiCS [1], we asked: ‘What do large language models (LLMs) know?’. We answered by granting LLMs instrumental knowledge: that is, knowledge gained through using the instrument of next-word generation. We then explored how this type of instrumental knowledge could be related to the more ordinary kind of worldly knowledge exhibited […]

In their recent Opinion in TiCS [1], Yildirim and Paul propose that large language models (LLMs) have ‘instrumental knowledge’ and possibly the kind of ‘worldly’ knowledge that humans do. They suggest that the production of appropriate outputs by LMMs is evidence that LLMs infer ‘task structure’ that may reflect ‘causal abstractions of… entities and processes […]