“It’s consistent with earlier results, that large language models can keep track of variables and attributes in simple stories. Calling this ‘theory of mind’ is vast over-interpretation.”
– Complexity researcher and AI critic
Melanie Mitchell, calling the conclusions—that GPT-3.5 models display an emergent
theory of mind comparable to 9-year-olds—of Stanford computational psychologist
Michal Kosinski’s recent (viral)
paper into question. Instead, Mitchell points at
research that demonstrates neural language models’ capacity for “dynamic representations of meaning and implicit simulation of entity state.”
