The 'microworld' problem in early AI: researchers built simplified simulated environments (e.g., simulated robot warehouses) where their systems worked well, but failed to transfer solutions to real-world problems because they had abstracted away all the difficult complexity.