Recent generative AI successes have led some to believe that Artificial General Intelligence is imminent. This optimism often focuses on multimodal approaches, where massive modular networks are optimized for multiple data types together.
However, Terry Winograd argues this strategy will fail in the near term. He states that projecting language back as the model for thought loses sight of tacit embodied understanding.
True AGI must be general across all domains, including physical tasks like repairing a car or driving. These problems require intelligence situated in a physical world model, not just symbol manipulation.
The Limitations of Next-Token Prediction
Winograd suggests Large Language Models (LLMs) learn bags of heuristics to predict tokens rather than learning a true model of the world. This creates a superficial understanding of reality.
Evidence for this includes OthelloGPT, which learned rules that do not hold for all possible games. For example, it assumed "if the token for B4 does not appear before A4 in the input string, then B4 is empty."
