AI Predicts the Text of Answers

This article challenges the common critique that AI models like LLMs merely perform next-token prediction without true understanding. The author argues this is analogous to human thought processes and highlights that AI predicts the text of answers to specific queries, not random strings. To prove AI can understand novel scenarios, the author created a website with original whodunit mysteries involving made-up physics rules, demonstrating that AI can solve problems not present in its training data. The article delves into philosophy of AI, comprehension, and the implications of token prediction, making it relevant to AI, machine learning, and tech industry discussions.