Hinton and LeCun argue once again, with LeCun's latest paper revealing that GPT-4 has an accuracy of only 15%, far below human performance. LeCun proposes the GAIA benchmark for general AI assistants, which evaluates reasoning, multimodal processing, and other issues, with results showing GPT-4's accuracy at only 15%. The research emphasizes that the design of GAIA is simple for humans but poses challenges for most advanced AI, with human accuracy rates reaching up to 92%. LeCun's study points out that the current autoregressive large models have not yet achieved human-level intelligence in reasoning and planning capabilities.