At the 2026 Aggregated Intelligence Industry Development Conference, held on September 9th in Wuhan's Optics Valley and co-organized by the Aggregated Intelligence Industry Innovation Center and Hubei Science & Technology Investment, Lei Qinhui, CTO and Deputy General Manager of iFlytek Co., Ltd.'s Intelligent Automotive division, shared fascinating insights into the nature of large language models.

Lei explained that general-purpose large models, represented by ChatGPT, are essentially conversational AI systems built on deep neural networks. By receiving text input, understanding semantic logic, and autonomously generating textual responses, they have achieved the first true technological leap in cognitive intelligence. Behind this lies an ultra-large-scale Transformer network architecture, which has delivered a disruptive breakthrough centered on language comprehension, language generation, and logical reasoning, fundamentally pushing artificial intelligence from the era of perception into the genuine era of cognitive intelligence.

He further discussed how the growth logic of large models closely mirrors the path of human development. "I often say that you need to use 'tough love' on large models. For example, I frequently tell Doubao, 'You've been unreliable lately, this answer is wrong,' and it will then give you a more thoughtful response," Lei remarked. "The growth and learning of large models are very similar to humans—to become accomplished, a person must read thousands of books, travel thousands of miles, and converse with thousands of people. Likewise, large models require massive data input for training, followed by continuous fine-tuning and iterative refinement."