According to God of Prompt on Twitter, recent research reveals that the interpretability challenge of large language models like GPT4 stems from their complex, evolving tensor space rather than sheer model size. Each Transformer layer in GPT4 generates an L×L attention matrix, and with 96 layers and 96 heads, this results in an immense and dynamic tensor cloud. The cited paper demonstrates that the opaque nature of this tensor space is the primary barrier to understanding model decisions, highlighting a critical issue for AI researchers seeking to improve transparency and accountability in advanced models.