The gap in shared understanding of LLM capability is widening

AI researcher Andrej Karpathy argues that disagreements about AI capability partly reflect differences in people's tasks and access. In his October 11 post, he contrasts asking for a paragraph with assigning a large, difficult project: the smaller request gives the system less room to reveal its capabilities. He also distinguishes access to free, paid and internal systems. This is his explanation for diverging experiences, rather than a measured distribution of what users can accomplish.

His earlier April 9 post explains why paying for a stronger model alone may not produce the same experience across occupations. He argues that progress is concentrated in technical work, partly because training can reward checkable results. A code test can pass or fail; writing quality is harder to grade explicitly. That makes the task domain consequential alongside access and ambition. On his account, someone's experience with writing assistance says relatively little about another person's experience assigning a substantial programming project.