No gradients + self-written code, AI finally no longer needs to stack cards for fine-tuning to evolve on its own—this is the true paradigm shift.

Diana Hu stated on X that the future frontier is a thin software layer, enabling AI to write executable world models like programmers—continuously testing, modifying, and streamlining code based on runtime results, without expensive fine-tuning. This confirms Weng Jiayi’s gradient-free learning: large models can write code and find bugs without tuning parameters. Paul Graham believes the cycle of coding—verification—compression is similar to scientific research, and as coding abilities explode, the next paradigm of AI self-evolution is being ushered in.