AI architectures evolve like biology, but with more good mutations

Researchers found that the way AI models change and improve follows the same statistical patterns as biological evolution. This means that even with human guidance, AI development isn't a completely unique process; it's just faster at finding good changes.

People thought AI development was a purely engineering problem, with humans in full control of every design choice. This paper shows that AI architectures evolve under fundamental statistical laws, much like living organisms, even when humans are guiding the process. It means that some aspects of AI's future development might be more predictable through an evolutionary lens than through traditional engineering roadmaps.