Can AI Improve Itself? [Chris Lu, Robert Lange, Cong Lu]

Automating scientific discovery and algorithm design through Large Language Models (LLMs) leverages these systems as evolutionary operators to explore vast, non-intuitive search spaces. By treating LLMs as mutation operators, researchers can discover novel preference optimization algorithms like DiscoPOP and complex agentic architectures that outperform hand-crafted solutions. These automated frameworks, such as the "AI Scientist," chain together ideation, experimentation, and paper writing, effectively turning compute and capital into scientific insights. While these models excel at interpolating across diverse fields—mixing concepts from physics, chemistry, and economics—they rely on human-defined metrics and curation to ensure quality. As these systems scale, the focus shifts from manual algorithm design to managing the "taste" and "vibes" of automated outputs, ensuring that the resulting scientific contributions remain meaningful and interpretable within the broader academic community.