Prof. Matthew Schwartz describes a new approach to AI-accelerated science in this Hacker News guest post. The article introduces "Vibe Physics" and the concept of "Claude-shaped" problems—tasks best suited to current LLM capabilities rather than forcing human scientist workflows. Schwartz details building Boot Loops, an open-source toolkit for exact calculations in quantitative science.
Because similar calculations appear across disparate fields like ecology and population genetics, Claude discovered unexpected connections. While technically correct, these were often scientifically unremarkable initially. Schwartz collaborated with domain experts to steer Boot Loops toward questions specific to each field.
The piece argues that academic scientists often feel a disconnect with AI models, which excel at well-scoped mathematical problems but struggle with the messy nature of most science. The core conflict is an "impedance mismatch" between human and AI capabilities. By providing a harness like Boot Loops, AI can handle agentic-shaped problems effectively.
The toolkit is model-agnostic and open-source, allowing use with various LLMs. Current AI tools excel at certain problem types but cannot solve most scientific challenges. However, when guided by expert input, they can make substantive advances across geology, biology, economics, and linguistics, resolving the mismatch between human intent and AI capability.
Source: Hacker News · Summarized by HeadlinesBriefing