25 days ago
- The author previously believed AI would not benefit from new programming languages, as AI can generate code in many existing languages.
- However, AI often generates good local code (e.g., functions) but struggles with global program understanding, leading to unnecessary defensive checks that can cause exponential state complexity.
- If AI's global reasoning weakness persists, programming language design may help by enforcing global properties through local reasoning, similar to Rust's data race prevention.
- Rust's ownership types and Send/Sync traits prevent data races statically, allowing local reasoning to ensure global data-race freedom without a new sublanguage.
- This demonstrates that programming languages can enforce surprising global properties (like data race freedom) with minimal burden on programmers.
- AI-generated code often includes excessive defensive checks, which are unnecessary, hurt performance, and mislead about program states, causing exponential state explosions.
- Future programming languages might incorporate features like effects systems to enforce more global properties, potentially benefiting AI code generation and human review.
- While not all global properties can be enforced, there is evidence that usable languages can provide more guarantees, changing incentives for language design in the AI era.