Koby Lewis advocates for improved type hints in AI safety

Koby Lewis argues that AI safety tooling needs better type hints. Good type hints make code more maintainable, easier to understand, and have fewer bugs. The article points out that many popular AI safety libraries—like TransformerLens, MACHIAVELLI, and HuggingFace datasets—use vague type hints that can confuse beginners and lead to errors. Lewis addresses objections: AI code generation doesn't eliminate the need for type hints (they can "tame hallucinations" and improve results), and even research code benefits from maintainable types when reused. Inspect is highlighted as a library that does this well.