Bend – A language that blocks AI mistakes via proof, on CPU and GPU

If you use AI to write code, it will now block mistakes that would have previously shipped.

Bend, a new programming language, has been released with the goal of preventing AI-generated code mistakes through formal proofs. Unlike natural language prompts, Bend uses "laws" to specify intents precisely, which are then verified by proofs to ensure AI implementations are correct. The language is designed for speed, compiling to native code that runs comparably to C on a single core and significantly faster on multi-core CPUs and GPUs, up to a hundred times quicker. Bend's compiler is also fast, completing type-checking, which also functions as proof-checking, in under a second. This rapid verification allows AI agents to check code after every modification. The language supports automatic parallelization, eliminating the need for manual thread management or lock handling. Key to Bend's functionality is its "LAWS.bend" feature, where developers define rules that AI cannot violate, with violations mathematically proven as impossible before code is merged. This system aims to prevent bugs from entering production by enforcing strict logical constraints on AI-generated code.

Bend's introduction signals a pivotal shift in how AI-generated code will be validated and integrated into development workflows. By embedding formal verification directly into the language and compiler, Bend addresses the critical trust deficit in AI-assisted programming, moving beyond syntax checking to semantic correctness enforced by mathematical proof. This could significantly accelerate AI adoption in safety-critical software development and complex systems where bugs have high costs, provided the language gains traction and its proof system proves robust and user-friendly for AI agents.

The focus on both speed and provability suggests a future where AI agents can rapidly iterate on code, with the assurance that fundamental logic errors are prevented by the language itself, not just by human review or testing. This has implications for development cycles, potentially reducing debugging time and increasing the reliability of software produced by AI. Developers will need to adapt to defining these "laws," effectively building a declarative contract with the AI that guarantees certain properties of the resulting code, which may redefine roles in software engineering.