Bend is a new programming language designed to block AI mistakes through formal proof, while delivering fast performance on both CPU and GPU. It compiles to native code and runs nearly as fast as C on a single core, with the ability to parallelize across sixteen cores or GPUs for up to 100 times speed improvements, according to bend-lang.com.
The language integrates a proof checker similar to those in Lean and Rocq, but Bend’s type checker completes verification in under a second, enabling AI agents to validate code after every change. Users can install Bend via a simple shell script and follow guides to write code that adheres to strict laws and proofs, ensuring ambiguity-free instructions for AI implementations.
Bend’s approach addresses the challenge of ambiguity in AI-generated code by allowing developers to specify precise intents and verify that AI implementations meet those intents correctly. This is particularly relevant in the post-AGI economy where humans will rely more on AI to build software. Bend’s combination of fast compilation, parallelism, and formal verification distinguishes it from traditional languages and proof systems that often take minutes to check code.
The language supports running the same binary on CPUs and GPUs, leveraging CUDA parallelism without requiring explicit thread or kernel management. Bend’s fast compile and proof-check cycle on Apple M4 Max chips demonstrates its efficiency, making it suitable for AI-driven software development workflows that demand correctness and speed.