Bend – A language that blocks AI mistakes via proof and runs on GPUs

AI-generated code will now be demonstrably correct if you verify its laws.

Bend, a new programming language, has been released with the goal of preventing AI-generated code mistakes through formal proofs and high performance on GPUs. Designed for an era where AIs are expected to write most code, Bend uses 'laws' written in a special syntax to declare invariants that must always hold true. If an AI attempts to write code that violates these laws, the compilation process, which includes a proof checker, will block the change. Bend compiles to native code, offering speeds comparable to C on a single core and achieving up to a hundred times faster performance on GPUs by automatically parallelizing computations. The language also boasts fast compilation times, aiming to allow AI agents to verify code changes rapidly. Developers can install Bend using a curl script and integrate it into their AI agent configurations by adding instructions to their AGENTS.md file, specifying its use for writing 'laws' and parallelizing code.

Bend introduces a novel approach to AI code generation by embedding formal verification directly into the development workflow. By requiring AI agents to satisfy declared 'laws' through mathematical proofs before code can be committed, Bend aims to fundamentally shift the paradigm of trust in AI-generated software. This could significantly reduce debugging time and the propagation of errors in complex systems, especially as AI takes on more critical development tasks. The language's performance characteristics, including its efficient GPU utilization and fast compilation, are designed to make this rigorous verification process practical for rapid AI iteration cycles.

The implications of Bend extend to how AI agents themselves will be developed and integrated. By providing a language that guarantees correctness through proof, Bend enables developers to delegate more complex coding tasks to AIs with a higher degree of confidence. This could accelerate the adoption of AI in software development, particularly for safety-critical applications where traditional testing methods may prove insufficient. Companies that heavily rely on AI for code generation will need to assess how Bend aligns with their existing toolchains and development methodologies, potentially leading to a new standard for verifiable AI-assisted programming.