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

Bend is a novel language designed to prevent AI errors through the use of proofs, targeting both CPU and GPU environments. In an era following the advent of Artificial General Intelligence (AGI), humans may no longer need to write or read code, but the need for a precise and unambiguous method to instruct AI systems remains crucial.

Laws can provide a high level of specificity in human intent, while proofs allow for verification that AI systems have correctly implemented those prompts. Bend compiles to native code and runs at speeds comparable to C, with the added benefit of being able to utilize multiple cores, including up to 16 or a GPU, achieving speeds up to a hundred times faster than a single core.

Bend's type checker functions as a proof checker, similar to systems like Lean and Rocq. Although these can take minutes to verify a mid-sized codebase, Bend manages to do so in just a second or less, allowing AI agents to check after every change. Unlike traditional programming, Bend does not require threads, locks, or kernel writing.

Instead, it splits the workload and distributes calls to available cores, then reassembles the results. This streamlined approach allows Bend to verify a code snippet like pow2 on 4,096 GPU cores, ensuring that no AI-generated code can violate declared laws.

LAWS.bend is a central component of Bend, where users declare the laws that must never be broken. Once these laws are established, no AI can execute a single line that contradicts them. This mechanism was demonstrated in a game scenario where a bug that went live was subsequently caught and rectified by Bend's LAWS.bend feature.

Attempting to merge a bug into the system is mathematically impossible, as it would require proving the law's adherence, which Bend makes virtually impossible to breach. The concept behind Bend is essentially an extension of LAWS.bend, backed by proof, ensuring "Make no mistakes" is type-checked at every step of the AI generation process.

For those who encounter issues or wish to contribute improvements, Bend provides guidance through GUIDE.md, which contains the entire language. BendTT is the core of Bend, an affine dependent type theory akin to Lean and Rocq. BendRT serves as a parallel runtime for both CPUs and GPUs, forming the virtual machine (VM) that executes Bend code. While Bend is still in its early stages and may contain bugs, users are encouraged to report them to help refine the language further.