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

Bend is presented as a fast language that blocks AI mistakes via proof, combining C-level speed, CUDA parallelism, Lean-style proofs, and Python-like syntax. Its premise is that in a post-AGI economy humans may eventually stop writing and reading code, but still need an ambiguity-free way to tell AIs what to build. Laws are meant to be more precise than natural language, proofs verify that the AI implemented prompts correctly, and a fast compiler runs the result at speed.

The language emphasizes speed in two ways. It compiles to native code and, on one core, runs nearly as fast as C; the same binary can also run on sixteen cores or on the GPU, where it can run up to 100 times faster than one core. Its type checker is also a proof checker, like those in Lean and Rocq, but whereas those can take minutes on a mid-sized codebase, Bend takes at most a second, allowing an AI agent to check after every change. The page cites Apple M4 Max benchmarks with lower-is-better charts.

Bend is parallel by design: there are no threads, locks, or kernels to write. Split the work in two, and Bend spreads the calls over every core it can find, then joins them back. The page shows pow2 running on 4,096 GPU cores.

The core safety claim is that you can trust code you never read by demanding a proof. Developers declare laws in LAWS.bend, and after that no AI can ship a line that breaks them. The page walks through a game-guarding example: a law states that winning is impossible. When a new feature is requested—'Claude, make the board wrap around'—without LAWS.bend the laws are broken and the AI mistake is merged; with LAWS.bend the laws stay intact and the mistake is blocked. The AI had to retry until it built a wall and proved the law holds. Merging a bug becomes mathematically impossible because it is a theorem. The example law defines you_cant_win over any List<Move>: replay moves from start and assert is_won(board) == False; the PROOF.bend file contains a Laws.you_cant_win proof 'written by the AI.' LAWS.bend is described as AGENTS.md backed by proof, so 'make no mistakes' becomes type-checked. The page invites skeptics to try breaking the game.

Getting started: install Bend with curl -fsSL https://bend-lang.com/install.sh | sh. Then tell the agent to use Bend by adding instructions to AGENTS.md: run bend guide to learn it, use LAWS.bend to keep important rules, run bend PROOF.bend before committing, and parallelize code whenever possible. After that, just say 'use Bend' and, per the page, enjoy bug-free, fast vibe-coded apps; a hint suggests asking the agent to write laws for whatever should never break.