Bend – A language that blocks AI mistakes via proof, on CPU and GPU
Points and comments are a snapshot, not live.
Bend is a language that uses proof-checking to block AI coding errors on CPU and GPU.
Bend compiles to native code, claiming nearly C-like speed on one core and up to 124x faster on GPUs via automatic parallelism. Its type checker also acts as a proof checker, completing in under a second for a 3,200-instantiation test where Lean took 19.2s. Users declare laws in a LAWS.bend file; the compiler then verifies any AI-generated code satisfies them, mathematically blocking bugs. The language uses Python-like syntax, an affine dependent type theory (BendTT), and a parallel runtime for CPUs and GPUs. Installation is via curl, and usage instructions are given for AGENTS.md files.
What commenters are saying
The dominant reaction is skepticism about the project's maturity and trustworthiness. Many commenters pointed out that the GitHub repo was squashed to a single commit, erasing history, and that the compiler is a single TypeScript file of AI-generated code. The author explained the squash was due to private data and proprietary code but restored history after pushback. Some commenters engaged with the formal verification concept, citing EC2 and Apple examples, but others argued that specifying laws for real-world programs is impractical. A few commenters expressed genuine interest and appreciation for the underlying interaction combinators research.