With a vision for a future where AI writes code, a new programming language called "Bend" has been developed with the goal of preventing AI errors through "Proofs."
New Programming Language "Bend" Prevents AI Errors via "Proofs," Enabling High-Speed Execution and Parallel Processing
This article is a translation. Read the Japanese original
Bend features "Laws" to accurately express code intent and "Proofs" to verify whether an AI has correctly implemented a prompt. For example, if a rule stating "winning is impossible" is defined in a game, it becomes mathematically impossible for an AI to generate new code that introduces changes violating that law. This makes it possible to improve the reliability of coding using AI agents.
It also possesses high execution performance. Through compilation to native code, it achieves speeds close to C on a single core, and by utilizing 16 cores or GPUs, it can execute at speeds nearly 100 times faster than a single core.
Furthermore, compilation speed is extremely rapid. Since the type checker also serves as the proof checker, even a medium-sized codebase containing 12,800 definitions can be compiled in under one second. Compared to existing languages like Lean or Rocq, which can take anywhere from several seconds to several minutes, this means AI agents can rapidly check source code changes.
In addition, Bend includes the capability to efficiently execute parallel processing by distributing tasks across multiple cores without the need to explicitly write threads or locks.