Bend 2 is a programming language designed for the AI coding era, where humans write formal specifications called “laws” and AI systems generate implementations and proofs that a compiler then verifies. While the concept sounds compelling, a detailed analysis reveals it exemplifies a broader pitfall called “vibe-coding”—using AI to rapidly prototype solutions without sufficient domain research.
According to the source material, the problem becomes apparent when examining Bend’s demo code. The specification requires 58 lines of code just to establish that a player cannot touch a flag or win a game. The AI-generated proof then requires 442 additional lines to validate these properties. This verbose approach suggests the language lacks awareness of established formal verification practices.
The article argues that the same program can be implemented far more efficiently in SPARK, an existing open-source language designed specifically for formal verification. When the same demo is recreated in SPARK—even using an LLM with minimal guidance—the implementation becomes substantially more concise. More importantly, SPARK’s built-in verification tools can prove all correctness properties automatically, requiring only 12 verification checks rather than the multi-hundred-line proof Bend demands.
The core issue is that Bend appears to have been developed without engaging with formal verification as an established academic and professional field. The words “formal verification” appear nowhere on Bend’s webpage or codebase, despite building an entire language around that concept.
This pattern illustrates a broader risk with vibe-coding: developers can create complete systems—languages and compilers included—while remaining unaware that better, more mature solutions already exist in their domain. An LLM will readily help build something from first principles without suggesting that decades of research have already solved the problem more elegantly. The author notes that prior research into formal verification would have immediately revealed SPARK and similar tools, potentially redirecting the entire project toward leveraging existing standards rather than reimplementing them verbosely.
Key facts
- Bend 2’s demo specification is 58 lines; the AI-generated proof is 442 lines
- The same program in SPARK requires far less code and automatically proves all properties with 12 checks
- Formal verification as a field does not appear in Bend’s documentation or codebase
- The author describes this as a ‘vibe-coding trap’ where rapid AI-aided development bypasses domain research
- SPARK is an existing open-source language specifically designed for formal verification