Bend pitches AI-era coding

Bend 2 is a new language where humans write specifications called laws, AI writes implementations and proofs, and the compiler verifies correctness. The pitch sounds compelling but masks deeper design problems.

The demo requires massive code

Bend's homepage demo needs 58 lines just to state that a player cannot touch a flag or win. The proof to verify this takes 442 additional lines of code.

Vibe-coding skips research

Building substantial solutions with AI before understanding the problem domain risks missing better approaches that existing research already provides. Bend's creator seemingly built an entire language without discovering formal verification exists.

Formal verification field exists

The words formal verification appear nowhere on Bend's website or codebase, despite being the established field for proving program correctness. This omission suggests the creator never researched the domain.

SPARK does it better

The same game demo written in SPARK, an established formal verification language, requires only a fraction of Bend's code and proves all correctness checks automatically with GNATprove.

SPARK eliminates verbose proofs

SPARK's approach lets developers write specifications and loop invariants that the compiler automatically verifies, eliminating the need for LLMs to generate hundreds of lines of manual proof code.

Bend reinvented existing work

Bend's verbose specification and proof system recreates problems that formal verification research solved decades ago. A little domain research before building would have revealed this.

AI won't suggest alternatives

When asked to build a language for formal proofs, an LLM will happily generate verbose specifications and proofs without ever suggesting that computers can already build complex proofs automatically.

“it will never stop to suggest to you that computers can already build complex proofs without the need for a LLM”

The broader vibe-coding risk

AI-assisted development makes it too easy to implement designs that are broken or decades behind current practice because you get immediate results without doing research first.

LLMs miss domain context

An LLM will never tell you that what you're building already mostly exists as established work you could build on. It generates solutions without understanding the field's history or best practices.

Research before building matters

Vibe-coding works best when paired with domain knowledge. Without understanding existing solutions, AI can help you build something impressive that solves yesterday's problems inefficiently.