A developer essay examining the Bend 2 programming language has renewed debate over whether AI-assisted “vibe coding” can accelerate implementation faster than developers acquire the domain knowledge needed to choose a sound design.
Author Liam Powell presents Bend 2 as a case study rather than a neutral review. Bend 2 is promoted as a language for an era in which people define rules while AI systems generate implementations and proofs that a compiler checks. Powell argues that the project’s public demonstration requires disproportionate effort: 58 lines to express that a game player cannot touch a flag or win, followed by a 442-line generated proof.
The essay contrasts that approach with SPARK, an open-source programming language and toolset used for formal verification. Powell says he asked a language model to recreate the demonstration in SPARK and found that its established proof tooling could verify the relevant properties without constructing a long proof from first principles. The comparison is the author’s experiment and interpretation, not an independent benchmark of either system.
Powell’s broader argument is that rapid AI generation can conceal an early research failure. A developer may build a substantial language, compiler or framework before learning that mature techniques already solve much of the same problem. Because a model generally follows the requested architecture, it may readily produce a complicated implementation instead of challenging the premise or directing its user toward decades of relevant work.
The post specifically situates Bend’s design within formal verification, a field concerned with mathematically establishing that software meets stated properties. It criticises the project for not clearly connecting its approach to that established discipline. Powell also acknowledges the limits of his assessment: he does not know the Bend author’s background or whether the trade-offs were considered and rejected deliberately. He asks readers to treat some criticism as applying to a hypothetical creator of the same design.
That qualification is important because the evidence supplied is a single critical essay, not a response from Bend’s maintainers or a peer-reviewed comparison. The article does not establish that Bend 2 is unusable, nor does it measure developer productivity across representative projects. Its documented examples instead support a narrower lesson about process.
AI coding tools reduce the cost of turning an idea into working code, but they do not automatically reduce the cost of selecting the right idea. Powell’s critique suggests that research, comparison with established systems and explicit design review remain necessary before implementation. Faster generation can make those steps more important because a flawed direction can accumulate convincing code and tooling much sooner than before.
