Liam Powell used Bend 2 to show a common vibe-coding trap, where AI builds a whole language for a job that SPARK already solves in a few lines

Liam Powell published a critique of Bend 2 on 18 September titled "Bend 2 and the Vibe-Coding Trap." It reached 308 points on Hacker News, and its argument is that Bend's demo is an entire language reinventing a category of software, formal verification, whose current standard tool would have solved the same demo in a fraction of the code. Powell treats Bend as an example, not the target: the same trap catches any team that builds a substantial system before learning the field it sits in.

Bend 2 asks a developer to write "laws" and lets an LLM write proofs the compiler then checks. The demo on Bend's homepage states that the player can never touch the flag or win the game. Powell counted 58 lines in the LAWS.bend file that state the property, and 442 lines in the accompanying PROOF.bend file that the model produced to prove it.

The comparison

Powell vibe-coded the same demo in SPARK, an open-source language and compiler for formal verification, with no guidance beyond "recreate this demo." The resulting Ada program declares a state, a wall function, a Safe ghost function, and a Step procedure with a postcondition. Running GNATprove returns "Success: all checks proved (12 checks)." The complete file is short enough to read on one screen.

Powell's point is about the field itself. Bend's author appears to have built an entire language around formal verification without noticing that formal verification is an established field with mature tools. The phrase "formal verification" appears nowhere on Bend's site or in its codebase.

The class of failure

Powell's title names the failure: vibe coding lets a developer produce a substantial solution before learning enough to recognise a better one exists. When a person asks an LLM for a system that proves function correctness from first principles, the model will build it. The model will not stop and say the field already has tools that do this without the model, and that the person is asking for the wrong thing.

The Bend author later commented on the Hacker News thread, and Powell added a note pointing to that comment. Powell also says he does not know the author's history with language design and that his real objection is the vibe-coding pattern, not the person.

What this changes for a team

The reader-facing consequence is a change in the order of work when using AI to build a new system. One hour reading the field, listing existing solutions, and choosing which of them the new system is meant to beat comes before opening the editor. Code produced without that step still runs, so the failure is silent: the model builds a 500-line answer where 30 lines would have done, and nobody sees the shorter version because it was never asked for.

Bend 2 launched earlier this month with a claim that AI writes the implementations and the compiler checks the proofs. The launch story is a real one, and it ran on this AI News on 17 September. Powell's critique is a separate story: it is not about Bend but about the pattern that catches every team building with a model.

Source

• Bend 2 and the Vibe-Coding Trap, by Liam Powell, 18 September 2026.
• Hacker News discussion, 308 points at time of writing.
• Comparison target: SPARK Pro, AdaCore.