1. Trust and transparency (commit history)
Many commenters reacted strongly to the repository being squashed to a single commit, seeing it as a red flag for reproducibility and openness.
- AlexErrant: “Erm, because it looks weird? Do you know any other language projects that squash their repos down to 1 commit? That's a destruction of trust, which is kinda important if you want people to build on your language.”
- LightMachine admitted the move was to hide personal data and “AI slop,” later restoring the history after feedback.
- NortySpock noted that a single commit makes it impossible to gauge the project’s age, update frequency, or community size.
2. AI involvement and hype (“vibe‑coding”)
The discussion repeatedly questioned how much of Bend was produced by large language models versus hand‑crafted work, and whether the project’s claims were overstated.
- boxed called the compiler “a single typescript file 6k characters long of AI slop. Jesus.”
- AlexErrant warned that “nuking your history is one hell of a way to raise eyebrows” in an era where AI trust matters.
- LightMachine conceded: “Yes, there's a lot of vibe‑coding in many places, but the critical parts (compiler, runtime, kernel) are human designed.”
- stschaef remarked, “I'm glad you're having fun vibecoding…”, highlighting the skepticism around AI‑generated documentation and papers.
3. Technical claims (GPU parallelism, proof checking, performance)
Debates centered on whether Bend truly delivers on its promises of GPU‑accelerated proof checking and competitive speed.
- LightMachine explained: “we implemented a complete allocator, garbage‑collector, closure evaluator and functional evaluator, on the GPU (with zero interaction net overhead this time).”
- stschaef questioned: “How does this benefit from GPU parallelism? … these tasks aren't amenable to running on a GPU.”
- LightMachine replied that GPU proof checking would come later “when we implement Bend in itself.”
- ModernMech stressed that destroying history makes it hard to reproduce the benchmark SHAs cited in the paper.
4. Practical usability of “laws” and specification burden
Several users illustrated how underspecified laws can lead the AI to find loopholes, questioning the approach’s scalability.
- pdpi observed: “having only the one single 'you can't win' law is severely underspecified… writing the laws becomes a bigger problem than writing the code itself.”
- holden_nelson showed the AI making the flag “unenterable” by adding a force field after being told to keep walls away.
- LightMachine agreed: “'you can't win' is grossly under‑specified… they only protect what you remember to write.”
- The exchange highlighted the tension between lightweight specifications and the need for precise, maintainable laws.