Michael Malis, creator of pgrust, has been advocating that developers should build more ambitious software—a pitch backed by concrete evidence of what’s now possible.
Pgrust is a Postgres rewrite project that, according to Dan Luu’s analysis, has demonstrated substantial quality through rigorous testing. During development, the project has uncovered real bugs in Postgres itself, including a numerical bug causing valid SP-GiST operations to fail, a format cache bug producing incorrect JSONPath results, a pgcrypto bug affecting later operations, multiple memory safety issues in the dictionary loader, and a money overflow bug. While pgrust itself may not succeed as a product, Luu notes that even failed ambitious projects can leave lasting improvements—demonstrated optimizations could potentially make their way into Postgres, Clickhouse, and similar databases.
Luu acknowledges widespread skepticism toward such projects, as low-quality AI-generated rewrites are common. However, he argues that builders shouldn’t be deterred by this landscape. The real change enabling ambition is capability advancement in tools and AI systems.
Luu illustrates this with personal examples. He built a custom regex engine, a ripgrep fork, and a Rust compiler fork—projects that would have required a year or more to reach working quality in the past, but took only single-digit hours. Similarly, he developed a chess AI that achieved superhuman performance at matched time controls in approximately five hours using GPT-5.1/5.2, then refined it to compete at the strongest level in about 20 hours.
The pattern extends across domains. Luu conducted evaluations of verification and test techniques, initially with GPT-5.6 implementing zstd, then repeated them with GPT-6 Astra. The newer model performed so well—succeeding at passing all registered tests more than 99% of the time across conditions—that Luu had to find different evaluation criteria because models had advanced beyond the original benchmarks. Notably, implementing zstd cost approximately $1 with the newer model.
Luu notes that capabilities have advanced such that bottlenecks he previously encountered—regex performance, Rust compilation time, and syscall/filesystem overhead—can now be addressed. While skepticism about quality remains justified given the prevalence of poor AI-generated projects, he argues that builders applying reasonable testing and benchmarking rigor can create genuinely usable work.
Key facts
- pgrust’s testing has uncovered multiple real bugs in Postgres, including numerical, format cache, and memory safety issues
- Projects that would have taken a year to complete in the past now take single-digit hours to reach working quality
- A chess AI reached superhuman performance in approximately 5 hours; subsequent refinement took about 20 hours
- GPT-6 Astra successfully implemented zstd passing all registered tests 99% of the time across various conditions for approximately $1 in costs