Hardly Promethean

The piece challenges the claim that LLMs have erased software development bottlenecks, arguing that generating code is only one step in a longer pipeline that still includes shipping, maintenance, coordination, and preventing regressions. It draws on critiques from Dan Luu (agents overfit benchmarks and can’t replace rigorous experimental design) and Varun Gandhi (the LLM cost argument mainly helps expert, self-directed projects) to show that DHH’s native-frontends-plus-Rust pitch is a best‑case example, not a general rule. Concrete specifics include Hey Next being a week‑old, nonproduction experiment; Basecamp 5’s “Swiss cheese” architecture as an artifact of coordination costs; and Bun’s LLM‑assisted Zig fork compiling faster but being non‑deterministic and therefore unmergeable upstream. Removing the coding bottleneck merely exposes the next constraint.

The author also highlights a cultural shift in tolerance: handing tasks to agents lets creators accept verbose, brittle, or opaque outputs they’d never allow in shared infrastructure like Rails. DHH’s projects - a calculator, video editor, presentation tool, Linux distro, and a rewrite of his own product - are offered as evidence of scarcity of ideas rather than transformative vision; LLMs excel at reproducing existing artifacts, not sparking novelty. The central claim is that design, debugging, and idea generation remain scarce and valuable, so the era of handwritten, carefully shaped software is not obviated.