Building Real Things with AI Takes Longer Than Promised

A founder posted about spending a full year trying to build a real app with AI assistance and the thread drew significant engagement. Separately, 'Why Software Factories Fail' argued that automated code generation pipelines break down because harness engineering alone is not enough, and that the constraints come from how models are trained, not just how they are prompted. The computational chemistry SSH rant also surfaced a version of this: someone concluding that Claude makes local tasks easier, while commenters pointed out that Claude would make cluster tasks equally easier if they applied the same thinking.

The pattern here: there is a growing body of first-person evidence that AI coding tools change the shape of the work rather than eliminating it. The year-long app build is the starkest version of this. The 'software factory' framing is the enterprise version. Both point at the same thing: AI shifts where the friction is, it does not remove it.

The HN discussion on software factories was notably thoughtful, with people pointing out that model behavior is downstream of RL training choices, and that understanding that layer matters more than optimizing prompts.