Reddit essay argues one-shot AI coding demos are huge token waste with poor decisions

rsvallen · reddit · 2026-09-07

The author criticizes the viral "AI built Minecraft/CoD in one prompt" genre on YouTube/Twitter, calling it one of the highest capability and token wastes possible: agents just loop generate-and-fix under a prompt, producing demos whose underlying decisions are poor.
Core argument: in a general model's training data, high-quality game-dev code is vanishingly rare (est. 0.005-0.05%); the more niche the domain, the thinner the coverage. GitHub's Octoverse notes 81.5% of 2025 contributions happened in private repos, so the best real-world code is largely inaccessible.
He frames AI as a multiplier: each suboptimal decision biases subsequent output, diluting quality multiplicatively (1×0.5×0.5).
Proposed alternative: build reusable core working skills first, then layer domain skills in context, so agents can add, subtract, and compound improvements—producing work valuable to humans instead of bigger demos.

More from coding & agent