An engineer's case for canceling his AI subscription

A developer catalogs roughly twenty projects he built with Claude and Codex — speech recognition in Rust, a Jellyfin clone, a Windows 95 Notepad port, a regional news site that accidentally attracted real traffic, and a sizeable SaaS — and concludes that almost none of it is useful or maintainable. Sessions that began as ‘write a quick script for X’ routinely metastasized into hour-long builds that never solved the original problem. He describes the tooling as a ‘thermonuclear ADHD amplifier’ and reports seeing the same scattered, low-commitment output across his adult friends and in job interviews where managing five parallel agent sessions is framed as restrained usage.

His core argument is that vendor incentives push the wrong direction: every product is tuned for more tokens, more output, more follow-up questions, while judicious application requires focus the tooling actively erodes. An experiment piping voice notes through Claude into blog posts produced garbage because removing effort also removed the commitment that gives writing its signal. He ties this to Cal Newport’s ‘pseudo-productivity’ — visible busyness mistaken for value — and argues that frictionless cheap rewards are a liability regardless of how capable the underlying model becomes.

The piece resonates as a counterweight to industry boosterism: a heavy user who acknowledges the technology is genuinely impressive at honed tasks, but who can find no management strategy short of curtailing use. The implication for engineering culture is uncomfortable — if commercial environments are simultaneously generating 10,000-line untested codebases in five minutes, the aggregate maintenance burden and attention cost may swamp the productivity gains.