The solution might be cancelling my AI subscription

Twenty projects nobody asked for

David Wilson, who writes at thoughts.hmmz.org, published a reckoning on 31 May: the list of things he has built with AI assistants is long, and almost none of it should exist. A speech recognition system in Rust, a Jellyfin desktop clone with GStreamer and Qt Quick, an Invidious clone in Python and yt-dlp, a Windows 95 notepad.exe clone in FLTK ported from the Wine sources, an OpenCV tool that counts traffic flows from public street cameras, a 3D car game built on an existing multiplayer title's protocol, and a Lightroom UI clone that never got a backend.

Only the SaaS he wrote in Rust has continuing value. The rest he cannot afford to maintain — "not in terms of time, commitment, belief, attention or willingness to spend on tokens". A regional news site he never meant to build, running on Python and Flask, now gets real traffic; he calls it "surely a liability".

A thermonuclear ADHD amplifier

The problem is not the models but what they do to attention. Wilson calls the technology "horrific for attention" and a "thermonuclear ADHD amplifier", and says he has watched it happen to every adult friend: three screens, unrelated projects, and too little commitment for the time to pay off. In a recent interview the host said the team was "quite light" on AI — everyone had up to five rooms where they manage their agents. Wilson felt a tightness in his stomach.

Friction equals focus

He tested the idea on himself. A pipeline that turned Telegram voice notes into finished Opus-formatted posts via speech recognition produced "unbridled garbage": remove the effort and you remove the commitment, and with it the focus that makes anything worth reading. Conversational English, he argues, is low-bit-rate noise; good writing carries high-bit-rate information in well-formed concepts. His conclusion: as long as quality matters, handwriting can never be obsolete.

More usage is the product

Wilson downgraded his own subscription to Pro hoping a quota would restrain him, then moved to Codex during a bad service period at Claude — the CLI is nicer and noticeably faster, he notes — and his usage crept back up anyway. The tools are built for the opposite of restraint: ask ChatGPT a yes/no question and it still appends a follow-up to keep you talking. He cites Cal Newport on "pseudo productivity" and the digital productivity paradox, in which frictionless tools multiply shallow tasks and the appearance of busyness. His remedy is blunt: he sees no way to manage AI except by curtailing use, because a tool that returns a cheap reward for minimal input can only be a liability.