What's this about?
Two posts from the same person this week. On writing: he fed a long passage to a model and asked for a 1,500-word article. It spent two hours revising, kept getting it wrong, kept embellishing and adding strange things. He gave up, opened a markdown editor, pasted in what he had — and found it was already 1,200 words. He reckons writing it himself would have taken under half an hour. His conclusion: for writing, the productivity gain does not really hold. On the Cursor cutoff, more tersely: "OpenAI banning Cursor from using its models perfectly explains why we need open source models. we need open not openai."
Why does writing resist the productivity gain?
Because there is no external criterion. In code the test passes or fails, so an agent can iterate against a signal. In writing, the only judge is the author's taste, which is not in the loop — so the model optimises for something adjacent (fluency, completeness, hitting a word count) and drifts. The same reason self-improving skills work for code review and, as this week's Warp discussion noted, negatively optimise for writing.
What's the community saying?
- His broader complaint the same week is the most-liked of the three: "using AI is so painful — waiting, revising, irritation, repetition; time slips away; you can't get into flow at all. If it weren't for making a living, who would use this junk?"
- Small engagement numbers, included because they are the counterweight to a week of capability announcements, from someone who uses these tools daily and professionally.
- "We need open not openai" was the tersest reaction to the Cursor decision and the one most repeated in Chinese-language circles.