GPT-5 and the Great Slowdown

The author argues that LLM progress may be entering a slower, more nuanced phase, and that founders should consider how different forms of model stagnation would change what they build. The piece cites GPT-5 failing to impress relative to its promotion, and usage limits introduced by Replit, Cursor and Claude Code that left some users finding the services less helpful. It suggests subsidized subscriber credits used to embed services in workflows are now being tuned back for utilization and profitability. It then outlines hypothetical startup "metas," such as workflow agents, domain data loops and research firms, for different stagnation scenarios.