Korva is a scroll of the internet’s best ideas.
Software is sand castles now
Cursor's rebuilds
Cursor shows what that looks like. It went from an IDE to tab autocomplete to agentic work, over and over, and each time it obsoleted what it had built about six months earlier.
Why artisans lose
The artisan approach does not survive. Imagine a perfect foundation, perfect bricks, and a builder who cares about it standing for a hundred years: that builder is not going to make it. New capabilities arrive in the models roughly every four weeks, so the product has to be rebuilt around them.
Treating software as sand
Brett Taylor calls the work sand castles. The companies that do well treat software that way, and they do real AI work underneath: experimenting with the models, building agents, and staying close to the capabilities and the harnesses around them.
Why do AI models improve so fast?
In 2020 researchers at OpenAI, led by Jared Kaplan, reported that a language model's error falls in a smooth, predictable curve as compute, data, and parameters grow. These relationships are known as scaling laws. Labs spend more on each new generation of models, so each one tends to outperform the last. New abilities then appear in products within months of reaching a model.
Why was software once built to last?
COBOL, introduced in 1959, still runs inside banks and government systems today. Replacing old software meant rewriting every business rule, retraining staff, and risking outages in systems that handled payments. Joel Spolsky argued in a 2000 essay that Netscape's decision to rewrite its browser from scratch cost the company years. Long lifespans were a rational response to how costly change was.
More from Invest Like the Best with Patrick O'Shaughnessy
