Korva is a scroll of the internet’s best ideas.
New capabilities arrive in the models every four weeks, so work a team finished last season can be obsolete before it pays off. An artisan who lays a perfect brick foundation expecting it to last a century won't make it. The teams that do well treat what they built as temporary and keep rebuilding as the models change.
Cursor shows the pattern. It went from an IDE to tab autocomplete to agentic work, and each step retired the work from six months before. Its team understood the uneven capability of the models from the beginning and built a product that translated what developers need into it, then kept iterating as the edge moved.
Why do new model capabilities arrive so often?
Frontier labs run pretraining, fine-tuning, and evaluation pipelines in parallel, so improvements ship as soon as they clear testing. Competition between labs adds pressure to release. OpenAI released GPT-3.5 in late 2022 and GPT-4 in March 2023, only months apart.
How did Cursor start from an existing editor?
Cursor, made by the company Anysphere, began as a fork of Visual Studio Code, Microsoft's open-source code editor. Starting there gave it an established editor, extension system, and user habits, so the team could spend its effort on AI features. Its early AI features completed a developer's typing, and later ones edited code across files on their own. Anysphere was founded in 2022.
Why does hand-built work lose to scale?
Rich Sutton's 2019 essay The Bitter Lesson looked back at seventy years of AI research. It found that general methods that use more computation eventually beat approaches built on hand-crafted human knowledge. Hand-built work looks better at first, then gets overtaken as models and compute grow. Chess and speech recognition are among the examples Sutton cited.
More from Invest Like the Best with Patrick O'Shaughnessy
