The best AI products are sand castles, not stone castles
Models pick up new skills every few weeks, so the best teams expect to throw out their own work.
Invest Like the Best with Patrick O'Shaughnessy · Everyone Is Still Undersizing the AI Market | Eric VishriaEric Vishria, a partner at the venture firm Benchmark, borrows a line he credits to software executive Brett Taylor to describe how to build in artificial intelligence.
"We used to be building castles. Now we're building sand castles that are going to get washed away," Vishria says.
By that, Vishria means that the old virtue in software was durability. Teams laid a careful foundation, brick by perfect brick, and expected it to stand for years. In AI, he argues, that kind of craftsmanship will sink you, because models gain new capabilities about every four weeks.
His exhibit is Cursor, a coding tool that helps software developers write programs. Vishria says Cursor went from offering a full editing setup, to finishing lines of code with tab autocomplete, to doing longer stretches of work on its own. Again and again, he says, the company scrapped what it had built six months earlier and started over, and embraced it.
Staying in the game, Vishria argues, takes more than lightly adding AI on top. AI is powerful in some spots and surprisingly weak in others, a pattern he calls the jagged edge. Winners, in his view, are the teams that study that edge closely, see where it leaves gaps or low spots for their customers, and fill them in.
That inverts the old rules of product work, Vishria says. It used to be standard advice that a product manager, the person who decides what to build, should understand customers and leave implementation details to engineers. He calls that approach terrible now. Builders have to understand what models do well, where they fail, and what customers need, and then bridge the two. He says Cursor's founders did that unusually well from the start, and kept redoing it as the edge moved.
That creates an irony: knowing the technology matters more even as models are supposed to make that knowledge less important. Vishria says job titles matter less than three traits, which are an understanding of customer problems, taste, and curiosity about that jagged edge. If you have those, he says, your formal role does not matter.
Questions people ask
What does Vishria mean by the jagged edge?
It means AI is not uniformly smart. It can write a difficult piece of code and then fail at something simple, and the boundary between the two moves with each new model. Builders have to test constantly to find where the model is strong, where it breaks, and what customers will therefore pay for.
What exactly is Cursor?
Cursor makes an AI assistant for computer programmers. It started as a modified code editor, added predictions that complete the next bit of code when you press tab, and later added agents that can carry out multi-step tasks like editing files and fixing errors.
Why do AI models change so fast?
Lab researchers regularly release larger or better-trained models and new training methods. Each release can unlock behaviors the previous version could not do reliably, so a product built around last month's limits can look dated within weeks.