Korva is a scroll of the internet’s best ideas.
Build for two to three months out
Too old and too early
Building for a past model's capabilities is wrong entirely, because the product ends up overly focused on what is already out of date. Building for a year out is wrong too: a product that is too early is simply wrong.
The two-to-three-month window
The window that works is two to three months ahead. That puts model capability at the center of the product and asks the builder to get out of the model's way in the product constructs created. Each design choice gets tested against one question: will it be right for the model in two to three months?
Research as the forecast
Research supplies the forecast. The trajectory isn't a black box, since there are focused efforts to make models better at particular things, such as coding in specific ways or writing in specific ways. Product development stays as tied as possible to the research agenda and roadmap, and constant, tight communication with research shows where the next few months are heading.
Why do products that arrive too early fail?
General Magic, founded in 1990 by Apple veterans, set out to build pocket communicators with email, apps, and wireless messaging. The networks, batteries, and chips of the early 1990s could not carry that vision at a price people would pay. Its Magic Cap devices shipped in 1994, and the company shut down in 2002. Tony Fadell and Andy Rubin both worked there before going on to help create the iPod and Android.
Why does hand-built scaffolding age so badly?
AI researcher Rich Sutton described the pattern in his 2019 essay The Bitter Lesson. Across seventy years of AI research, he wrote, methods that lean on more computation have beaten methods built on human-designed knowledge. Clever scaffolding gives a quick gain, then gets overtaken when the underlying model improves. Sutton later shared the Turing Award for 2024 with Andrew Barto for their work on reinforcement learning.
More from Lenny's Podcast: Product | Growth | Career
