Korva is a scroll of the internet’s best ideas.
Four and a half years, no product
Simulating customers
Without customer signal, the team had to simulate users. During the Dota 2 days, when they were using reinforcement learning to beat people at a video game, they put up a leaderboard. People could see how different ideas were performing, the ranking was objective and real, and people wanted to climb it.
Learning from old labs
Starting a research lab was a vanity project for everyone in Silicon Valley then, and books about the heyday of Bell Labs and Xerox PARC were popular. Few people had a living memory of how to run one, so the team talked to Alan Kay and a handful of others. Some of the advice was really good, and some of it did not fit the current moment.
What worked and what didn't
Leaderboards worked. So did external demos for an eminent person the researchers wanted to impress. Fake deadlines did not work. Bell Labs had a giant profit machine behind it, and OpenAI had nothing like it: the dominant memory of the early days is trying and failing to raise money.
How did Bell Labs fund long-term research?
Bell Labs belonged to AT&T, which held a regulated monopoly on American telephone service and had steady revenue. That income let the lab fund work with no promise of a near-term product. Its researchers invented the transistor in 1947. John Bardeen, Walter Brattain, and William Shockley shared the 1956 Nobel Prize in Physics for it.
How did Xerox PARC's ideas reach everyday computers?
Xerox opened its Palo Alto Research Center in 1970. By 1973 its researchers had built the Alto, a computer with a mouse, a graphical display, and Ethernet networking. Steve Jobs toured PARC in 1979 and saw its graphical interface in action. Ideas from that visit shaped Apple's Lisa and Macintosh, which brought windows and a mouse to a wide audience.
Why do leaderboards motivate researchers?
A leaderboard gives everyone one shared, objective measure, so progress is visible and ideas can be compared fairly. The ImageNet challenge, which ran each year from 2010, worked this way for image recognition. According to its organizers, the best top-five error rate fell from roughly 28 percent in 2010 to about 2.3 percent in 2017. Public rankings also invite researchers to build on each other's results.
More from David Senra
