Korva is a scroll of the internet’s best ideas.
A five-month project in minutes
A year of experiments
The owner of a group of small businesses hired a PhD in AI and spent all of last year running experiments. Not all of them were complete failures, but things kept breaking and nothing turned out the way the team expected. The technology was not there until about November or December of that year, when everything changed.
The rebuild
When Claude Cowork came out, it brought the layman what coders had been able to do two or three months earlier. In late January, two big AI projects that had been shelved were still sitting there. One of them took five months and hundreds of hours to build the first time. With the new tool, it was recreated in about fifteen to twenty minutes, and it worked flawlessly.
Three months later
In the three months since, the group has built more software, and improved its businesses more, than in the previous ten years combined. Its decision making and its relationship-first approach stayed the same.
Why did coders get AI tools first?
Code comes with a built-in way to check the answer: a program either runs and passes its tests or it does not. That gives AI systems fast, clear feedback, and public code repositories supplied huge amounts of examples to learn from. Benchmarks such as SWE-bench, released in 2023 by researchers at Princeton and the University of Chicago, score models on whether they can resolve real software issues. Most other kinds of office work have no comparably clean pass-or-fail test.
Why is a second build so much faster?
The first build of a project includes discovery: working out what the system needs, hitting dead ends, and finding edge cases. A second build starts with all of that already known, so only the writing remains. Software engineer Fred Brooks wrote in The Mythical Man-Month in 1975 that teams should plan to throw one away, because the first version teaches what the real problem is.
More from Joys of Compounding
