Korva is a scroll of the internet’s best ideas.
On the first attempt, the latest generation of Anthropic's model, Fable, can create very sophisticated things that look really good. 3D games are one example. On some tasks, more intelligence is great.
As a project evolves, it works the same as before: a person often should plan together with the agent about what the right thing to do is, and that planning is more of the bottleneck. Figuring out the right strategic directions, or the experiments that would improve outcomes for a business, is not changing as fast.
Many engineers no longer write code or look at it, so creating software no longer requires being an engineer. What moves a product and a business forward is knowing how to use the tool and plugging in all the right data to take the right decisions.
Why is deciding what to build the hard part?
Frederick Brooks drew a line in his 1986 essay No Silver Bullet between accidental difficulty, such as languages and tools, and essential difficulty, which lies in deciding what the software must do. Tools can remove the accidental part but leave the essential part untouched. Choosing requirements means understanding people, markets, and trade-offs that no code editor shows. Brooks predicted that no single tool would deliver a tenfold gain in productivity within a decade.
How are AI coding models measured?
One widely used benchmark is SWE-bench, introduced in 2023 by researchers at Princeton University. It gives a model real issues from popular open-source projects and checks whether its changes pass the projects' tests. Scores on it rose from a few percent at launch to well above half on later releases. Benchmarks like it measure the writing of working code and say little about picking which problem to solve.
Why do experiments matter more as building gets cheaper?
When a version of a product takes days instead of months, a company can try many more ideas in the same time. Each idea still needs a measurement to tell whether it helped, such as signups or retention. Controlled experiments compare a change against an unchanged control group at the same time. The number of experiments a team can run per quarter is limited by the traffic and time needed for each result.
More from All-In with Chamath, Jason, Sacks & Friedberg
