Korva is a scroll of the internet’s best ideas.
Cheap models make frontier AI valuable
Cheap models, strong director
Picture a frontier model at 250 IQ directing open-source models at 150 IQ. The stronger one can farm out tasks, so each cheap model extends what the expensive one can get done.
The Manhattan Project had the same structure. Robert Oppenheimer led it with perhaps 30 physicists on par with Nobel laureates, and the project still needed some 2,000 more of the top 10,000 physicists in the world.
The price gap
The price gap is already wide. An open model called GLM costs about 90 percent less than Claude Opus, and founders and large companies are starting to use it. Corporate America made the same move with open source over the last 20 years, wary of being rug-pulled or held hostage by a large technology company.
Cheaper intelligence raises the value of the frontier. One plausible outcome has frontier tokens at 65 to 85 percent of the economic value while open-source tokens make up about 80 percent of the volume.
Why do cheap models pair well with strong ones?
Researchers have tested this pairing directly. A 2023 Stanford paper called FrugalGPT described a cascade that sends each query to cheaper models first and escalates to a costly one only when the answer looks unreliable. The authors reported matching a leading model's accuracy on some tasks with cost cuts of up to 98 percent. Modern agent systems use the same pattern, with a large model planning and smaller models carrying out the steps.
How did Los Alamos organize its scientists?
Los Alamos Laboratory was split into divisions for theoretical physics, experimental physics, chemistry and metallurgy, and ordnance, each with its own leader reporting to the laboratory director. Weekly colloquia let the groups trade results across divisions. At its peak the Manhattan Project employed roughly 130,000 people across sites including Oak Ridge and Hanford, according to the US Department of Energy. Los Alamos itself grew from a small group in 1943 to several thousand people by 1945.
Why do big companies adopt open source software?
Linux shows the pattern. The TOP500 project, which ranks the world's fastest supercomputers, has listed only Linux systems since November 2017. Companies adopt open software partly because they can inspect the code, modify it, and move to another vendor without rewriting everything. Android, built on the Linux kernel, carried the same code to billions of phones.
More from All-In with Chamath, Jason, Sacks & Friedberg
