Korva is a scroll of the internet’s best ideas.
Jason Calacanis set one up under the name Improve My Bots Bot. It went through his bots, picked out two or three that overlapped, and asked whether combining them into one instruction set would serve him better. Then it reported every change it made.
By a rough estimate the bots came out 30 to 40 percent better on average. The result was a multi-level civilization of bots all making one another better.
The same move works on a single assistant. Asking an agent what else it should be used for brings back suggestions, and asking what to do next while building a product brought back more features.
How can an AI improve its own instructions?
A 2022 paper by researchers at the University of Toronto, Large Language Models Are Human-Level Prompt Engineers, had a model write many candidate instructions, test each on examples, and keep the best. The instructions it found matched or beat human-written ones on most of the tasks tested. Later systems repeat this loop automatically, scoring each rewrite against a set of test cases.
Why does merging overlapping instructions help?
Overlapping instructions can contradict each other or repeat, and every extra line competes for the model's attention. A 2023 study by Nelson Liu and colleagues, Lost in the Middle, found that models use information at the start and end of a long prompt more reliably than information in the middle. The study tested multi-document question answering and key-value retrieval tasks.
How do teams check a prompt got better?
Teams run evals: a fixed set of example inputs, each with a known good output or a scoring rule. They run the old and new instructions on the whole set and compare the scores. The approach also underlies public benchmarks such as MMLU and SWE-bench, which apply it at larger scale.
More from All-In with Chamath, Jason, Sacks & Friedberg
