Why old code is the hottest data in AI right now

Models learn to write software by studying software. Since 2023, more and more new code has been written with the help of AI assistants: in Stack Overflow's 2023 survey, 70% of developers said they were already using AI tools in their work or planning to.

That matters because of what happens when models learn from other models. Researchers writing in Nature in 2024 showed that when models are trained on text produced by earlier models, they gradually lose the rare and unusual cases, a problem they called model collapse. A later study found the cure: keep the original, human-made data in the mix.

So code written by people before 2023 is scarce, and it only gets scarcer. Epoch AI estimates that AI training could use up the public text written by people between 2026 and 2032, and the 2026 Stanford AI Index finds that AI-made data is still not replacing real data in training.

The AI data buyers we work with ask for human-written code from before 2023, usually hundreds of thousands of lines for each product, not counting third-party libraries, configuration or generated files. Some languages are scarce enough to be wanted in smaller amounts, among them COBOL, Rust, Go, Swift, MATLAB, C and C++. Its history adds to it: the commits, pull requests, review comments and tickets that show how real engineers reasoned and decided.

It does not have to be live code. A retired system, or the archive of a company that has closed, can count, as long as whoever holds the rights agrees. Software Heritage, which preserves public source code, warns that software is already being lost, and private repositories are beyond its reach.

If you lease or license it, the code stays yours. Before anything is shared, secrets and personal data come out, and you check that no customer contract or open-source licence limits what you can offer.