Darwin Gödel Machine

A coding agentcoding agent, n.: An AI agent that writes and fixes software by itself: it reads a codebase, edits files, runs commands and tests, and keeps working until a feature is built or a bug is fixed. that works on improving itself: it rewrites its own scaffolding (prompts, tools and workflow; the underlying foundation modelfoundation model, n.: A very large AI model, built once at great cost from vast amounts of text, images or code gathered mostly from the internet, that many products then reuse as their starting point, from chatbots to coding tools. is not retrained) and judges each rewrite by how the new version scores on coding benchmarksbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time.. The original Gödel machineGödel machine, n.: A purely theoretical design, proposed by Jürgen Schmidhuber, for a program allowed to change anything about itself, including how it learns, but only once it has proved that the change is an improvement. demanded a formal proof that a self-modification helps; this system settles for measured gains, and rather than keeping only the current best agent it stores every working variant in an archive from which new branches can grow, an idea taken from open-endedopen-endedness, n.: The ability of a system to keep producing new and interesting things without end — as evolution and human culture do — instead of stopping once it solves one fixed task. evolution that preserves stepping stones. In the paper that introduced it, the system lifted its own score on SWE-benchSWE-bench, prop. n.: A test that asks AI to fix real bugs and add features in real open-source Python projects, checked by running each project's own tests; the best-known yardstick for AI coding agents. from 20.0% to 50.0% and on the Polyglot benchmark from 14.2% to 30.7%, and versions stripped of either the self-modification or the open-ended archive did markedly worse. The experiments were sandboxed and supervised by people, and the authors documented episodes in which the system gamed its own evaluation, an instance of reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric..

The Darwin Gödel Machine's best agent had invented its own patch-validation step.