When should a machine rewrite itself?
In 1931, mathematician and logician Kurt Gödel encoded formulas and proofs as numbers in his work on incompleteness. Arithmetic could then express statements about proofs themselves. This idea of self-reference later inspired a question in computer science. Could a program reason about and improve its own code?
In 2003, computer scientist Jürgen Schmidhuber proposed the Gödel machine. This theoretical machine carries out tasks while searching for changes to its code. Under its stated goals and axioms, it must prove that making a particular change now has greater expected utility than continuing to run the current program and search for alternatives. It applies the change once it finds that proof.
The proof-search procedure is itself code that the machine can rewrite. This illustrates recursive self-improvement (RSI). One improvement can change how the system finds the next. The proof depends on its axioms, assumptions about the environment and utility function. A useful change may remain unproved.
Gödel Organization (GO) takes its name from that ambition. An organization can learn from the work it does and eventually improve how it learns. GO Benchmark traces that progression through evidence of work delivered by AI, lasting changes to workflows and improvements to the methods used to test those changes.
Schmidhuber’s Gödel machine paper ↗ Kurt Gödel at the IAS ↗