Jürgen Schmidhuber Publishes a Four-Decade Case for Recursive Self-Improvement

Jürgen Schmidhuber’s new technical note makes a sweeping case: the core ideas behind recursive self-improvement have been under development for decades, and software-based versions are now practical. But the note also places a hard boundary on that promise, arguing that full recursive self-improvement will require machines able to improve their own hardware in the physical world.

Published September 17 as “Recursive Self-Improvement (RSI) Since 1987,” the note is a survey of Schmidhuber’s own work rather than a report of a new working system. It traces a sequence of methods that he presents as attempts to make a learning system improve the method by which it learns.

The idea is to improve the improver

Schmidhuber defines true meta-learning, or recursive self-improvement, more narrowly than carrying knowledge from one dataset to another or tuning a few settings. In his account, the learning algorithm itself must be represented in a universal programming language and allowed to alter its own code, while a recursive framework filters for modifications that are useful.

The retrospective begins with what Schmidhuber calls Meta Evolution: applying genetic programming to evolve better genetic-programming methods. He says his 1987 diploma thesis used multiple meta-levels, with each level capable of searching for improvements to the level below it.

Different mechanisms, one recursive ambition

The methods in the note do not rely on one mechanism. Schmidhuber describes self-modifying policies for lifelong reinforcement learning, where the policy includes the means by which it may modify itself. He also describes neural systems whose outputs can set the weights of other neural networks, and recurrent networks designed to inspect and modify their own weight-change procedures.

Another branch is the Optimal Ordered Problem Solver, introduced in 2002. Schmidhuber says it addresses a sequence of problems by testing programs that can invoke, adapt, or reuse earlier solutions. The aim is not only to solve the next task, but also to search for better search procedures as experience accumulates.

A proof before a rewrite

The note’s most formal proposal is the Gödel Machine. Schmidhuber describes it as a reinforcement-learning system that can rewrite any portion of its code only after its embedded proof search finds a proof that the change will improve its utility. He characterizes the approach as mathematically optimal under the conditions laid out in the system’s initial axioms and utility function.

That framing highlights the distance between a formal model of self-improvement and an unconstrained machine that continually makes itself better in the world. The note attributes greater software practicality to falling computing costs, yet argues that the full version of the project cannot stop at code. Hardware, in this telling, must become subject to improvement too.