Schmidhuber traces four decades of recursive self-improvement

While frontier AI labs treat recursive self-improvement as a newly unlocked frontier post-LLMs, Schmidhuber delivers a timely reminder that the mathematical foundations of self-modifying architectures were mapped out decades ago, waiting primarily for compute and expressive representations to catch up.

• –**Validation of Early Theory by Modern Scale:** Algorithms like self-modifying policies and meta-evolution were computationally prohibitive when developed in the late 1980s and 1990s; modern LLM agents and massive distributed compute finally provide the execution layer needed to realize these recursive loops in practice.
• –**Resurgence of the Gödel Machine:** Modern agent architectures—such as Sakana AI's Darwin-Gödel and Huxley-Gödel machines—are directly reviving Schmidhuber’s 2003 blueprint of proof-search and self-rewriting code, shifting from theoretical proofs to empirical code generation and verification.
• –**In-Context Learning Reframed:** Framing LLM in-context learning and test-time training through the lens of fast-weight programmers (1991) demystifies dynamic adaptation as a continuous meta-learning process rather than an emergent black-box anomaly.
• –**The Hardware Reality Check:** The assertion that software self-improvement alone cannot trigger an unconstrained intelligence explosion without physical self-replicating robotics grounds the current ASI hype in material and manufacturing realities.