Infinities, impossibilities, and the man in the white linen suit
17 hours ago
- #mathematical logic
- #artificial intelligence
- #computability theory
- Kurt Gödel, a foundational logician, died from self-starvation due to paranoia about poisoning, highlighting a tragic personal struggle.
- Gödel's incompleteness theorems proved that any mathematical system powerful enough for arithmetic contains true statements it cannot prove, undermining Hilbert's program for a complete and consistent foundation of mathematics.
- Alan Turing built on Gödel's work, defining mechanical computation and proving the halting problem's impossibility, which led to the theoretical blueprint for modern computers.
- The Darwin Gödel Machine represents a shift from provably optimal self-improvement (as in Schmidhuber's Gödel machine) to empirical, benchmark-driven adaptation, illustrating a trade-off between safety guarantees and practicality in AI.
- Research by Ben-David and colleagues showed that some machine learning problems, like certain ad targeting, are undecidable—their learnability depends on unresolved mathematical questions like the continuum hypothesis.
- Colbrook's theorem demonstrated that for certain problems (e.g., medical image reconstruction), a stable and accurate neural network exists but cannot be found by any training procedure, challenging the 'more data' assumption.
- Alfonseca's work argued that guaranteeing a superintelligent AI will not cause harm is provably impossible, akin to the halting problem, emphasizing inherent limits in AI safety certification.
- Gödel's legacy reveals that rule-based systems, including AI, cannot fully certify their own trustworthiness, boundaries, or safety, posing fundamental challenges for the industry despite commercial scaling efforts.