Compression, Beauty, and Schmidhuber: Can Neural Networks Invent Something New?

In November 2025, Jürgen Schmidhuber, a pioneer of modern AI, referenced his 2008 research paper to comment on the current limitations of neural networks. The article explores the nature of creativity and whether AI is capable of generating truly novel ideas. Schmidhuber proposes a theory suggesting that the drive for data compression and pattern recognition underlies everything from humor to fundamental laws of physics. The piece examines why, according to this concept, modern Large Language Models (LLMs) can simulate creativity but lack the capacity for genuine creative thought. The author analyzes how the theory of 'beauty' as a measure of information compression explains the constraints of current architectures and what this implies for the future development of artificial intelligence.