Jürgen Schmidhuber, a prominent AI researcher and co-creator of the LSTM architecture, recently challenged the current understanding of AI creativity. In November 2025, he referenced his 2008 paper, 'Driven by Compression Progress,' to argue that modern large language models lack the fundamental mechanisms for true innovation.
Schmidhuber’s theory posits that curiosity and creativity are computable processes. He defines beauty as the subjective compressibility of data and interest as the rate of learning progress. According to his framework, an intelligent agent should be rewarded not for external results, but for the progress made in compressing its observations of the world.
The proposed architecture for a creative agent consists of two modules: a compressor that predicts observations and a controller that seeks out data that improves compression efficiency. Schmidhuber argues that while modern LLMs function as effective compressors, they lack this internal reward mechanism for curiosity, which prevents them from moving beyond imitation.
Recent studies support the link between intelligence and compression. For instance, DeepMind’s Chinchilla model demonstrated superior compression capabilities compared to traditional algorithms like Gzip. However, research published in PNAS Nexus suggests that commercial LLMs lack the diversity and intentionality found in human creative processes.
