AGI as a cognitive OS: what if we look for programs, not weights

The article examines the current state of Large Language Models (LLMs) through the lens of their lack of architectural transparency. The author draws an analogy to binary files where the source code is unavailable: we successfully 'compile' intelligence from massive datasets, yet we do not understand the principles of its internal organization. Instead of focusing solely on optimizing neural network weights, the author suggests rethinking the approach to AGI by treating it as a cognitive operating system. The main thesis is the need to shift from purely statistical training methods to searching for algorithmic structures that could serve as the 'source code' for artificial intelligence. This would make models more interpretable, reliable, and manageable, transforming them from 'black boxes' into full-fledged software systems with understandable logic.