Show HN: Progressive Cognitive Architecture – Training LLMs in 4 Phases (medium.com)

A new approach to training large language models (LLMs) called "Progressive Cognitive Architecture" has been proposed, aiming to model AI learning more like human cognitive development. This framework consists of a four-phase developmental pipeline: foundational learning of exact calculations, structural pruning for intuitive understanding, teaching delegation to external tools, and finally, orchestrating these tools for complex problem-solving. The significance of this model lies in its potential to transform how AI systems handle tasks, moving from brute-force computation to a more nuanced understanding that mimics human intuition. The architecture is evident in the experimental setup utilizing a 1.5 billion parameter model, Qwen2.5, and incorporates strategies like structured pruning and progressive learning across various cognitive modes. By emphasizing intuitive reasoning over deterministic computation, this research not only challenges conventional AI training methods but also suggests a path for creating more sophisticated systems capable of higher-level reasoning without losing fundamental competencies. This aligns with established ideas in cognitive science and has implications for the future of AI, particularly in enhancing the efficiency and effectiveness of learning mechanisms across different tasks.