While many believe that monolithic models are the best approach, DeepSeek appears to adopt a multi-model architecture, delegating tasks to smaller, specialised models rather than relying on a single, monolithic system. Each model is purpose-built for specific functions, enhancing efficiency. This specialisation not only boosts performance but also reduces data and resource requirements, resulting in a streamlined and cost-effective design.
Such a modular approach is consistent with how the brain operates. Though seemingly monolithic, the brain is inherently modular, with distinct regions dedicated to specific functions—much like specialised models in a multi-model system.
The real challenge lies in making everything work together efficiently. Reinforcement learning likely plays a pivotal role in DeepSeek’s development, as it enhances coordination, decision-making, and adaptability, ensuring cohesive and effective collaboration among specialised models within its multi-model framework. Similarly, the brain utilises feedback loops, allowing the output from one module to influence the activity of another, thereby promoting integrated functioning.
Anders Hast 