What vs. How (5boiryzrsw74q.ok.kimi.link)

Researchers propose a "What vs. How" framework that reconceives AI intelligence as a collaboration between a passive, high-dimensional compressed world model ("what") and active procedural systems ("how"). The world model stores encyclopedic knowledge; an innate policy network—shaped by frequency and fitness—provides hardwired response tendencies; designed operational protocols (user commands, task-specific rules) steer behavior toward goals; and a self-correcting mechanism reflects on errors to rewrite protocols and accumulate wisdom. Rather than treating AI as just a brain or a library, this model emphasizes a clean separation between knowledge representation and procedural execution, and sees true intelligence emerging from their interaction. For the AI/ML community this has practical and conceptual implications: architectures should focus less on packing ever more "what" and more on scaffolding robust "how" mechanisms—protocol design, meta-learning, and error-driven protocol rewriting. Large context windows let models consume gigabytes of navigational data (rich operational context) to produce precise, goal-directed behavior from general capabilities. The framework suggests incremental paths toward reliable, general intelligence by layering operational protocols and self-correction on top of powerful world models, shifting research priorities toward modularity, interpretable procedures, and continuous protocol optimization rather than raw knowledge accumulation alone.