MIT proposes open-ended AI representation framework

While LLMs can generate impressive combinations within their training distribution, they are fundamentally trapped in a representational jail of human-defined tokens and concepts. Genuinely autonomous scientific and creative discovery will remain out of reach until AI can dynamically invent, store, and verify its own novel conceptual building blocks. The vocabulary problem restricts current models to static token vocabularies and predefined semantic spaces, making the invention of truly novel primitives extremely difficult. Furthermore, the verifier bottleneck complicates validating new concepts, as their utility might only be proven after multiple downstream uses requiring long-term planning and adaptive verification. To address this, the proposed framework re-imagines intelligence as cognitive discrepancy reduction, shifting the target from matching human outputs to modifying the representational frame itself.