Import AI 456: Meta and KAIST Propose 'Neural Computer' — Replacing OS-Level Computation with a Single Learned Substrate

Import AI·medium signal

Researchers from Meta and KAIST (led by Jürgen Schmidhuber) explore whether neural networks can function as traditional computers, demonstrating early prototypes using generative video models to create command-line and graphical interfaces. The paper envisions a 'Completely Neural Computer' where computation, memory, and I/O unify in a single learned substrate rather than being engineered separately — a radical departure from von Neumann architecture.