Mingchen Zhuge, founder of research lab Metauto, has published an essay proposing what he calls the Neural Computer (NC). The idea: AI models absorb runtime responsibilities that currently belong to the program stack, toolchain, and control layer. Instead of agents using computers, AI becomes a kind of computer. The completed form, dubbed Completely Neural Computer (CNC), would organize around neural runtime rather than explicit programs, tasks, or environments. Zhuge calls it a "pre-consensus" idea. The kind of concept multiple researchers are likely arriving at independently as agents move from experiments to infrastructure.
And agents have come a long way. 2023's MetaGPT struggled to spit out a few hundred lines of code. Now AI coding tools like Cursor, Claude Code, and OpenClaw are genuine productivity tools. World models made a similar jump. Waymo uses them for autonomous driving simulation. NVIDIA is building Cosmos for physical AI. Meanwhile, conventional computers are straining under open-ended, long-horizon tasks where the traditional software stack feels heavy.
NC pushes past "build a better agent" or "make a world model for computer environments." Zhuge is asking whether models can handle responsibilities that have always belonged to the machine itself. Staying stable over long tasks. Retaining and reusing capabilities. Preserving workflows. Current approaches bolt on structure through scaffolds and tighter action loops. NC suggests moving control logic into the model's runtime layer instead. It's distinct from Alex Graves' Neural Turing Machine lineage and hardware plays like Taalas, which turns specific models into deployment units but stops short of redefining what a machine is.
Whether this is prescient or premature is anyone's guess. Some Hacker News commenters questioned whether the direction is even necessary. Companies like Google DeepMind, World Labs, and OpenAI are exploring related directions through world model and agent projects. Nobody has shipped anything resembling a Neural Computer. The essay is more framework than product roadmap. But it nails a question the industry is wrestling with: if capabilities can enter runtime and stay governable there, what exactly counts as a computer anymore?