Meta and KAUST want to turn AI models into computers

For the past two years, the AI playbook has been pretty consistent: build a smarter model, give it a browser and mouse, and have it use your computer like a remote intern. But what if we just skipped the computer entirely and turned the model itself into the machine?

Researchers from Meta and KAUST just published a paper proposing exactly that: Neural Computers (NCs). The core idea is to fold computation, memory, and I/O directly into a single learned runtime state, rather than using external software stacks.

Rolling out screens as video frames

In standard computing, explicit programs execute deterministic code. AI agents operate tools inside an external OS, while world models simulate physical environments. Meta and KAUST argue that these scattered layers should eventually collapse into a unified neural runtime.

To test whether this is even feasible, the team built early NC prototypes purely from recorded I/O traces. Instead of emulating CPUs, they instantiated them as video models that roll out screens frame by frame based on user instructions, pixels, and keystrokes.

They demonstrated two setups: a CLI terminal simulator and a GUI desktop simulator. The terminal setup can simulate typing, command execution, and Python sessions, faithfully rendering text scrolls and window shifts. The GUI version mimics cursor trajectories, clicks, and menu expansions.

That said, it is very early. The prototypes only sustain interactions for a few to tens of seconds. Run them any longer, and they drift off course. Symbolic stability also degrades rapidly during complex calculations, and because the model lacks durable memory, it essentially hallucinates state anew.

The road to Completely Neural Computers

The researchers frame this prototype as step one toward what they call the Completely Neural Computer (CNC). A true CNC would need to solve four massive challenges: Turing completeness, general-purpose programmability, behavioral consistency, and machine-native semantics.

If those hurdles are ever cleared, the computing experience could fundamentally shift. Traditional files, directories, and pre-packaged applications would fade away, replaced by an adaptive neural runtime that reshapes itself into whatever tool you need on demand.

Why it matters

For builders and end users, this is not something you can download and use for real work today. It is a foundational research paper, and running basic symbolic logic through a video generation pipeline is currently far too unstable and compute-heavy to replace real operating systems.

Yet the conceptual pivot is worth paying close attention to. While the rest of the industry is obsessed with building better scaffolding to let agents click buttons in Windows or macOS, Meta and KAUST are questioning whether those legacy human-centric interfaces should exist at all.

Rendering a terminal via video frames might look wildly inefficient on the surface. But we think the core value here is rethinking how software executes. If models can eventually master stable internal computation, the boundary between the AI and the computer running it will completely disappear.