Meta/KAUST propose 'neural computer'
Researchers at Meta AI and KAUST proposed a 'neural computer' architecture that folds computation, memory and I/O into a single learned model. (marktechpost.com) The idea underscores active research interest in model architectures and systems questions about how models compute and manage memory. (marktechpost.com)
Computers usually split work across a processor, memory, and input and output. A paper posted on April 9, 2026 from Meta AI and King Abdullah University of Science and Technology asks whether one neural network could do all three jobs inside a single learned state. (arxiv.org) The paper calls that setup a “Neural Computer,” or a model that updates an internal runtime state from what it sees and what a user does, then produces the next screen frame. The authors say the model itself would act as the running computer, not as software sitting on top of one. (arxiv.org) That is different from an artificial intelligence agent, which uses an existing operating system, terminal, or application programming interface to get work done. It is also different from a world model, which mainly learns to predict how an environment changes over time. (arxiv.org) The team says the long-run target is a “Completely Neural Computer,” a general-purpose system that is Turing complete, programmable, stable unless explicitly changed, and built around machine-like rules for execution. The paper also proposes a “run/update contract,” which would separate ordinary use from explicit behavior-changing updates. (arxiv.org) That contract tries to solve a current problem in machine learning systems: models often mix inference, memory, and adaptation in ways that make behavior hard to inspect or roll back. The authors argue a usable neural computer would need updates that leave traces and can be reviewed, rather than silent drift during normal use. (arxiv.org) For a first test, the researchers built two video-based prototypes instead of a full new machine architecture. They say video models are the most practical starting point because screen activity already bundles visual state, user input, and short-horizon transitions into one stream. (arxiv.org) One prototype, NCCLIGen, works in a command-line interface, where text commands and screen changes arrive in sequence. The other, NCGUIWorld, models a graphical user interface, where buttons, windows, and pointer actions have to stay aligned from frame to frame. (marktechpost.com) Both prototypes were built on Wan2.1, a video generation model, with added conditioning and action modules, and the two systems were trained separately rather than sharing parameters. The paper says those early systems show learned input-output alignment and short-horizon control, while leaving long-term memory, symbolic stability, and general reprogramming unsolved. (marktechpost.com) The authors also draw a line between this work and older “Neural Turing Machine” and “Differentiable Neural Computer” research, which added differentiable external memory to neural networks. Here, the question is broader: whether a learned system can begin to replace the external runtime stack itself. (arxiv.org) For now, the paper is a proposal with two narrow prototypes, not a replacement for conventional hardware or operating systems. But it puts architecture, memory, and control back at the center of artificial intelligence research by asking whether the model can become more than an app running on a machine. (arxiv.org)
