Researchers propose 'neural computers'
Meta AI and KAUST researchers proposed a 'neural computer' concept that folds computation, memory and I/O into a single learned model, according to a recent research summary. The work is early-stage research rather than an operational deployment, focused on reducing inefficiencies from moving data between layers. (marktechpost.com)
A computer usually splits work across separate parts for processing, memory, and input or output. A new paper from Meta AI and King Abdullah University of Science and Technology argues those jobs could be learned inside one model instead. (arxiv.org) The paper, titled *Neural Computers*, was submitted to arXiv on April 7, 2026, by 19 authors including Mingchen Zhuge, Changsheng Zhao, Vikas Chandra, Yuandong Tian, and Jürgen Schmidhuber. The authors describe “Neural Computers” as a machine form that “unifies computation, memory, and I/O in a learned runtime state.” (arxiv.org) In plain terms, the proposal asks whether a model can be the machine that runs a task, not just the software sitting on top of a conventional machine. The paper contrasts that idea with standard computers, which execute explicit programs, and with agents, which act through external tools and environments. (arxiv.org) The first prototype is not a new chip or operating system. The researchers built early versions as video models that predict screen frames from instructions, pixels, and sometimes user actions in command-line and graphical interfaces. (arxiv.org) That setup targets a familiar bottleneck in computing: moving information back and forth between separate components. The paper says current systems pay repeated costs when data has to shuttle among compute, memory, and input or output layers, and it proposes a learned runtime as one way to reduce that separation. (arxiv.org) The authors say the early systems learned limited interface skills, especially lining up inputs with outputs and handling short sequences of control steps. They also say bigger problems remain unsolved, including routine reuse, controlled updates, and symbolic stability over longer runs. (arxiv.org) The paper draws a line between this work and older “neural computer” research such as Neural Turing Machines and Differentiable Neural Computers. Here, the goal is not an external differentiable memory module attached to a network, but a model that begins to take on the role of the running computer itself. (arxiv.org) The researchers frame the end goal as a “Completely Neural Computer,” or a general-purpose system with stable execution, explicit reprogramming, and durable reuse of capabilities. In the current paper, that remains a roadmap rather than a deployed system or benchmark-beating product. (arxiv.org) For now, the contribution is a research agenda with an initial prototype: use learned models to handle more of what today is divided among programs, memory, and interfaces. The paper says whether that becomes a practical computing architecture depends on solving the stability and control problems it lays out. (arxiv.org)
