Schmidhuber’s 'Neural Computers' paper shared

A new arXiv paper from Jürgen Schmidhuber and Meta-led coauthors argues that a neural network could act as the computer itself, not just software running on one. (arxiv.org) The paper, “Neural Computers,” was submitted to arXiv on April 7, 2026, by 19 authors including Mingchen Zhuge, Yuandong Tian, Vikas Chandra and Schmidhuber. It defines a “Neural Computer” as a system that unifies computation, memory and input-output inside a learned runtime state. (arxiv.org) In plain terms, the proposal asks whether a model’s internal activations — the shifting numbers inside the network while it runs — could do the jobs now split across code, memory and operating-system interfaces. The authors’ long-term label for that idea is the “Completely Neural Computer.” (arxiv.org) The first prototypes are narrower than that label suggests. The paper says the team instantiated the idea as video models that predict screen frames from instructions, pixels and, when available, user actions in command-line interface and graphical user interface settings. (arxiv.org) That means the model is trained on what goes in and what appears on screen, rather than on the hidden program state inside a conventional machine. The authors write that these early systems learned interface primitives such as input-output alignment and short-horizon control, while “routine reuse, controlled updates, and symbolic stability” remain unsolved. (arxiv.org) The paper also draws a line between this work and earlier “neural computer” research such as the Neural Turing Machine and Differentiable Neural Computer from Alex Graves and colleagues. It says the new proposal is not mainly about attaching differentiable external memory, but about whether the model can become “the running computer itself.” (arxiv.org) That distinction landed at a moment when large models are already being used to write code, call tools and operate desktop interfaces. The authors explicitly place their proposal against three existing frames — conventional computers, agents that act in external environments, and world models that learn environment dynamics. (arxiv.org; metauto.ai) The public release around the paper pushed the idea beyond a PDF. A GitHub repository described as the project’s “data engine” says its first open-source release in April 2026 covers data generation for command-line interface and graphical user interface trajectories. (github.com) The paper is also listed for the International Conference on Learning Representations 2026 Workshop on World Models, where OpenReview shows it was posted on March 1, 2026 and modified on March 4. That places the current burst of attention after the workshop posting and after the April 7 arXiv release. (openreview.net; arxiv.org) Schmidhuber’s name gave the paper extra reach because he remains one of the field’s best-known neural-network researchers and is now a professor and generative artificial intelligence center co-chair at King Abdullah University of Science and Technology. The immediate claim in this paper, though, is modest: the model can imitate pieces of a computer’s interface behavior, and the harder parts of stable reusable computation are still open. (kaust.edu.sa; arxiv.org)