Researchers have developed Neural Computers, a novel machine architecture where a neural network directly serves as the running computer, eliminating the need for external software stacks.
Meta AI and the King Abdullah University of Science and Technology (KAUST) have pioneered Neural Computers (NCs), a fundamentally different approach to machine computation. Unlike traditional computers, AI agents, or world models, NCs integrate all the components of a running computer—executable context, working memory, and interface state—within a single neural network. This is achieved through a system defined by an update function (F θ) and a decoder (G θ) operating on a latent runtime state (h t), which updates based on observations (x t) and user actions (u t), and subsequently generates the next frame (x t+1). The ultimate goal is a Completely Neural Computer (CNC), possessing Turing completeness, universal programmability, consistent behavior, and machine-native semantics. A key aspect of this CNC is a “run/update contract” ensuring that inputs execute installed capabilities without modification, while behavior changes require explicit programming interfaces with traceable updates. To demonstrate this concept, the researchers built two video-based prototypes: NC CLIGen and NC GUIWorld. Both prototypes were constructed upon Wan2.1, a state-of-the-art video generation model, with specialized modules added to facilitate command-line interface (CLI) and graphical user interface (GUI) interaction. Crucially, the models were trained independently, lacking shared parameters. Evaluation was conducted in open-loop mode, meaning they were triggered by recorded prompts and action streams, not directly interacting with a live environment. NC CLIGen focuses on generating terminal interactions from text prompts and initial screen frames, treating CLI generation as a text-and-image-to-video task. DATA: * **Researchers:** Meta AI and the King Abdullah University of Science and Technology (KAUST) * **Technology:** Neural Computers (NCs) * **Key Feature:** Neural network acts as the running computer, integrating OS stack components. * **Prototypes:** NC CLIGen (CLI), NC GUIWorld (GUI) * **Base Model:** Wan2.1 (video generation) * **Training:** Independent, without shared parameters * **Evaluation:** Open-loop (triggered by prompts and action streams) * **Goal:** Completely Neural Computer (CNC) - Turing complete, programmable, consistent behavior, machine-native semantics * **Operational Requirement:** Run/update contract for predictable execution. ```
✨ This report was generated by AI News Assistant.
Tags:
United States