on Toshareproject.it - curated by Bruce Sterling
*It’s a proposal — replacing the programmable computers of the past with Turing-equivalent systems that are based on visual neural-net models.
https://arxiv.org/pdf/2604.06425
Introduction
Can a single set of weights act as a “computer”? We term this abstraction a Neural Computer (NC): a neural system that unifies computation, memory, and I/O in a learned runtime state. This usage is distinct from the Neural Turing Machine / Differentiable Neural Computer line (Graves et al., 2014, 2016): our concern is not differentiable external memory, but whether a learning machine can begin to assume the role of the running computer itself.
To implement this idea, we instantiate NCs as video models. At this stage, video models are the most practical substrate for this prototype, though we expect the long-term solution to require a fundamentally new neural architecture (Section 4). This implementation draws on several technical lines.
World models (Ha and Schmidhuber, 2018) show that neural networks can internalize environment dynamics and support predictive imagination, while high-capacity video generators such as Veo 3.1 (Google, 2025) and Sora 2 (OpenAI, 2025) show that such learned dynamics can be rendered into coherent frame sequences.
Frontier interactive video models such as Genie 3 (Bruce et al., 2024) further extend this trajectory toward action-controllable generative environments.
These lines provide practical machinery for current NC prototypes, but do not by themselves define the NC abstraction. In parallel, LLM-driven UI systems such as Imagine with Claude1 map
natural-language inputs to structured interface updates. Yet these capabilities remain split across different systems objects: conventional computers execute explicit programs, agents act through external execution environments, and world models render or predict environment dynamics, while executable state still resides outside the model.
NCs are motivated by this gap: they are not a smarter layer on top of the existing stack,
but a proposal to make the model itself the running computer.
The immediate question in this paper is whether early runtime primitives can be learned directly from raw interface I/O without privileged access to program state.
Throughout this paper, NC denotes this proposed machine form, while CNC denotes its mature, general-purpose realization….