🧠 AI⚪ NeutralImportance 6/10

Neural Computers

arXiv – CS AI|Mingchen Zhuge, Changsheng Zhao, Haozhe Liu, Zijian Zhou, Shuming Liu, Wenyi Wang, Ernie Chang, Gael Le Lan, Junjie Fei, Wenxuan Zhang, Yasheng Sun, Zhipeng Cai, Zechun Liu, Yunyang Xiong, Yining Yang, Yuandong Tian, Yangyang Shi, Vikas Chandra, J\"urgen Schmidhuber|April 10, 2026 at 04:00 AM

🤖AI Summary

Researchers propose Neural Computers (NCs), a new computing paradigm where AI models function as executable runtime environments rather than static predictors. The work demonstrates early NC prototypes using video models that process instructions and user actions to generate screen frames, establishing foundational I/O primitives while identifying significant challenges toward achieving general-purpose Completely Neural Computers (CNCs).

Analysis

The Neural Computer proposal represents a conceptual shift in how machine learning systems could function as computing substrates. Rather than treating AI as a tool that operates within traditional computational frameworks or as agents that interact with external environments, this work explores making the model itself the executing computer—a fundamental architectural departure that could reshape machine intelligence design.

The research grounds this ambitious vision in practical experiments using video generation models trained on I/O traces without explicit program state instrumentation. By training models to predict screen outputs from instructions, pixel inputs, and user actions in CLI and GUI environments, the authors demonstrate that neural systems can learn rudimentary interface primitives and short-horizon control. This empirical grounding distinguishes the proposal from pure theory, though the authors candidly acknowledge significant open problems: routine reuse, controlled updates, and symbolic stability remain unsolved.

For the AI research community, this work opens a new research direction that bridges classical computing architecture, machine learning, and world modeling. The vision of CNCs—mature, general-purpose neural computers with stable execution and reprogramming capabilities—would represent a paradigm as transformative as the von Neumann architecture. However, the gap between current video models and systems capable of complex, durable computation is substantial.

The roadmap outlined invites continued investigation into whether purely learned systems can acquire the stability, composability, and symbolic reasoning required for general-purpose computation. Success would reshape AI systems architecture; failure would clarify inherent limitations of neural approaches to computation.

Key Takeaways

→Neural Computers propose treating AI models as executable runtime environments that unify computation, memory, and I/O rather than as external tools or agents.
→Early prototypes using video models demonstrate learned I/O alignment and short-horizon control, validating core NC concepts but leaving major challenges unresolved.
→Achieving general-purpose Completely Neural Computers requires solving routine reuse, controlled updates, and symbolic stability—currently open problems with no clear solutions.
→This research establishes a new research direction distinct from world models and agentic AI, potentially defining computing paradigms for the next decade.
→The work emphasizes empirical validation through I/O trace learning rather than instrumented program state, demonstrating feasibility of learning computing primitives from raw behavioral data.