Neural Computers claims
A new paper from the SchmidhuberAI lab describes ‘Neural Computers’ that the authors say reach 98.7% GUI cursor accuracy and show an 83% gain in arithmetic reasoning on their benchmarks. (x.com) The summary frames these models as architectures that increasingly act like the computer itself for interface and reasoning tasks. (x.com)
A new arXiv paper from a Meta AI and KAUST team introduces "Neural Computers" and reports 98.7% GUI cursor accuracy and about an 83% arithmetic gain on their benchmarks. (arxiv.org) Neural Computers are defined by the authors as neural models that unify computation, memory and input/output in a single learned runtime state rather than executing separate programs. (arxiv.org) The 19-author preprint lists Mingchen Zhuge and Jürgen Schmidhuber among its contributors and was submitted to arXiv on April 7, 2026; the team prototypes NCs as video models for both command-line and graphical interfaces. (arxiv.org) The paper and independent summaries report that explicit pixel-level or SVG cursor supervision boosted positional accuracy to about 98.7%, while coordinate-only supervision reportedly performed poorly (around 8.7%). (paperium.net) On curated REPL (read–eval–print loop) arithmetic probes the authors show performance jumps under reprompting and report arithmetic-probe results near 83% on clean terminal traces. (paperium.net) The experiments used two data regimes the paper calls CLIGen (terminal traces) and GUIWorld (graphical interactions); outside summaries say the authors trained on roughly 1,100 hours of terminal data and about 1,510 hours of GUI recordings and consumed tens of thousands of GPU-hours. (paperium.net) The authors explicitly list remaining challenges — routine reuse, controlled updates, long-horizon state persistence, and weak native symbolic stability — and say NCs currently capture only early I/O and short-horizon control primitives. (arxiv.org) The paper frames a long-term goal called the "Completely Neural Computer" and includes links to a GitHub data pipeline and a blog post for replication and follow-up experiments. (arxiv.org) "The paper asks, 'Can a single set of weights act as a “computer”?' " the authors write, and they publish code and data to let the community test that claim next. (arxiv.org)
