「Neural Computer」 paper updated to v2 within 9 days, removes assertion that 「the model itself is a computer」

According to Dongcha Beating monitoring, on April 7th, Meta AI and KAUST (King Abdullah University of Science and Technology) jointly published a paper titled "Neural Computers," proposing the concept of a "Neural Computer" (NC): enabling a single model to simultaneously handle computation, storage, and I/O, without relying on external programs like agents or solely learning environmental dynamics like world models. The paper envisions the ultimate form as a Completely Neural Computer (CNC) and demonstrates two prototypes using a video model to predict the next frame from commands, screen pixels, and user interactions in both CLI and GUI environments.

On April 16th, the authors released version 2 of the paper, with no new experiments but with systematic softening of key terms in the abstract. The statement from version 1 defining "NCs aim to make the model itself the running computer" was completely removed. Two modifiers were changed: "early NC primitives" became "elementary NC primitives," and "remain open" was altered to "remain challenging." The opening of the abstract in version 1 originally presented a direct comparison with traditional computers, agents, and world models to argue its point; version 2 removed this comparative section, moving the "new computing paradigm" from the defining sentence to the end as a long-term roadmap.

In other words, while the core experiments of the paper remained unchanged, the authors shifted their positioning from "The goal of NC is to make the model a computer" to "NC is just the beginning, and many challenges lie ahead."