On April 7th, Meta AI and a team led by Jürgen Schmidhuber—one of the co-inventors of the LSTM architecture—published a paper titled "Neural Computers." It introduced the concept of a "Neural Computer" (NC): a model that would handle computation, storage, and input/output all by itself. No more calling external programs like an agent, and no more just learning environmental dynamics like a world model. The paper called the ultimate form a "Completely Neural Computer" (CNC) and built two prototypes using video models. These prototypes predicted the next screen frame from instructions, screen pixels, and user actions in both CLI and GUI environments.
https://twitter.com/MingchenZhuge/status/2046138600073441688
Then, on April 16th, the authors quietly swapped in a v2 of the paper. No new experiments were added, but the abstract's key language was systematically softened. The defining sentence from v1—"NCs aim to make the model itself the running computer"—was deleted entirely. Two modifiers got tweaked: "early NC primitives" became "elementary NC primitives," and "remain open" became "remain challenging." The v1 abstract had opened by positioning the idea in direct contrast to traditional computers, agents, and world models. The v2 version removed that contrast and moved the phrase "new computing paradigm" from the definition to the end, framing it as a long-term roadmap instead.
In other words, the core experiments stayed the same, but the authors dialed back their own positioning from "the goal of NC is to make the model the computer" to "NC is just a starting point, and many challenges remain."