Meta AI has introduced the concept of "Neural Computers" (NCs), a paradigm shift in autonomous system architecture. Rather than an AI model controlling an external computer, the Neural Computer integrates memory, computation, input, and output directly within the model itself, effectively transforming it into a self-sufficient computational apparatus. This idea, championed by researchers including Jürgen Schmidhuber, directly challenges the prevailing paradigm of AI agents.
The practical implementation of the NC concept involves training video models to simulate computer interfaces. A model trained on 1,100 hours of terminal usage videos and approximately 1,500 hours of mouse and keyboard interactions can generate videos mimicking operating system operations. For instance, it can 'see' a terminal screen, receive commands, and then generate the next frame showing those commands already executed. Visually, this looks convincing, with a realistic terminal appearance.
However, beneath the impressive surface lie significant limitations. The accuracy of text within the generated interfaces remains inconsistent. There are also logical gaps: the model’s accuracy in performing arithmetic operations is a mere 4%. The authors do note, however, that "clever prompts" can boost this figure to 83%. As researchers point out, the model currently "looks like a computer" more than it "thinks like a computer."
The advent of Neural Computers could radically reshape future IT infrastructure, paving the way for a new generation of autonomous systems. These systems would not only perform complex computational tasks but also possess the potential for self-modification, opening the door to entirely new computational platform architectures and potentially catalysing the emergence of fully neural computers. For businesses, this signifies the potential for a new tier of autonomous systems capable of self-sufficient operation, demanding a re-evaluation of current IT strategies.