The large language models are transforming numerous sectors, and yet their race toward supreme intelligence presents an insurmountable limit: the absence of a physical body.
A recent study published by Tom Zahavy of Google DeepMind, entitled “LLMs can’t jump“, summarizes this fundamental truth, stating in no uncertain terms that software will never take the place of human creative genius.
According to the research, the cognitive limits arising from the lack of a physical presence force machines to handle only partial forms of reasoning exceptionally well, preventing them from matching the true scientific invention.
L’ingegno umano è meglio dell’IA? Google sembra convinta
The document analyzes in depth the types of logical inference, dividing them into three categories: induction, deduction and abduction. Thanks to the statistical processing of billions of parameters and the compression of enormous volumes of data, modern software master induction impeccably.
At the same time, advanced systems such as AlphaProof have demonstrated excellent deductive abilities, managing to derive perfect logical proofs starting from predefined mathematical rules.
The real obstacle for algorithmic models emerges with abduction, namely the ability to formulate original explanatory hypotheses when faced with scarce or fragmentary information. This specific cognitive approach completely eludes current machines.
La smentita a Schmidhuber e il caso Einstein
The German computer scientist Jürgen Schmidhuber has long maintained that scientific discovery is nothing more than a highly advanced form of data compression. The researchers at DeepMind categorically reject this view, appealing to the General Theory of Relativity to demonstrate otherwise.
During the development of his deepest concepts, Albert Einstein did not have enormous datasets to study. Instead, the physicist relied on mental experiments, imagining, for example, the physical sensations of a person in free fall inside a closed elevator. By directly linking sensory experiences to an abstract thought, Einstein was able to formulate novel principles.
A modern language model could effortlessly perform 100% of the complex deductive calculations derived from Einstein’s equations, but it would never be able to conceive the original premises by analyzing preexisting texts.
La necessità di un mondo simulato
Adding millions of servers or massively increasing computing power will not be enough to imbue these systems with intelligence comparable to human intelligence.
Currently, automated tools are limited to applying rules within known boundaries, without creating truly new conceptual structures. The lack of grounding in the tangible world makes it, in fact, impossible to conceive intuitive conclusions about cause and effect.
The document suggests that in order to bridge the gap between logical computation and genuine invention, artificial intelligence architectures will need to move toward the multimodality applied to three-dimensional spatial models.
Only by endowing the systems with the ability to perform simulations within virtual environments will it be possible for them to develop a real physical intuition, an essential step before they can formulate formulas or theorems that did not exist until then.
