From Tokens to Atoms

So far, artificial intelligence has mainly been used to produce information: text, images, software, analysis, and recommendations. A new generation of systems, however, aims to move beyond the screen and participate in the development of medicines, materials, engines, medical devices, and other physical products.

The main purpose of this paper is to study the economic consequences of the transition from digital AI to what is often called physical AI. It separates product development into two essential stages: designing something and proving that it actually works.

Artificial intelligence can generate thousands or even millions of possible designs in a short period of time. The difficulty is that these designs must still be tested through prototypes, laboratories, clinical trials, safety assessments, and certification procedures. These physical processes do not improve at the same speed as computer models.

The paper’s main result is what it calls the “validation cost paradox.” As designing becomes cheaper, AI may generate so many candidates that the total cost of testing them increases. In other words, producing ideas at almost no cost does not mean that finished products will also become almost free.

Examples from materials science and AI-assisted drug development illustrate this difference. AI systems may greatly reduce the time required to identify a promising option, but physical testing, clinical trials, and regulatory approval can still take years.

The conclusion is that the real challenge of physical AI will not simply be producing better designs. It will be validating those designs safely and at scale. This will require better laboratories, reliable simulations, digital twins, and regulatory systems that can accept computational evidence. In the future, the ability to prove that an idea works may become more scarce and valuable than the ability to generate the idea itself.