Quanta Magazine

Artificial intelligence is the new Alchemy? Is artificial intelligence the new alchemy? Are the algorithms that control our lives, from social media feeds to internet searches, the modern equivalent of turning lead into silver? Furthermore: Is that a bad thing?
Ali Rahimi, an AI researcher, and others claim that today's popular neural networks and deep-learning techniques are based on a variety of tricks and a little bit of optimism rather than systematic analysis. According to modern engineers, they assemble their codes with the same wishful thinking as the ancient alchemists when creating their magical potions.

We don't have a deep understanding of the inner workings or limits of self-learning algorithms. These new forms AI are quite different from the traditional computer codes that can only be read line by line. They operate in a black box that is unknowable to both humans and machines.

All sciences will be affected by this discussion in the AI community. Deep learning has a huge impact on many areas of research, from drug discovery to smart material design to analysis of particle collisions. Science itself could be affected by this conceptual black box. It would be difficult to have a computer program instruct physics or chemistry classes. Are we abandoning the scientific method that has been so successful and turning to the dark arts of alchemy by delegating so much to computers?

Yann LeCun was co-recipient the 2018 Turing Award, for his pioneering work in neural networks. The current state of AI research in science is not new, he argues. This is a normal adolescent stage that many fields have gone through, marked by confusion, trial and error and an inability to fully grasp the issues. This approach is not something to be afraid of. It has much to offer. It's just that we are more familiar with the opposite.

It is easy to see knowledge flowing downstream from an abstract idea to the end of the experimentation process to the delta of practical applications. Abraham Flexner's seminal 1939 essay argued this famous idea of the usefulness of unutilized knowledge. It is a play on the American notion of useful knowledge that was created during the Enlightenment.

Albert Einstein's general theory about relativity is a great example of this flow. The basic idea that all observers should be able to observe the laws of physics independently of their movements was the basis of it all. This idea was then converted into the mathematical language curved space-time, and applied it to gravity and the development of the cosmos. The GPS on our smartphones would drift by 7 miles per day without Einstein's theory.

This paradigm of useless knowledge being useful is perhaps what Niels Bohr, a Danish physicist, liked to call a great fact. It is a truth that has an opposite. Perhaps knowledge can flow uphill, as AI has shown.

LeCun suggested that there are many examples of this effect in the history of science.

The best example is perhaps the discovery of the laws and principles of thermodynamics. This foundational science branch is perhaps the most striking. These laws, which describe the conservation of energy as well as the increase in entropy are natural laws that all physical phenomena must obey. These universal concepts were only discovered after long and confusing experimentation. This began with the 18th century's construction of steam engines, and continued through the improvement of their designs. Mathematical laws emerged slowly from the dense fog of practical considerations.

Another example is the history of hydrodynamics. Early humans faced a problem with transportation on various waterways. They did their best to solve it and didn't worry about fluid dynamics. People built and sailed ships over the millennia, relying only on empirical knowledge and their experience to create ever more efficient forms.

The famous Navier-Stokes equations, which describe fluid motion in mathematical precision, were discovered only in the 19th Century. As mechanical engines and faster speeds made it possible to study fluid motion, knowledge continued to flow. These complex equations are now a part of the Millennium Prize puzzle, and they have been placed at the frontiers of fundamental mathematics.

This is even true for science. Scientific research was largely based on non-systematic experimentation and theory until the advent of modern research methods in the 17th century. These ancient practices were long considered academic dead ends. However, they have been revived in recent years. Alchemy is now seen as a useful and possibly even necessary precursor for modern chemistry that is more proto-science than hokus-pocus.

Tinkering is a way to discover new insights and theories. This is especially relevant to current research that combines advanced engineering with basic science in innovative ways. Nanophysicists are working hard to create the next generation of steam engines at the molecular level. They manipulate individual atoms, electrons, and photons using cutting-edge technologies. CRISPR allows us to copy and paste the code for life using genetic editing tools. We are pushing nature to new horizons with structures of unprecedented complexity. We could be entering a golden age in modern-day alchemy with so many possibilities to explore new combinations of matter and information.

We should not forget the lessons learned from history. Alchemy was not just a proto-science but also a hyperscience that promised too much and didn't deliver. Astrological predictions were so serious that the world had to adapt to theory instead of vice versa. Modern society isn't immune to such magical thinking. They trust in magical algorithms without questioning their ethical or logical basis.

Science has always been characterized by a natural cycle of expansion and consolidation. Periods of unstructured exploration were immediately followed by periods that saw the consolidation of new knowledge and fundamental concepts. It is possible to believe that the current period for creative tinkering with artificial intelligence, quantum devices, and genetic editing will lead eventually to a deeper understanding.