The Uselessness of Useful Knowledge

Is artificial intelligence the new alchemy? That is, are the powerful algorithms that control so much of our lives — from internet searches to social media feeds — the modern equivalent of turning lead into gold? Moreover: Would that be such a bad thing?

According to the prominent AI researcher Ali Rahimi and others, today’s fashionable neural networks and deep learning techniques are based on a collection of tricks, topped with a good dash of optimism, rather than systematic analysis. Modern engineers, the thinking goes, assemble their codes with the same wishful thinking and misunderstanding that the ancient alchemists had when mixing their magic potions.

It’s true that we have little fundamental understanding of the inner workings of self-learning algorithms, or of the limits of their applications. These new forms of AI are very different from traditional computer codes that can be understood line by line. Instead, they operate within a black box, seemingly unknowable to humans and even to the machines themselves.

This discussion within the AI community has consequences for all the sciences. With deep learning impacting so many branches of current research — from drug discovery to the design of smart materials to the analysis of particle collisions — science itself may be at risk of being swallowed by a conceptual black box. It would be hard to have a computer program teach chemistry or physics classes. By deferring so much to machines, are we discarding the scientific method that has proved so successful, and reverting to the dark practices of alchemy?

Not so fast, says Yann LeCun, co-recipient of the 2018 Turing Award for his pioneering work on neural networks. He argues that the current state of AI research is nothing new in the history of science. It is just a necessary adolescent phase that many fields have experienced, characterized by trial and error, confusion, overconfidence and a lack of overall understanding. We have nothing to fear and much to gain from embracing this approach. It’s simply that we’re more familiar with its opposite.

After all, it’s easy to imagine knowledge flowing downstream, from the source of an abstract idea, through the twists and turns of experimentation, to a broad delta of practical applications. This is the famous “usefulness of useless knowledge,” advanced by Abraham Flexner in his seminal 1939 essay (itself a play on the very American concept of “useful knowledge” that emerged during the Enlightenment).

A canonical illustration of this flow is Albert Einstein’s general theory of relativity. It all began with the fundamental idea that the laws of physics should hold for all observers, independent of their movements. He then translated this concept into the mathematical language of curved space-time and applied it to the force of gravity and the evolution of the cosmos. Without Einstein’s theory, the GPS in our smartphones would drift off course by about 7 miles a day.

But maybe this paradigm of the usefulness of useless knowledge is what the Danish physicist Niels Bohr liked to call a “great truth” — a truth whose opposite is also a great truth. Maybe, as AI is demonstrating, knowledge can also flow uphill.

In the broad history of science, as LeCun suggested, we can spot many examples of this effect, which can perhaps be dubbed “the uselessness of useful knowledge.” An overarching and fundamentally important idea can emerge from a long series of step-by-step improvements and playful experimentation — say, from Fröbel to Nobel.

Perhaps the best illustration is the discovery of the laws of thermodynamics, a cornerstone of all branches of science. These elegant equations, describing the conservation of energy and increase of entropy, are laws of nature, obeyed by all physical phenomena. But these universal concepts only became apparent after a long, confusing period of experimentation, starting with the construction of the first steam engines in the 18th century and the gradual improvement of their design. Out of the thick mist of practical considerations, mathematical laws slowly emerged.

For another example, we can turn to the history of hydrodynamics. An immediate problem presented itself to early humans — transportation over various waterways — and they did what they could to overcome it, not worrying or even caring about a fundamental understanding of fluid dynamics. Over the millennia that followed, people built and sailed ships, engineering ever more efficient shapes based solely on empirical knowledge and experience.

Only in the 19th century did we stumble upon the famous Navier-Stokes equations that describe, in mathematical precision, the motion of fluids. Even then, the knowledge kept flowing uphill, as the advent of mechanical engines and higher speeds drove the need for theoretical considerations. Now the properties of these intricate equations form one of the unsolved million-dollar Millennium Prize problems, placing them at the frontier of fundamental mathematics.

One could even argue that science itself has followed this uphill path. Until the birth of the methods and practices of modern research in the 17th century, scientific research consisted mostly of nonsystematic experimentation and theorizing. Long considered academic dead ends, these ancient practices have been reappraised in recent years: Alchemy is now considered to have been a useful and perhaps even necessary precursor to modern chemistry — more proto-science than hocus-pocus.

The appreciation of tinkering as a fruitful path toward grand theories and insights is particularly relevant for current research that combines advanced engineering and basic science in novel ways. Driven by breakthrough technologies, nanophysicists are tinkering away, building the modern equivalents of steam engines on the molecular level, manipulating individual atoms, electrons and photons. Genetic editing tools such as CRISPR allow us to cut and paste the code of life itself. With structures of unimaginable complexity, we are pushing nature into new corners of reality. With so many opportunities to explore new configurations of matter and information, we could enter a golden age of modern-day alchemy, in the best sense of the word.

However, we should never forget the hard-won cautionary lessons of history. Alchemy was not only a proto-science, but also a “hyper-science” that overpromised and underdelivered. Astrological predictions were taken so seriously that life had to adapt to theory, instead of the other way around. Unfortunately, modern society is not free from such magical thinking, putting too much confidence in omnipotent algorithms, without critically questioning their logical or ethical basis.

Science has always followed a natural rhythm of alternating phases of expansion and concentration. Times of unstructured exploration were followed by periods of consolidation, grounding new knowledge in fundamental concepts. We can only hope that the current period of creative tinkering in artificial intelligence, quantum devices and genetic editing, with its cornucopia of useful applications, will eventually lead to a deeper understanding of the world.