Much to the chagrin of vacationers planning a summer picnic, the weather is an incredibly capricious and unpredictable thing. Small changes in rainfall, temperature, humidity, wind speed or wind direction can change outdoor conditions over hours or days. Therefore, weather forecasts are usually not made more than seven days into the future - and therefore picnics require contingency plans.
But what if we could understand a chaotic system well enough to predict how it will behave far in the future?
Can you predict the weather for the year?
In January 2018, scientists succeeded. They used machine learning to accurately predict the outcome of a chaotic system over a much longer period than was thought possible. And the machine did it simply by observing the dynamics of the system, having no idea of the equations behind it.
Awe, fear and excitement
We have already begun to get used to the incredible manifestations of artificial intelligence.
Last year, a program called AlphaZero learned the rules of the game of chess from scratch in just a day, and then beat the world's best chess software. She also learned to play Go and surpassed the former silicon champion, the AlphaGo Zero algorithm, which improved in the game through trial and error after being fed the rules.
Promotional video:
Many of these algorithms start with a pure state of blissful ignorance and quickly gain knowledge by watching the process or playing against themselves, improving at every step thousands of times per second. Their abilities inspire feelings of fear, awe, excitement. We often hear about the chaos into which they can plunge humanity one day.
But it is much more interesting what artificial intelligence will do with science in the future, with its “understanding”.
Perfect forecasting means understanding?
Most scientists will probably agree that prediction and understanding are not the same thing. The reason lies in the myth about the origin of physics - and, one might say, modern science in general.
The fact is that for over a thousand years people have used the methods proposed by the Greco-Roman mathematician Ptolemy to predict the movement of planets across the sky.
Ptolemy knew nothing about the theory of gravity or that the sun was the center of the solar system. His methods included ritual calculations using circles within circles within circles. And while they predicted planetary motion pretty well, no one understood why it worked or why the planets obey such seemingly complex rules.
Then there were Copernicus, Galileo, Kepler and Newton.
Newton discovered the fundamental differential equations that govern the motion of each planet. With their help, it was possible to describe every planet in the solar system. And that was great because we understood why the planets move.
Solving differential equations turned out to be a more efficient way to predict planetary motion in comparison with Ptolemy's algorithm. More important, however, is that our belief in this method has allowed us to discover new invisible planets, thanks to the law of universal gravity. He explained why rockets fly and apples fall, and also why moons and galaxies exist.
This basic pattern - finding a set of equations describing a unifying principle - has been used successfully in physics over and over again. This is how we defined the Standard Model, the culmination of half a century of particle physics research, that accurately describes the structure of every atom, nucleus, or particle. This is how we try to understand high temperature superconductivity, dark matter, and quantum computers. (The unjustified effectiveness of this method even raised questions about why the universe lends itself so well to mathematical description.)
Throughout science, understanding something means going back to the original schema: if you can reduce a complex phenomenon to a simple set of principles, you understand it.
Exceptions to the rule
And yet, there are annoying exceptions that spoil this beautiful story. Turbulence is one of the reasons it is difficult to predict the weather - a prime example from physics. The vast majority of problems from biology, from entangled structures in other structures, also defy explanation by simple principles of unification and simplification.
While there is no doubt that atoms and chemistry, and thus the simple principles underlying these systems, are described using universally effective equations, this is a rather ineffective way of generating useful predictions.
At the same time, it is becoming apparent that these problems lend themselves easily to machine learning methods.
Just as the ancient Greeks looked for answers from the mystical Delphic oracle, we will look for answers to the most complex questions of science from omniscient oracles with artificial intelligence.
Such oracles are already driving autonomous vehicles and choosing investment targets in the stock market, and very soon they will predict which drugs will be effective against bacteria - and what the weather will be like in two weeks.
They will make these predictions with the highest precision that we never dreamed of, without using any mathematical models and equations.
It is possible that, armed with data on billions of collisions at the Large Hadron Collider, they will do better at predicting the outcome of an experiment with particles than even the beloved Standard Model.
Similar to the inexplicable sources of the Delphi priestesses' revelation, our artificial intelligence prophets are also unlikely to be able to explain why they predict this way and not otherwise. Their conclusions will be based on many microseconds of what might be called "experience." They will be like an uneducated farmer who knows how to accurately predict how the weather will change, "because the bones ache" or other premonitions.
Science without understanding?
The implications of the work of machine intelligence in the field of science and the philosophy of science can be startling.
For example, in the face of ever more accurate predictions, albeit obtained by methods that are incomprehensible to humans, will we deny that machines have better knowledge than we do?
If forecasting is really the main goal of science, how should we modify the scientific method, the algorithm that has enabled us to identify errors and correct them for centuries?
If we give up understanding, is there any point in doing the science we were doing?
No one knows. But if we can't articulate why science is more than the ability to make good predictions, scientists will soon find that "trained artificial intelligence does their job better than themselves."
Ilya Khel
