[Tech Talk] Can AI Beat Einstein?

The Pillar of Theoretical Science 'Thought Experiment'
Attempted Reproduction Through AI Simulator

Human science and philosophy have advanced through 'thought experiments.' It is a method of inferring the essence of objects or phenomena through hypothetical scenarios. The most famous master of thought experiments is undoubtedly Albert Einstein. He derived the theory of relativity based on the assumption, "Can a person running at the speed of light see their own face in a hand mirror?"

The master of theoretical physics who established the general/special theory of relativity, Einstein [Image source=Rarehistoricalphoto Archive]

Thought experiments remain the core of all theoretical sciences, but now scientists employ computers. Unlike in the past, the process of analyzing the essence of the universe and matter has become exponentially more complex. But could computers go beyond being simple 'calculators' and conduct thought experiments themselves? Recent advances in artificial intelligence (AI) have set the first milestone for a 'machine that performs thought experiments.'

Video Generation AI 'SORA'... Aiming to be a World Simulator

What if video generation AI could truly predict wave flows identical to reality?
[Image source=OpenAI]

Recently, AI company OpenAI unveiled SORA, a generative AI that converts text into 60-second videos. While video generation AI technology itself is not new, SORA attracted attention for generating longer videos with much more stable quality than before.

However, SORA's true value is not in simple video generation. In fact, according to a report released by OpenAI, SORA was created with the goal of being a "video generation tool for world simulation." However, it still cannot fully understand the complex physical phenomena occurring in the real world.

The world we live in operates through countless forces. From microscopic gravity and nuclear forces to macroscopic perspectives including atmospheric and water flows. For OpenAI to realize SORA as a true 'world simulator,' the AI must understand and 'predict' the effects of all these forces.

The Core of World Simulation, Computational Fluid Dynamics (CFD)

Computational Fluid Dynamics (CFD) can simulate the effects of air, liquids, and heat on objects. Accordingly, it is an essential technology in modern precision mechanical engineering. The photo shows a CFD simulation of a SpaceX rocket. [Image source=YouTube]

In fact, efforts to create physical phenomenon simulators using computers have been ongoing for some time. Today, many industries rely on computer-based physical simulators. The most representative simulator is Computational Fluid Dynamics (CFD). It is a technology that implements the flow of liquids or gases in a virtual world to analyze their effects on specific objects.

CFD has had a particularly significant impact on the aerospace industry. When airplanes or rockets cut through the atmosphere and take off, calculating the air resistance, heat, and vibrations affecting the body manually requires enormous work. To build an environment similar to the air on the ground, huge wind tunnel test facilities must also be constructed, incurring tremendous costs. CFD automated these 'thought experiments' using computers. Meanwhile, it also greatly shortened the design time of complex gas turbines used in aircraft engines.

Engine wind tunnel testing facility of Rolls-Royce, one of the world's top three aircraft engine manufacturers. [Image source=Rolls-Royce]

However, current CFD is extremely limited and fragmented. It still cannot completely replace 'real-world testing,' and the types of fluids that can be simulated at one time are limited.

Genius-Level AI Thought Experiments, Still a Long Way Off

The ambition of AI companies like OpenAI is to complete a 'universal CFD.' This would be an AI capable of simultaneously predicting all kinds of dynamics?from atmospheric flows to rough waves, and the responses of solids exposed to heat and vibrations. If AI's CFD reaches this level, it could be called a 'world thought experimenter.' It would be the moment when computers surpass Einstein's intelligence.

Although OpenAI's SORA is in the spotlight, the leader in AI-based CFD is actually Google DeepMind. Last November, they introduced a weather prediction AI called 'GraphCast,' which attracted industry attention. Weather prediction is a highly advanced thought experiment that assumes numerous dynamic variables such as droplets, wind, and temperature in the atmosphere.

DeepMind's weather forecasting program 'GraphCast,' released last year, is also an example of a 'world simulator' that combines CFD and AI. [Image source=DeepMind]

DeepMind created GraphCast using a model called Graph Neural Network (GNN). GNN is a data structure specialized in inferring how distributed data points are related to each other, making it suitable for thought experiments that must capture flows from countless data points. In fact, GraphCast reportedly showed much higher accuracy than most meteorological weather prediction algorithms.

However, despite some achievements, AI still cannot truly replicate the world. Even OpenAI admitted that SORA's main weakness is "struggling to accurately implement physics in complex scenes and failing to understand the causes and effects of specific physical phenomena."

One day, an AI capable of performing truly sophisticated simulations may emerge to unravel the secrets of matter and the universe on behalf of humans, but for now, thought experiments remain a uniquely human talent.