Just read “Defeated by A.I., a Legend in the Board Game Go Warns: Get Ready for What’s Next”
https://www.nytimes.com/2024/07/10/world/asia/lee-saedol-go-ai.html?unlocked_article_code=1.6U0.j201.K0_Ul-SfrC0n&smid=url-share

As someone who did their PhD in game-tree search, I’ve watched in awe as AI has evolved from mastering board games to tackling some of humanity’s most pressing challenges. The journey from early game-playing algorithms to today’s multifaceted AI systems is nothing short of extraordinary.

Software Engineer at Google by day, Web3 degen by night

In fact, the impact of MCTS on the field was so profound that it accelerated my own academic journey. I was deep into my PhD research on forward pruning and selective search in game-tree search when MCTS emerged. It quickly became clear that complete game-tree search might not be the future of game AI. This realization prompted me to swiftly wrap up my research into a PhD thesis.

The last paragraph of my thesis (2007) reflects this pivotal moment:

“The game of Go can be considered as the grand challenge of game AI at this point in time. One interesting development in computer Go has been the introduction of Monte Carlo methods that combine game-tree search and randomly generated moves for evaluation [Coulom, 2006, Kocsis and Szepesvári, 2006]. The random nature of Monte Carlo methods corresponds well with the theoretical analysis of the properties of forward pruning presented in this thesis, and should extend to Monte Carlo tree search. More research on how to incorporate risk management strategies in forward pruning can be done to further improve the state of the art for Monte Carlo tree search.”

Looking back, it’s remarkable how accurate this assessment was. MCTS indeed became a cornerstone in advanced game AI, particularly in conquering the game of Go.

But the true revolution came with the integration of deep learning. Neural networks, once considered a relic of AI’s past, roared back to life. They brought with them the ability to learn complex patterns and strategies from raw data. This culminated in the creation of AlphaGo, a system that shocked the world by defeating Lee Saedol, one of the greatest Go players of our time.

Yet, AlphaGo was just the beginning. AlphaGo Zero and then AlphaZero demonstrated that AI could learn superhuman strategies without any human knowledge, purely through self-play. The implications were staggering.

Beyond the Game Board
What truly excites me is how these game-playing techniques have found applications far beyond the world of board games. Let me share some cutting-edge examples:

Weather Forecasting
The advancements in game AI have found significant applications beyond the game board, one of which is weather forecasting. For example, MetNet-3 leverages techniques reminiscent of those used in game AI to tackle the complexities of atmospheric dynamics. By integrating the data assimilation and simulation processes into a single neural network pass, MetNet-3 can provide high-resolution weather predictions up to 24 hours ahead, covering variables such as precipitation, temperature, wind, and dew point. This model outperforms traditional physics-based numerical weather prediction methods, achieving temporal resolutions as fine as two minutes and spatial resolutions of one to four kilometers.

Drug Discovery
In the realm of drug discovery, AI techniques refined in game-playing applications are accelerating research and development. Insilico Medicine’s AI-generated drug for idiopathic pulmonary fibrosis (IPF) is a prime example. Their AI system identified a novel therapeutic target (TNIK) and designed a small molecule (INS018_055) for IPF treatment in just 18 months, a process that traditionally takes years. This AI-driven approach mirrors game AI’s method of identifying key strategic points and developing optimal strategies to exploit them. The rapid progression of this AI-designed drug, now in Phase 2 clinical trials, underscores the potential of AI in pharmaceutical research.

Molecular Biology
In molecular biology, the impact of AI is profoundly illustrated by AlphaFold 3, which extends the capabilities of its predecessor by predicting the structure and interactions of a wide range of biological molecules, including proteins, DNA, RNA, and small molecules. This progression from predicting protein structures to modeling complex molecular interactions showcases the significant leap in AI’s application in biology. AlphaFold 3’s enhanced architecture, featuring an improved Evoformer module and a diffusion network, has achieved a substantial improvement in the accuracy of predicting protein interactions, further advancing our understanding of cellular biology.

The Common Thread
What fascinates me most is the common thread running through all these applications – the ability of AI to learn complex patterns and make decisions in high-dimensional, uncertain environments. Whether it’s navigating the vast possibility space of Go, predicting chaotic weather systems, or exploring the intricate world of molecular structures, AI is proving to be a powerful tool.

Looking Ahead
As we stand at this exciting juncture, I can’t help but wonder: what’s next? Which field will see the next AlphaZero-like breakthrough? How might these AI techniques reshape your industry?

One thing is certain – we’re not just spectators in this AI renaissance. We’re potential collaborators, innovators, and pioneers. The games AI has learned to play have prepared it for the greatest game of all – advancing human knowledge and capabilities across all fields of endeavor.

I’m eager to hear your thoughts. How do you see AI impacting your field in the coming years?

The Game-Changing Progression
The evolution of game AI tells a fascinating story of human ingenuity. We started with simple minimax algorithms and alpha-beta pruning, techniques that seemed cutting-edge at the time. Then came knowledge-based systems, leveraging human expertise to improve performance.

A significant leap during my PhD was the advent of Monte Carlo methods, particularly Monte Carlo Tree Search (MCTS). This probabilistic approach opened new horizons, allowing AI to handle the vast complexity of games like Go.