Strange Loop

A Funny Thing Happened On The Way To Reimplementing AlphaGo

In 2016, AlphaGo defeated the reigning Go world champion, Lee Sedol. As humanity sighs in resignation of its place to its robot betters, much work is done to understand why AlphaGo won. I was more upset that AlphaGo wasn't implemented in Go (the language), and so early this year, having a bit of spare time, some friends and I set out to reimplement AlphaZero, the successor to AlphaGo in Go (the language) to play and win at Go (the board game). Along the way, I learned a few new things about AI, deep neural networks and a few new tricks.

In this talk, I will explain briefly the AlphaZero algorithm in simple, understandable English. Due to working under constrained scenarios, we developed some novel techniques that were used to accelerate the training of our reimplementation of AlphaZero. This talk will recount some of them which may be useful to both programmers in general and AI developers.

Along the way, I will explore the implications of our experiments and what it means for the goal of building an artificial general intelligence. If an AI can self-modify its own code, what does it mean for us puny humans? Do we have any future on this planet?

Xuanyi Chew

Xuanyi Chew


Xuanyi is the Chief Data Scientist at Ordermentum. He is an avid Go programmer, and is the primary author of Gorgonia, a deep learning package for Go which he thinks of it as the bastard child of PyTorch and Tensorflow though it predates both slightly. He thinks the most realistic depiction of AGI so far is The Machine and Samaritan from Person of Interest and has been working towards an AGI since he was a child. In his spare time, he enjoys building new programming languages and alchemy (deep learning and cooking are no different from transmuation ;P)