Learning How AI Works from AlphaGo

November, 8 2018

Subscribe and stay up to date

No spam, we promise! You will only 
receive essential emails.
Written By
Alyssa Verzino


When the average consumer sees headlines like DeepMind's AlphaGo AI becoming the most formidable Go player in the world, they often wrongly conclude that all artificial intelligence solutions are as dominant in their respective fields, and that they achieve that dominance almost immediately.

In reality, AlphaGo is a great example of how unintelligent AI usually is "out of the box" and the amount of effort it takes to realize the staggering potential of AI systems.

AlphaGo is a deep learning AI system, which means it effectively programs itself by ingesting and analyzing training data. More simply, AlphaGo learned how to play Go by watching other games of Go, rather than being formally programmed with a Go rule-set and Go strategies. And while the development of the AI software that was capable of this type of learning took years, the actual training of AlphaGo took months more.

Just like championship athletes train for months and years prior to becoming champions, AlphaGo – and all deep learning AIs – need significant prep time to learn how to perform at an elite level.

When you purchase an AI solution with the potential to enhance your workflow, you're typically buying the version of AlphaGo that had seen only enough games of Go to understand the rules but not to win against any competent opponents. AlphaGo went through two stages of training to develop its unique Go skills.

AlphaGo began by analyzing records of past human games. Go has the advantage of natively producing annotated training data, which is to say data that is already marked as a success or as a failure. Go games have winners and losers, and each turn of the game is strictly documented, giving a deep learning algorithm a rich set of data to work with to train itself.

When you purchase a commercial AI, it has typically been trained against a proprietary data set, which gives it enough competence to "play the game" as soon as you deploy it but it likely won't start out any more effective that a typical human employee.

To achieve excellence, AlphaGo had to move past static data and play games against live human opponents. AlphaGo's developers disguised it as a human player in online Go competitions and let the AI improve itself by playing against increasingly skilled Go opponents.

This gave AlphaGo a sense of how an opponent reacts to any of its strategies and counter-adapts. By playing against humans, AlphaGo learned how to apply it skills in "the real world" and thus developed novel Go strategies that made it a grand champion.

Your own AI solutions likely won't deliver truly significant advantages until they've spent time in deployment, adapting to the realities of your business environment. Like AlphaGo, they'll be competent at the start, but will take time to become a superior solution.

For it's part, AlphaGo has already been replaced by AlphaGo Zero, a new AI designed to play games of Go against itself. Like a Generative Adversarial Network (GAN), AlphaGo Zero it trying to jump-start excellence by pitting AIs against AIs, rather than waiting out the long process of training a competent AI against human counterparts.

If AlphaGo Zero succeeds we may finally unlock the key to out-of-box excellence from AI solutions. Until then, we'll need to temper expectations for initial AI performance and give our artificial intelligence agents time to achieve AlphaGo levels of business success. And there's no time to wait; the sooner you adopt an AI solution, the faster it will get up to speed and start delivering AlphaGo-level results.

If you'd like to learn how these AI principles can improve your customer support, check out Talla's AI -enhanced support agent and knowledge base. If you'd like to get started on your own AI learning curve, contact Talla today.

New call-to-action

View all posts

Recommended Reading

Subscribe and stay up to date

No spam, we promise! You will only 
receive essential emails.

Subscribe and stay up to date