What’s So Special About AI?

There is no doubt that in the past few years AI has radically expanded computer capabilities. Adoption of AI by businesses and governments is accelerating, but we are nowhere near saturation. It’s the Wild West out there in terms of applications and business models, with lots of unexplored country remaining.

But what is so different about AI compared to traditional computer programming approaches?

Unrealistically high or scary expectations

First, one major difference about AI is expectations. Anything labeled “Artificial Intelligence” comes with a lot of baggage. Expectations range paradoxically from unrealistically high to ridiculously scary. There is much hype and fear swirling around the technology in the news media and science fiction.

But alarms and praise has been raised by very smart, knowledgeable people as well. Depending on who you ask, it will either bring the end of humanity or solve all of our problems. According to Elon Musk, “[AI] is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five-year timeframe, 10 years at most.” Even more alarming, famous physicist Stephen Hawking said that “The development of full artificial intelligence could spell the end of the human race” On the other hand, the CEO of Google, Sundar Pichai, says AI will do more for humanity than “fire or electricity.” Tech visionary Peter Diamondis claims AI is “a massive opportunity for humanity, not a threat.”

Most of these comments are related to general AI, which is probably decades out, but these fears and expectations will affect anything labeled as Artificial Intelligence.

Computing scale matters

Another way AI is different is that scale really matters. By scale, I mean processing speed and memory capacity.

IBM’s Deep Blue beat chessmaster Garry Kasparov in 1996. Kasparov would later say that he felt he was representing all of humanity in the competition. But in reality, it was simply a matter of time before computers beat humans in chess. Building a computer to win at chess was a difficult problem, but the possible moves of the game are not infinite. Deep Blue simply used brute force (in terms of processing capacity and memory) to calculate every possible move and counter move.

Because of the mathematical intensity of many AI techniques, especially artificial neural networks, the computational power of computers really matters. Big Blue was a large computer system by the standards of the day. The continuation of Moore's Law over the past 30 years has only increased the raw capability of AI even more.

Alchemy of data science

Another way AI is different, is that the successful application of AI is often a complex alchemy of approaches. Progress towards a goal is lumpy and hard to predict. Different problems require different approaches and algorithms determined through a trial and error process.

20 years after IBM computers beat the reigning champion of chess, DeepMind’s AlphaGo won at the game of Go in 2016. Unlike chess, the number of possible Go games is extremely large. It is often compared to the number of atoms in the universe, but it is in fact much much larger. So a brute force method here will not work. AlphaGo was successful by using a novel combination of algorithms, including a Monte Carlo tree search and an artificial neural network. The choice and configuration of these algorithms made all the difference.

Data-driven learning systems

Data is the fuel for AI learning machines. This is different from traditional computer programming which simply uses data, but does not learn from it.

At a simple level, programming a computer is giving it instructions about how to process inputs to generate desired outputs. If you know the relationship between inputs and outputs, say converting celsius to fahrenheit temperatures, then a traditional approach to programming works great.

The difference with AI systems, is that we are trying to solve a problem where we don’t know the exact relationship, say the impact of company financials and economic factors on share price. We can guess at the list of potential input variables, but we don’t know how much they should be weighted or combined to predict an output. This is why we need the data to train the AI system and develop a model for this relationship between inputs and outputs. Generally, the more training data the better the AI model, up to a point.

As you may imagine the importance of data may drive a number of strategic business, policy and ethical questions. Who owns the data, and was it intended for this purpose? How can startups compete for AI applications if the big tech companies have all the data? When does it make sense to share data between companies? How do we ensure security and privacy of data being used by AI? Is data access a source of strategic advantage? We’ll discuss these issues in subsequent articles, but suffice to say that access to training data is an essential and distinguishing characteristic of AI systems.

Uncertainty and unpredictability

The fifth and last way AI is different from traditional computing is that the systems tend to lack certainty and predictability.

Most of the techniques of AI deal in probilities. This is driven in part by the complex nature of the problems being solved. For example, as in the real world, diagnosing a disease is not a binary yes or no, it’s a probability based on lots of data, somewhere between zero and 100%.

Dealing with uncertainties is a distinctive advantage of AI. As mentioned above, we deploy AI learning systems because we don’t know the exact relationship between inputs and outputs nor the optimal strategy for achieving a goal. This means the algorithm may and should take novel approaches not imagined by the programmers. This is sometimes called emergent behavior, and usually is helpful and in some cases remarkable. One example of emergent behavior also comes from Google’s DeepMind. When DeepMind used reinforcement learning to teach their AI to play the Atari video game Breakout, it quickly developed successful strategies for winning the game without anyone expressly teaching them.

Also, some AI techniques do not make it easy to explain the reasoning behind their result. This lack of explainability is particularly challenging in regulated or public policy areas, where it is essential that we understand the basis of decisions. The European Union recently passed a law requiring that decisions made by a machine must be readily explainable, which may limit what kinds of AI techniques could be used.

And finally, many AI models are non-deterministic. In computer science, a nondeterministic algorithm is one that, even for the same input, can exhibit different behaviors on different runs.