The Limits of Artificial Intelligence Today
Guest Blogger Introduction
I’m Emily Zhao, a summer intern at Quantellia. I am an undergraduate at the University of California, San Diego, studying Applied Math, Cognitive Science, and Computer Science.
As co-sponsor of the Responsible AI/DI summit, Quantellia has asked me to help write a few blog posts to explain the summit topics from an “outsider’s” perspective. After interviewing Lorien and meeting with Ferose and a few other speakers in Palo Alto this week, I’ve been pulling together my thoughts for this blog series. The articles will be posted at both www.aidisummit.org as well as at www.lorienpratt.com. This second blog in the series answers my question as to why we need a new discipline that goes beyond the current generation of AI systems.
I am thrilled to have been given this opportunity, and I hope that I and those who read these blog posts alike can gain insight into the effects of current AI on society, learn how we can utilize the beneficial potential of new AI/DI systems, and understand various perspectives that we haven’t considered before. I hope you find these valuable!
The Limits of Current AI
Our brains make at least hundreds, perhaps thousands, of decisions every day. Many require context and information processing from countless inputs and outputs, and we make many of them with little conscious effort.
AI decisions are different: this technology excels at performing single tasks and single-link reasoning, but when it comes to making more complicated decisions that require understanding multiple factors, such as the context of the decision, there remain drastic differences in the way that humans and machines process information.
From AI to DI: “multi-link think”
Yann LeCun, chief scientist at Facebook’s Artificial Intelligence Research lab, explains how computers that are good at processing information and making calculations are still lacking in “common sense”:
“We rely, at a minimum, on four interconnected capabilities to successfully navigate the world: (1) to perceive and categorize things around us; (2) to contextualize those things for understanding and learning; (3) to be able to make predictions based on past experience and present circumstances and (4) to make plans based on all of the above.”
Though these abilities are inherent to humans, machines don’t understand cause and effect, aren’t able to contextualize relationships between objects and the environment, and can’t make decisions from both internal and external factors. This is where current AI systems are limited.
Take, for instance, the perspective of a company that sells sprinkler products. An “AI classic” system might predict how much calling a customer on the phone could increase their likelihood of buying sprinklers. A prediction like this is typical of most AI systems today.
But this one question is not enough, because making the decision to call the customer requires multiple predictions: How many calls do we think it will take for the customer to answer the phone? How long will we need to talk take to convince them? How much will they spend on our products over their lifetime — can we use their location to determine that they’re likely to have a bigger house, which means more acres, which equates to more money ultimately spent on sprinklers? How likely is this customer to recommend our product to his/her friends? And will they be costly to support?
Single-link systems are today often useful in producing results for large organizations, such as for Amazon’s recommended items list.
Going from single- to multi-link systems like these give us a better, more big-picture approach to determining the true cost and benefit of calling a sprinkler customer. The answer cannot be found from one direct link of cause and effect, but rather multiple links that come from multiple predictions.
Lorien Pratt, Chief Scientist and Co-Founder at Quantellia and author of the forthcoming Emerald Press book Link: How Decision Intelligence Makes the Invisible Visible, sees an increasingly greater need for multi-link AI to extend the limitations of single-link AI, using the same ideas of contextualization and cause-and-effect reasoning identified by companies like Facebook and increasingly covered in the media, including the important new book called The Book of Why and the “network of causes” / “systemic causation” work of George Lakoff, as described in the video below.
(If you can’t see the video above, then view the web page version of this post)
Adding intangibles to the multi-link picture
However, says Pratt, many companies become stuck on predicting only the tangible, easy-to-quantify characteristics, like dollars. This overlooks long-term factors that are equally, if not more important in the decision-making processes.
“If we step back and we look at the whole field of AI, we’re at this transition point where we see diminishing returns from single-link use cases. So now people are starting to say, okay, how can we go to the next level?”
To incorporate these many factors into a model requires measuring the intangible and expanding the map to a see the bigger picture. Customer satisfaction is an example of an intangible yet highly important factor to consider; a company may not be able to directly measure the likeliness of a customer to recommend a product to a friend, but they can use well-designed surveys that consider user experience in the decision-making process.