What if artificial intelligence, as we understand it today, is based on a decade-old misunderstanding about what intelligence is? And what if some of the problems we’re dealing with can be traced back to that misunderstanding? Computer scientist Stuart Russell, co-author of the most popular AI textbook of our time, explored these questions in a lecture organized by the Alan Turing Institute last week.
When pioneering work on artificial intelligence began in the 1940s, our understanding of intelligence was connected to the ability to achieve objectives. Scientists translated that notion to the emerging field of AI, encoding goals into machines.
To this day, a lot of machine-learning technology is based on that principle: Engineers and computer scientists define an objective for the computer. Then they essentially sit back and watch the machine learn and teach itself how to get there.
This standard model is actually flawed, Russell says, and this has led to some devastating consequences. It’s easy to see what’s wrong with the current state of the AI by looking at the algorithms responsible for selecting content on YouTube. The objective coded into them is to maximize the probability that users want to click on the next video they’re shown. But the algorithm, instead of really “learning” what people want to see, is modifying people to be more predictable.
This, of course, also has consequences for legislation as well: Russell has proposed codes of practice and standards for defining objectives and identifying actions that can be taken by systems. The goal is that if a system interacts with users, it is ensured that it has no incentive to manipulate or modify the user's beliefs and preferences.
Not least because of the limitations described in his lecture, most of today’s AI systems will hit a wall in terms of what they’re able to do, Russell says. He recommends a different approach to AI development called probabilistic programming which combines probability theory with lessons from the world of programming languages or logic. “In the future what we’ll see is probably some kind of merger between probabilistic programming and deep learning and this will fuel the next phase of growth in AI,” he predicts.
Turing Lecture: Provably beneficial AI
Author: Stuart Russell, Janosch Delcker