There is a lot of fiction circulating regarding the best use of artificial intelligence in human affairs, but many questions, and more importantly, failures, still remain unresolved, even today. A group of machine learning researchers at Stanford University worked together to answer a set of questions that have probably received more research attention than others, but which have remained unanswered.
The real question is, could artificial intelligence overcome these limitations at least on a limited basis? The answer appears to be “yes”: their AI-centric study attempted to see how humans and AI programs performed on different aspects of solving a Rubik’s Cube.
In their paper, they proposed, on the one hand, that AI would indeed be a better choice for solving the Rubik’s Cube. One would probably think that a Rubik’s Cube is kind of straightforward for humans as well as AI, but in a confusing and exhaustingly difficult problem, humans, indeed, were proven better than AI.
However, even when we assume that the success rate of AI in solving a Rubik’s Cube was 99 percent, we still find that humans still outperformed AI, something that they again confirmed in another study published by Cassilis recently.
Real and simulated difficulty
Let’s break it down: when humans analyzed the cube in the real world, they could easily see the cube and manipulate it – they were experienced Rubik’s Cube users. For the same study, though, they simulated 10,000 different Rubik’s Cube experiences, with 10 models of AI used.
The real human results stood out as obviously more complex than those recorded using the simulated AI. The tests revealed that the 10,000 Rubik’s Cube experiences not only required a higher level of cognitive ability, it involved more “social cues,” which translates to more “mental modeling” when resolving a problem. For an example of their skill level, computers performed better than humans in finding new angles. They were considered just as good at finding the best possible option as humans.
In another type of study the authors made up a set of 1,004 Rubik’s Cube configurations, and put them to the test using real human players. The results were again shown as being easier for human players than for the artificially intelligent versions.
Problems with performance on the target were found to be an area of strong research that could benefit from both AI and human bias. Yet another example is the length of matches, which, surprisingly, was harder to program than human-versus-human bouts – a fact that researchers were completely surprised about.
The authors began their research in order to explore the problems faced by human technology. They began with the fact that humans have a tremendous amount of performance biases. Other studies have been indicating such biases for years. So, they worked on building their own “learning algorithm,” designed for unsupervised perception and perception of decoders.
Does it still matter for AI?
These researchers were clear about their intention to bring real artificial intelligence into the Rubik’s Cube problem. In the real world, a user doesn’t actually solve a Rubik’s Cube; they simply find the solution and store it in a computer, and if it’s just a part of a sample of a particular puzzle, or not much different from others, users would never create their own solutions.
In fact, by just looking at this bit of data, we can see that humans tend to show the kind of cognition and processing power that machines don’t have today. However, humans in our study still favored human-training AI as a solution, since they did not believe that their capability can be increased much at the individual level, and AI is not able to perform tasks currently limited to humans.
Hopefully, a change is not too far off, and then we will be able to use the knowledge of humans, and also the light of ethics, to bring about the kinds of powerful artificial intelligence solutions that we need. Until then, we will just have to be thankful that we have a keyboard and mouse, and a Rubik’s Cube!
Duff is a freelance science writer and storyteller. When he’s not exploring the mysteries of artificial intelligence, he writes about mind games, snake oil salesmen, seances, “grandpa-to-grandma” child reunions, the predictive personality traits of active geriatrics, and the almost-real nature of DD Tiles.