Recently there has been a spate of articles in the mainstream press, and a spate of high profile people who are in tech but not AI, speculating about the dangers of malevolent AI being developed, and how we should be worried about that possibility. I say relax. Chill. This all comes from some fundamental misunderstandings of the nature of the undeniable progress that is being made in AI, and from a misunderstanding of how far we really are from having volitional or intentional artificially intelligent beings, whether they be deeply benevolent or malevolent.
By the way, this is not a new fear, and we’ve seen it played out in movies for a long time, from “2001: A Space Odyssey”, in 1968, “Colossus: The Forbin Project” in 1970, through many others, and then “I, Robot” in 2004. In all cases a computer decided that humans couldn’t be trusted to run things and started murdering them. The computer knew better than the people who built them, so it started killing them. (Fortunately that doesn’t happen with most teenagers, who always know better than the parents who built them.)
I think it is a mistake to be worrying about us developing malevolent AI anytime in the next few hundred years. I think the worry stems from a fundamental error in not distinguishing the difference between the very real recent advances in a particular aspect of AI, and the enormity and complexity of building sentient volitional intelligence. Recent advances in deep machine learning let us teach our machines things like how to distinguish classes of inputs and to fit curves to time data. This lets our machines “know” whether an image is that of a cat or not, or to “know” what is about to fail as the temperature increases in a particular sensor inside a jet engine. But this is only part of being intelligent, and Moore’s Law applied to this very real technical advance will not by itself bring about human level or super human level intelligence. While deep learning may come up with a category of things appearing in videos that correlates with cats, it doesn’t help very much at all in “knowing” what catness is, as distinct from dogness, nor that those concepts are much more similar to each other than to salamanderness. And deep learning does not help in giving a machine “intent”, or any overarching goals or “wants”. And it doesn’t help a machine explain how it is that it “knows” something, or what the implications of the knowledge are, or when that knowledge might be applicable, or counterfactually what would be the consequences of that knowledge being false. Malevolent AI would need all these capabilities, and then some. Both an intent to do something and an understanding of human goals, motivations, and behaviors would be keys to being evil towards humans.
Michael Jordan, of UC Berkeley, was recently interviewed in IEEE Spectrum, where he said some very reasonable, but somewhat dry, academic, things about big data. He very clearly and carefully laid out why even within the limited domain of machine learning, just one aspect of intelligence, there are pitfalls as we don’t yet have solid science on understanding exactly when and what classifications are accurate. And he very politely throws cold water on claims of near term full brain emulation and talks about us being decades or centuries from fully understanding the deep principles of the brain.
The Roomba, the floor cleaning robot from my previous company, iRobot, is perhaps the robot with the most volition and intention of any robots out there in the world. Most others are working in completely repetitive environments, or have a human operator providing the second by second volition for what they should do next.
When a Roomba has been scheduled to come out on a daily or weekly basis it operates as an autonomous machine (except that all models still require a person to empty their bin). It comes out and cleans the floor on its schedule. The house might have its furniture re-arranged since last time, but the Roomba finds its way around, slowing down when it gets close to obstacles, it senses them before contact, and then heading away from them, and it detects drops in the floor, such as from a step or stair with triply redundant methods and avoids falling down. Furthermore it has a rudimentary understanding of dirt. When its acoustic sensors in its suction system hear dirt banging around in the air flow, it stops exploring and circles in that area over and over again until the dirt is gone, or at least until the banging around drops below a pre-defined threshold.
But the Roomba does not connect its sense of understanding to the bigger world. It doesn’t know that humans exist–if it is about to run into one it makes no distinction between a human and any other obstacle; by contrast dogs and even sheep understand the special category of humans and have some expectations about them when they detect them. The Roomba does not. And it certainly has no understanding that humans are related to the dirt that triggers its acoustic sensor, nor that its real mission is to clean the houses of those humans. It doesn’t know that houses exist.
At Rethink Robotics our robot Baxter is a little less intentional than a Roomba, but more dexterous and more aware of people. A person trains Baxter to do a task, and then that is what Baxter keeps doing, over and over. But it “knows” a little bit about the world with just a little common sense. For instance it knows that if it is moving its arm towards a box to place a part there and for whatever reason there is no longer something in its hand then there is no point continuing the motion. And it knows what forces it should feel on its arms as it moves them and is able to react if the forces are different. It uses that awareness to seat parts in fixtures, and it is aware when it has collided with a person and knows that it should immediately stop forward motion and back off. But it doesn’t have any semantic connection between a person who is in its way, and a person who trains it–they don’t share the same category in its very limited ontology.
OK, so what about connecting an IBM Watson like understanding of the world to a Roomba or a Baxter? No one is really trying as the technical difficulties are enormous, poorly understood, and the benefits are not yet known. There is some good work happening on “cloud robotics”, connecting the semantic knowledge learned by many robots into a common shared representation. This means that anything that is learned is quickly shared and becomes useful to all, but while it provides larger data sets for machine learning it does not lead directly to connecting to the other parts of intelligence beyond machine learning.
It is not like this lack of connection is a new problem. We’ve known about it for decades, and it has long been referred to as the symbol grounding problem. We just haven’t made much progress on it, and really there has not been much application demand for it.
Doug Lenat has been working on his Cyc project for twenty years. He and his team have been collecting millions, really, of carefully crafted logical sentences to describe the world, to describe how concepts in the world are connected, and to provide an encoding of common sense knowledge that all of us humans pick up during our childhoods. While it has been a heroic effort it has not led to an AI system being able to master even a simple understanding of the world. Trying to scale up collection of detailed knowledge a few years ago Pushpinder Singh, at MIT, decided to try to use the wisdom of the crowds and set up the Open Mind Common Sense web site, which involved a number of interfaces that ordinary people could use to contribute common sense knowledge. The interfaces ranged from typing in simple declarative sentences in plain English, to categorizing shapes of objects. Push developed ways for the system to automatically mine millions of relationships from this raw data. The knowledge represented by both Cyc and Open Mind has been very useful for many research projects but researchers are still struggling to use it in game changing ways by AI systems.
Why so many years? As a comparison, consider that we have had winged flying machines for well over 100 years. But it is only very recently that people like Russ Tedrake at MIT CSAIL have been able to get them to land on a branch, something that is done by a bird somewhere in the world at least every microsecond. Was it just Moore’s law that allowed this to start happening? Not really. It was figuring out the equations and the problems and the regimes of stall, etc., through mathematical understanding of the equations. Moore’s law has helped with MATLAB and other tools, but it has not simply been a matter of pouring more computation onto flying and having it magically transform. And it has taken a long, long time.
Expecting more computation to just magically get to intentional intelligences, who understand the world is similarly unlikely. And, there is a further category error that we may be making here. That is the intellectual shortcut that says computation and brains are the same thing. Maybe, but perhaps not.
In the 1930’s Turing was inspired by how “human computers”, the people who did computations for physicists and ballistics experts alike, followed simple sets of rules while calculating to produce the first models of abstract computation. In the 1940’s McCullough and Pitts at MIT used what was known about neurons and their axons and dendrites to come up with models of how computation could be implemented in hardware, with very, very abstract models of those neurons. Brains were the metaphors used to figure out how to do computation. Over the last 65 years those models have now gotten flipped around and people use computers as the metaphor for brains. So much so that enormous resources are being devoted to “whole brain simulations”. I say show me a simulation of the brain of a simple worm that produces all its behaviors, and then I might start to believe that jumping to the big kahuna of simulating the cerebral cortex of a human has any chance at all of being successful in the next 50 years. And then only if we are extremely lucky.
In order for there to be a successful volitional AI, especially one that could be successfully malevolent, it would need a direct understanding of the world, it would need to have the dexterous hands and/or other tools that could out manipulate people, and to have a deep understanding of humans in order to outwit them. Each of these requires much harder innovations than a winged vehicle landing on a tree branch. It is going to take a lot of deep thought and hard work from thousands of scientists and engineers. And, most likely, centuries.
The science is in and accepted on the world being round, evolution, climate change, and on the safety of vaccinations. The science on AI has hardly yet been started, and even its time scale is completely an open question.
Just how open the question of time scale for when we will have human level AI is highlighted by a recent report by Stuart Armstrong and Kaj Sotala, of the Machine Intelligence Research Institute, an organization that itself has researchers worrying about evil AI. But in this more sober report, the authors analyze 95 predictions made between 1950 and the present on when human level AI will come about. They show that there is no difference between predictions made by experts and non-experts. And they also show that over that 60 year time frame there is a strong bias towards predicting the arrival of human level AI as between 15 and 25 years from the time the prediction was made. To me that says that no one knows, they just guess, and historically so far most predictions have been outright wrong!
I say relax everybody. If we are spectacularly lucky we’ll have AI over the next thirty years with the intentionality of a lizard, and robots using that AI will be useful tools. And they probably won’t really be aware of us in any serious way. Worrying about AI that will be intentionally evil to us is pure fear mongering. And an immense waste of time.
Let’s get on with inventing better and smarter AI. It is going to take a long time, but there will be rewards at every step along the way. Robots will become abundant in our homes, stores, farms, offices, hospitals, and all our work places. Like our current day hand-held devices we won’t know how we lived without them.