Foresight’s mission is essentially an educational one. In simplest terms we are here to point out foreseeable technological developments that not only will make the future different from the past, but make it different in ways that aren’t obvious and which everyone isn’t already planning for. Nanotechnology — true nanotech in Drexler’s original sense of having a thorough control over the structure of matter at the atomic scale and thus being able to build productive machinery — is such a development, even though the word “nanotechnology” is widely used for much more mundane, predictable, linear, and non-revolutionary progress.
Similarly, the term “Artificial Intelligence” is widely used for predictable, linear progress in software engineering. The field has come a long way, so that it is getting close to the point that any well-specified human skill, such as driving a car, can be implemented given an appropriate application of talent and resources. Just like “nanotechnology,” though, it originally meant something more revolutionary:
Some years ago, Ben Goertzel coined the term “AGI” — artificial general intelligence — to distinguish the original, revolutionary goal of AI as originally seen by such pioneers as McCarthy and Minsky, from the more mundane, incremental work that the term AI had come to cover. This was very similar in spirit to the term MNT — molecular nanotechnology — coined by Drexler and Foresight for essentially the same reason.
Within the past couple of years, the Productive Nanosystems Roadmap was organized and published, under the names of a wide sampling of people from academia, industry, and the national laboratories. This had the effect of making it clear that the ultimate goal of nanotechnology research is indeed “MNT”-style capabilities, and is one that is ultimately feasible and worth working toward.
While the “diaspora” in AI may have been deeper than the one in nanotech, it was also longer ago — there was no need for the AGI Roadmap to re-establish the possibility of an artificial intelligence in the full sense, but to try and make some sense of the state of the art with respect to it, figure out some milestones and metrics that might be used to judge progress, and so forth.
The meeting last weekend at the University of Tennessee, organized by Ben Goertzel and Itamar Arel, served to bootstrap the process and begin to work out what kind of roadmap might be possible. The main problem, of course, is that we don’t really know how intelligence works, which pieces are essential and which ancillary, or indeed whether there are a few powerful underlying principles or a huge kludge of random techniques.
To that end we began by trying to define the kind of tasks that we felt a general intelligence could do but that no hand-coded “narrow AI” could do. The classic such task, or course, is the Turing Test, which has many points in its favor but is also considered (a) too high a bar, and (b) a test of the wrong thing, since it requires fooling a judge as well as exhibiting basic intelligence.
To give some of the flavor of the scenarios, here’s the one I proposed:
The Wozniak Test
In an interview a few years ago, Steve Wozniak of Apple fame opined that there would never be a robot that could walk into an unfamiliar house and make a cup of coffee. I feel that the task is demanding enough to stand as a pons asinorum for embodied AGI.
A robot is placed at the door of a typical house or apartment. It must find a doorbell or knocker, or simply knock on the door. When the door is answered, it must explain itself to the householder and enter once it has been invited in. (We will assume that the householder has agreed to allow the test in her house, but is otherwise completely unconnected with the team doing the experiment, and indeed has no special knowledge of AI or robotics at all.) The robot must enter the house, find the kitchen, locate coffee-making supplies and equipment, make coffee to the householder’s taste, and serve it in some other room. It is allowed, indeed required by some of the specifics, for the robot to ask questions of the householder, but it may not be physically assisted in any way.
The state of the robotics art falls short of this capability in a number of ways. The robot will need to use vision to navigate, identify objects, possibly identify gestures (“the coffee’s in that cabinet over there”), and to coordinate complex manipulations. Manipulation and physical modelling in a tight feedback learning loop may be necessary, for example, to pour coffee from an unfamiliar pot into an unfamiliar cup. Speech recognition and natural language understanding and generation will be necessary. Planning must be done at a host of levels ranging from manipulator paths to coffee-brewing sequences.
But the major advance for a coffee-making robot is that all of these capabilities must be coordinated and used appropriately and coherently in aid of the overall goal. The usual set-up, task definition, and so forth are gone from standard narrow AI formulations of problems in all these areas; the robot has to find the problems as well as to solve them. That makes coffee-making a strenuous test of a system’s adaptiveness and ability to deploy common sense.
I claim that this test addresses the bulk of the aspects of general intelligence that are missing from AI today. Although standard shortcuts might be used, such as having a database of every manufactured coffeemaker built in, it would be prohibitive to have the actual manipulation sequences for each one pre-programmed, especially given the variability in workspace geometry, dispensers and containers of coffee grounds, and so forth. Transfer learning, generalization, reasoning by analogy, and in particular learning from example and practice are almost certain to be necessary for the system to be practical.
Coffee-making is a good test of generality because, although it would be possible to hand-code most of the skills needed, it would be much cheaper simply to build a coffeemaker into the robot! Thus the only economical way to approach the task is to build general learning skills and have a robot that is capable of learning not only to make coffee but any similar domestic chore.
Coffee-making is a task that most 10-year-old humans can do reliably with a modicum of experience. I would guess that a week’s worth of being shown and practicing coffeemaking in a variety of homes with a variety of methods would provide the grounding for enough generality that a 10-year-old could make coffee in the vast majority of homes in a Wozniak test.