Supermodelling

One question that always nags at me is what exactly should an AGI do? I think others are puzzled by my confusion on this subject. Surely the answer is everything! That’s not very helpful in a practical way though. So in an earlier post I tried to sketch out what I thought an AGI should be built to do. Even so, it is too much for me to wrap my mind around.

How can progress toward AGI be measured? It’s tricky: defining a task on which performance can be measured causes a focus on something specific when generality is what’s needed. Even the famous Turing Test (in its current laughingstock incarnation, the Loebner Prize) falls prey to this… clever trickery is so far more effective than generality.

After much thought, I have started to focus my thinking on two different tasks. Both of them are inspired from early work in AI and more recently by J Storrs Hall’s excellent book Beyond AI. Both tasks seem like reasonable steps along a pathway toward the previously described end goal for an AGI; both encourage a certain degree of generality and are open-ended enough to have no simple solution. Both lead to interesting demos even at early stages of progress, and both are immediately understandable by other researchers.

Task 1: Super-SHRDLU. Given an environment with objects in it and an AI-controlled manipulator, the goal is to learn how to interact with and model that environment, and discuss it with a human partner. This task will go beyond Winograd’s original SHRDLU (one of the most impressive early AI projects) in the following ways:

• The environment is more complex than Blocks World. The AI interface to the environment will be via normal sensors (cameras, maybe audio, haptic and proprioceptive sensors on the actuator). I envision a first environment being a tabletop with various objects… although blocks may be used, I also imagine things like towels, keys, stuffed animals, string, and so on. Eventually I would like to use a kitchen as the environment, with a fully mobile robot as the manipulator.
• As little knowledge about the environment as possible should be hard-coded into the AI. As an ideal, “everything” would be learned — how to control the manipulator, sensor data interpretation, and the whole ladder of concepts that human infants climb on their way to effectiveness.
• Similarly, as little as possible should be built in regarding language and communication. Certainly no words for objects should be pre-specified. Best if no specific grammar rules are built in either. The AI should learn to communicate with its human partner solely through interaction.

It is not hard to imagine different levels of impressiveness from such a setup. I don’t see a good way to make a “test” that cannot be gamed through special-purpose cheats, but I think that fellow AGI researchers, if shown the results and the code/data built into the system beforehand, could make reasonable judgments about whether the accomplishments of the system are impressive or not.

A final interesting point about this task: It could be done with a robot and a real physical environment, or it could be done in a virtual environment. If a virtual environment is chosen, I would like to try as hard as possible to make it a good virtual environment — the most detailed possible models of objects, lighting, physics, and so on.

Task 2: Super-AM. A different way of looking at the path to the scientist/inventor/engineer AGI I’d like to see built starts by noting that even though formal methods appear to be useful only in certain unusual circumstances, those circumstances revolve around science and engineering. That is, detailed formal models (or maybe “semi-formal models” is a better phrase) are the way advanced science and engineering gets done. The most formal of all things is mathematics. And yet, computers pretty well suck at being mathematicians! I think there’s a good chance that figuring out why will lead to interesting progress.

There’s nothing new about this idea, of course. Doug Lenat wrote a celebrated program called AM in the 1970’s to explore exactly this task.

Three obvious extensions to the task come to mind:

1. Add a language component similar to the Super-SHRDLU task so that conversations can be held about the mathematics, and the system could be taught known mathematics.
2. Add the ability for the system to apply mathematics to the physical world, and focus mathematical efforts on formal models useful for science and engineering.
3. Extend the formal systems of interest to include computer programs, resulting in an automated programming system.

I had always assumed that I would focus on something like the first task as a lens for studying AGI issues — although the exact form of the task I wanted was not clear to me until I read Beyond AI.

At the moment I’m intrigued by the second task, though. While I continue the project of thinking about concepts and models in general terms I’d like to spend some time musing on math.

Question for any readers out there:  is there a better task than these? In an ideal world a substantial number of AGI thinkers would agree that some particular task is interesting enough to be worked on from multiple angles; I think such a focus of attention would help discover the strengths and weaknesses of different efforts. But AGI researchers are kind of like cats, and herding them in this way is probably impossible!