Web Robots not Physical Robots are the Key to AI

A commentary on Hans Moravec's Essay Ripples and Puddles
Ben Goertzel
August 2000

Moravec gives a generally sound analysis of the challenges of AI and the strengths and weaknesses of various approaches to overcoming them. But he stumbles, in my view, on one crucial point -- his emphasis on robotics as the magic solution to AI. This focus on the physical world may have seemed obvious in the 70's, but here in the 21'st century, as communication networks expand and the era of virtual reality dawns, it's a severely limited perspective.

It's absolutely correct that true intelligence requires embodiment. Physical-world robotics provides a familiar approach to embodied intelligence. But embodying an AI mind in an Internet agent sidesteps a lot of nasty mechanical and electrical engineering issues, and allows one to focus on mind design and mind engineering.

I believe that the magic solution to AI, insofar as there is one, is not robotics but -- the Net.

Of course, embodying AI in Internet agents doesn't solve all the hard problems of AI. It only solves two of them: where to find a big enough brain for an AI system, and how to embody an AI system within a world that it can fluidly perceive and manipulate. This still leaves the problem of how to actually structure a digital mind -- how to build the software (or specialized hardware). In this respect, all the Net offers is a metaphor: the mind as a self-organizing network.

The "Internet as AI brain" is a fairly simple point, but Moravec chooses not to emphasize it. He points out, correctly, that simulating the detailed functioning of the human brain on contemporary computer hardware is very difficult, requiring a scale of processing power equal to millions of PC's. But he doesn't note that, through distributed processing across the Internet, it's possible to actually harness the power of millions of PC's, right now. Distributed.net and SETI@home started using the latent computing power of the Net, various start-up firms are now following in their footsteps -- and this is only the beginning.

The "mind as network" metaphor is a powerful one. Mind is a massively parallel self-organizing system of interacting, intertransforming actors, many of them specialized to particular domains or particular processes. It demands a complex-systems-theoretic analysis. If a sufficiently deep and careful analysis of mental processes is carried out, in this vein, one discovers that the divison between reasoning-based AI and neural-net-based AI is largely bogus; reasoning emerges in a clear and detailed way as a statistical emergent from neural net dynamics. The network approach cuts through the apparently unresolvable knots set up by traditional AI theorists.

Moravec suspects that "devising such programs requires lifetimes of work by world-class geniuses." My claim, on the other hand, is that these lifetimes of work have already been done by very clever computer scientists working outside the accepted mainstream of AI. I and my collaborators at Webmind Inc. believe that we've cracked the conceptual problems involved in building real AI. Our proposed solutions are embodied in the Webmind system, now under development for 2 years and scheduled for completion in 2001, Webmind will indeed require hundreds of PC's, but not millions -- though it will be able to enhance its intelligence by dispatching hundreds of small learning problems to millions of distributed PC's operating in a peer-to-peer network.

In short, Moravec foresees a path to AI that begins with simple robots like robotic lawnmowers and vacuum cleaners, and progresses eventually to human-level intelligence. I say: Sure, this can work, but it's an unnecessarily long and difficult path. There is another, shorter one, which is going to be followed first. The incremental development of intelligent robots which Moravec describes will take place in the context of an increasingly intelligent population of Internet minds.