Unlike previous generations of AIs, which has more specialized functions covering a constrained domain, these entities have to be trainable, robust, social, general and affordable, he said.
When will we have robots serving lattes at Starbucks? The consensus is between 30 and 100 years, and the big factor will be money--massive investment in research and development to create advanced general intelligence (AGI). Pell said that the major milestones on the path to AGIs are unclear. There is no roadmap or significant funding mostly because it too far out--no low hanging payback fruit.
"At some point the creation of better AGI will become essential to creating better products, and it will be the basis for competition and a massive industry," Pell said. "Things have come far enough along in development of basic markets, the costs and Moore's Law to fuel development of AGIs." He gave an example of autonomous vehicles, the driverless car, as a kind of milestone.
As in past times, U.S. Department of Defense will be a major source of funding for technology development. The Internet, for example, came out of DoD DARPA, the Defense Advanced Research Projects Agency, as well as much of the current efforts on autonomous ground vehicles and robotics to produce improved ways to conduct wars...or keep the peace. Consumer and entertainment industries are increasing turning to robotics and artificial intelligence to make virtual worlds more realistic.
The money will come when there is plenty of low hanging fruit,
as opposed to blue sky ideas, to pick.
Pell outlined several approaches to creating AGI. A brain-inspired model would take the scan of the brain and simulate the human brain and crank it up along the path of Moore's Law. However, we may not have enough understanding of how the brain works to create good enough AGIs.
A top down engineering approach may have nothing to do with humans. In this approach you don't suffer from the idiosyncracies of the human brain of the human brains, but you are very far removed from the one existence proof--that the approach would ever yield results, Pell said.
A fusion of the brain simulation and top down engineering approaches is the most likely outcome, Pell said.
With intelligence augmentation, humans would be transformed into hybrid systems.
A radical architecture would be simulated evolution, in which researchers "stumble" into an artificial environment, simulate it massively and then natural artificial beings evolve and generate own languages and social structures.
Listen to my podcast with Pell about pathways to advanced general intelligence