The virtual-worlds aspect addresses the fourth of these points: teaching methodology. The other three points are addressed in the Novamente software design itself, which is really the hard part, and comes out of the last 6 years of collaborative design and prototyping work between myself and the rest of the Novamente team, as well as about 15 years of theorizing and prototyping on my part before that.
The really hard part is making an AGI design that is conceptually correct according to philosophy and theoretical psychology, and is also computationally tractable on current computers. That's what the Novamente design accomplishes, and there are some papers about this on novamente.net, which are conference papers given at various academic conferences on Novamente. Granted they don't really tell the whole story, as Novamente is a big idea and hard to capture in a brief conference paper. I have a 350-page book manuscript describing the design and its motivations, and keep debating whether to publish it or keep it secret!!
Once you have a workable AGI design, though, the next step is figuring out how to teach it. This is where virtual worlds come in. They give you a wonderful combination. You have a relatively simple-to-deal-with setting in which perception, action, cognition, socialization and language can all be dealt with as a unified whole. And then you have potentially millions of people to teach your dumb baby AI and help it get smarter and smarter. The latter point is really important: look how smart Google got just by utilizing the collective intelligence in the links people put in their Web pages. Directing human intelligence into artificial intelligence is an important trick to use ... and the combination of the Novamente AGI design with virtual world embodiment will enable us to use it very effectively.
MLU: Would you care to speculate about how intelligent Novamente might become in this environment, and how quickly it may learn?
BG: How intelligent: It's hard to say what the upper limit might be, but I'm sure it'll be well beyond human intelligence. My strong feeling is that the simplicity of the virtual environment is not going to be our limitation in terms of achieving superhuman levels of intelligence, and nor is hardware. The limitations are going to be: do we have the right AGI design, and have we taught it long and enough.
About how long we'll need to teach it: this is harder to say. It partly depends on how many people have the incentive to simultaneously teach it the right stuff in the right way. We don't really know how much it will learn through explicit teaching versus through general, ambient interactions, for example. My gut feeling is that the learning will be faster, rather than slower, than the learning of a human child. Which is why I have posited that a Singularity by 2015 or so is not an absurdity, and one by 2020 should be patently achievable, if we put a concerted focus on it now. Build the AGI, put it online, create a situation where the residents of the online world are motivated to teach it -- and from that point on, it may be years rather than decades before we have something like a fully-fledged "artificial scientist." Now, let this scientist read some advanced computer science and math books, asking questions as it goes along, and see how fast it learns to improve its own self in a manner consistent with its initial goal structure.