There are no problems. UFAI could be constructed by a few people who know what they are doing on today’s commodity hardware with only a few years effort.
The outside view on this is that such predictions have been made since the start of A(G)I 50 or 60 years ago, and it’s never panned out. What are the inside-view reasons to believe that this time it will? I’ve only looked through the table of contents of the Goertzel book—is it more than a detailed survey of AGI work to date and speculations about the future, or are he and his co-workers really onto something?
My prediction / contrarian belief is that they are really onto something, with caveats (did you look at the second book? that’s where their own design is outlined).
At the very highest level I think their CogPrime design is correct in the sense that it implements a human-level or better AGI that can solve many useful categories of real world problems, and learn / self-modify to solve those categories it is not well adapted to out of the box.
I do take issue with some of the specific choices they made in both fleshing out components and the current implementation, OpenCog. For example I think using the rule-based PLN logic engine was a critical mistake, but at an architectural level that is a simple change to make since the logic engine is / should be loosly coupled to the rest of the design (it’s not in OpenCog, but c’est la vie. I think a rewrite is necessary anyway for other reasons). I’d swap it out for a form of logical inference based on Bayesian probabalistic graph models a la Pearl. There are various other tweaks I would make regarding the atom space, sub-program representation, and embodiment. I’d also implement the components within the VM language of the AI itself, such that it is able to self-modify its own core capabilities. But at the architectural level these are tweaks of implementation details. It’s remains largly the same design outlined by Goertzel et al.
AI has been around for almost 60 years. However AGI as a discipline was invented by Goertzel et al only in the last 10 to 15 years or so. The story before that is honestly quite a bit more complex, with much of the first 50 years of AI being spent working on the sub-component projects of an integrative AGI. So without prototype solutions to the component problems, I don’t find it at all surprising that progress was not made on integrating the whole.
There are no problems. UFAI could be constructed by a few people who know what they are doing on today’s commodity hardware with only a few years effort.
The outside view on this is that such predictions have been made since the start of A(G)I 50 or 60 years ago, and it’s never panned out. What are the inside-view reasons to believe that this time it will? I’ve only looked through the table of contents of the Goertzel book—is it more than a detailed survey of AGI work to date and speculations about the future, or are he and his co-workers really onto something?
My prediction / contrarian belief is that they are really onto something, with caveats (did you look at the second book? that’s where their own design is outlined).
At the very highest level I think their CogPrime design is correct in the sense that it implements a human-level or better AGI that can solve many useful categories of real world problems, and learn / self-modify to solve those categories it is not well adapted to out of the box.
I do take issue with some of the specific choices they made in both fleshing out components and the current implementation, OpenCog. For example I think using the rule-based PLN logic engine was a critical mistake, but at an architectural level that is a simple change to make since the logic engine is / should be loosly coupled to the rest of the design (it’s not in OpenCog, but c’est la vie. I think a rewrite is necessary anyway for other reasons). I’d swap it out for a form of logical inference based on Bayesian probabalistic graph models a la Pearl. There are various other tweaks I would make regarding the atom space, sub-program representation, and embodiment. I’d also implement the components within the VM language of the AI itself, such that it is able to self-modify its own core capabilities. But at the architectural level these are tweaks of implementation details. It’s remains largly the same design outlined by Goertzel et al.
AI has been around for almost 60 years. However AGI as a discipline was invented by Goertzel et al only in the last 10 to 15 years or so. The story before that is honestly quite a bit more complex, with much of the first 50 years of AI being spent working on the sub-component projects of an integrative AGI. So without prototype solutions to the component problems, I don’t find it at all surprising that progress was not made on integrating the whole.
Any evidence for that particular belief?
What do you think is missing from the implementation strategy outlined in Goertzel’s Engineering General Intelligence?
Haven’t read it, but I’m guessing a prototype..?
If you had that then you wouldn’t need a few years to implement it now would you.