Could you please provide some insight into the current stage of developments in AGI area? Are there any projects that had breakthroughs recently? Maybe some news source to follow on this topic?
The state of AGI research is pursuing the few problems that we have been able to break off from the gigantic research problem. These are terms which can be more thoroughly looked into.
A few of the main focuses are:
One Shot learning - You know how a person can sometimes learn to do something by something by seeing literally 1 example of it? Well current learning methods on the whole are not able to accomplish this to the extent that we can easily take for granted. Work is being done to find ways to approach this feat of learning and it’s on its way to becoming much more influential.
Transfer Learning - If you have ever played a side scroller like Mario, and then I gave you a slightly different game like Sonic, odds are you’d could learn to play Sonic faster than it took you to learn to play Mario. This is because of learning “savings” you have by transferring your Mario knowledge to the new Sonic domain. This is a much more popular research arm than one-shot, probably because it’s easier to think about but also because there have been promising results of pretraining a network on one set of data and focusing it an another task.
Creativity/curiosity - Although one could say that GANs have really changed how humans can be more creative, it is difficult to quantify curiosity and creativity This paper gives an okay overview. Moreover, allowing an agent to take the chances and make some mistakes as is the nature of creativity concerns many people who are focused on AI Safety.
Understanding concepts - This is subtle but very very important. Current AI methodologies struggle with imbuing AI with the ability to have concepts. By concepts, I don’t mean “it kinda looks like this neuron in the second last layer is sensitive to tires”. I mean that understanding what a tire looks like is just a small part of understanding what a tire is, what it is used for, what it affords someone to do etc. This research direction is in it’s infancy but will be much more influential as more theories and ideas are brought forward.
Despite the progress made in these fields and in the many other areas in AI, there is still much to be done and understood before we can finally have s̶k̶y̶n̶e̶t̶ Wall-E.
In the paper Artificial General Intelligence: Concept, State of the Art, and Future Prospects (2014) Ben Goertzel gives an overview of the AGI field and its progress. He describes the main approaches
- symbolic (e.g. cognitive architectures, like SOAR)
- emergentist/subsymbolic (neural networks)
- hybrid (combination of symbolic and emergentist)
- universalist (AIXI and Godel Machines)
to AGI and metrics to assess human-level intelligence and partial progress.
There's also an older book by Cassio Pennachin and Ben Goertzel called Artificial General Intelligence (2007). A chapter of the book, Contemporary Approaches to Artificial General Intelligence, gives a brief history of the AGI field.