Carlos, I think you are quite right that most people are unaware of advances toward AGI and will be surprised as rapid improvements emerge. As you point out, it's advancing on many fronts. I think of it as like a jigsaw puzzle being assembled, but without a clear picture of the finished puzzle. You are others are trying to sketch out what that picture will ultimately look like.
I have been working on one small, but I think important, part of the puzzle: Natural language understanding in limited pragmatic domains. As I see it, a big advance is coming as we finally build truly "expert systems" (a term from the 1980s when it was over-used, mainly for marketing purposes) in various subject areas based on accurate models of the concepts and techniques that people use in those areas, and then combine them with natural language understanding. That should enable such systems to communicate with people knowledgeably about those subjects, both to learn from us, to assist us, and ultimately to teach other people about those subjects.
An example of the kind of system I have in mind is described in my article Recipe Expert --https://medium.datadriveninvestor.com/recipe-expert-an-ai-system-that-understands-recipes-f1bf36b92a99 . Food recipes are quite well-structured, compared with most other kinds of written English. They involve lists of ingredients of various kinds, pieces of equipment, measurements, techniques of various kinds for using equipment to manipulate ingredients, to make measurements of them, and so on. I think it won't be too hard to hand-craft knowledge of the relationships between those basic things and build the basic English grammar and semantics for reading and writing about them.
Then the system could begin to grow, by reading basic recipes and writing about them, and, with human assistance, adding to its syntactic, semantic, and pragmatic knowledge, until its human assistants can also use English to improve its knowledge, with sentences like "A kilogram is about 2.2 pounds", "A whisk is a hand tool, like a spoon, for fine mixing", "'Rocket' is another name for arugula", "Anice smells like licorice", and so on. (The system will know that smell and taste are properties of ingredients although it can't experience those properties directly.)
As the system grows it will be able, more and more, to learn directly by reading recipes, requiring help from more knowledgeable people (or from other recipe expert systems) when it encounters misunderstandings, just as a person would who was also a novice at cooking and at understanding recipes. But gradually the frequency of assistance it required would diminish.
The key requirements for building a system like that are (1) a limited pragmatic domain (like cooking recipes) for which an adequate starting knowledge model can be hand-crafted, and (2) a broad enough English grammar and semantic model to be coupled with the model of the pragmatic domain.
Systems like the ones I am describing are only small pieces of the overall AGI puzzle, but I think many people will be surprised by their capabilities as they emerge.
