Leading AGI theories?

AGI by definition would be capable of exhibiting all of the modes of thought that people do to solve real-world problems. That implies AGI would be capable of things LLMs are not, such as perception of the passage of time, experiencing boredom, metacognition, and intrinsic motivation from internal values. It would need these things to perform metacognitive monitoring, analogical abstraction, and interactional repair. I think LeCunn's world-modelling would be a necessary but insufficient part of a system capable of these things. Beyond the ability to form and use discrete models, an AGI would also need to build an ontological hierarchy for itself, drawing upon more-primitive concepts in its hierarchy to form more-abstract concepts by way of analogy (like George Lakoff's ontological metaphors). It would also need the ability to form internal (personal) values, from which motivation and initative are derived, and internal values are dependent upon a subjective experience of "good" and "bad", which in turn is dependent on embodiment. Again, Lakoff's theories are relevant to this, this time his theories about embodied intelligence. I don't think embodiment needs to literally take the form of a physical body, but at least something that is capable of experiencing sensations from which to form metaphors and heuristics about "good" and "bad" would need to be simulated. As for the perception of the passage of time, it seems to me like you might be able to approximate such a thing in an LLM's context buffer by having specialized tokens which represent the age of the tokens near it. You would need to update those and move them around periodically, which would invalidate K and V caches and force you to regenerate them, but depending on the update frequency that might not be a horrible burden. It's tempting to consider putting timestamps in the context instead and calculating age relative to the current time, but that's not how LLM inference works. To teach a model how to behave in respect to aged tokens, you would really need time tokens which were relatable to time tokens in its training data. The requirement for interactional repair (making note of how one's environment is responding to one's behavior and changing behavior in mid-stride) implies a very high rate of inference anyway, so perhaps invalidating your K and V caches with dynamic age tokens wouldn't be a huge burden after all. The interactional repair mechanism (and probably other modes and behaviors) implies the need for a Blackboard Architecture -- you would need a state space where you have the most-recently inferred plan (like something the AGI is saying to a person) that independent, asynchronous processes could inspect, and to which they could contribute ancillary information while the plan was in action (like observing the person's facial expressions and raising a flag for "they are not reacting as expected"). The main process would only continue following the plan for as long as there wasn't ancillary information in the state space of sufficient import to ditch the plan and infer something new. Wikipedia has an okay description of Blackboard Architecture here -- https://en.wikipedia.org/wiki/Blackboard_(design_pattern)