Ideally there should be certain abilities AGI would have in order for it to be truly useful. These would not be a condition for an initial AGI to be considered AGI, but utility-wise should be developed.
It should be able to build little pieces of software/algorithms to solve certain specific problems when they come up. Like how a human would build an internal model for a specific math problem, engineering problem, social problem etc... These algorithms should be re-usable, storable. An ideally shareable with other AI's to use, and even communicate ideas as a whole.
Modular idea sharing. Could be collected over time, and mass integrated at certain times (eventually weekly) with new iterations of the AGI version
The ability to integrate ideas wholesale with other AGI's.
- Package all relevant tokens/vectors/context, without having to copy the complete LLM and database(s)
- Other AGI processes and learns new idea/algorithm
- New ability is integrated
- Integrity is tested