Should be verifyable by human intelligence, no artificial or other intelligence can be used to verify the integrity or degree of the achieved intelligence; since humans are the only original creator intelligence, no substitute may fulfill the role of ultimate validation and verification. Context is the primary determinator of meaning, and if context is assumed it should always be mentioned, and clarified when ambigious. Should be based on a physical world model from its inception, since all language stems from physical actions and are abstractions from simple spatial logic (e.g. from, with, of, towards all have different levels of meaning from literal physical to figurative abstract. Even the word figurative stems from the idea of a physical figure where shape is the determining factor in its meaning in context) Without a rigorous understanding of the essential physical nature of language, the LLM blackbox is wide open for uncaught ambiguities that are not apparent to human curators and can mean something entirely different to an AI than we would assume it meant.
Shouldn't be required to be prompt engineered with things like: "You are an experienced programmer and use matter.js", it should figure out by context which expertise and libraries to use for the optimal outcome. The only prompt engineering should be in specificity, or style. If there are no context clues for the AGI, it should inquire it, and never assume something.
It should always have the primary sources of its information.
Ability to provide a backtracable tree to go over and verify the reasoning and origins of the outputs.
AGI should have a certain extent of "omniscience" or the ability to know everything that it has in its training set, and do anything with it that a human could, or a human with a computer could, like calculate correctly, comprehensive reading, reasoning, critical thinking etc. Should be able to answer things like
- What notable patterns can be detected about X in song lyrics in the last t period in subject Y (globally).
- Should probably respond first with clarifying questions to narrow down user intent and scope. Which signifies critical thinking.
A special kind of "turing-test" would have to have been created to verify true general intelligence
It seems that LLM's are way more useful at turning instructions into code, rather than text output. The exception being creating a perfect artificial word-salad of corporate marketing goo, which humans are very good at detecting and generally dislike. A perfect code that does something as best as possible is what you want, and it doesn't matter how artificial it is, that was the whole point; to be artificial. The same goes for artificial images and video which often have a certain "AI-glow" to it, and exacerbates the corporate fakeness of things in general.