Yes, "understanding" is a process, but it's not well defined. And anyway, if it's a requirement for those things, and the AI did those things, then the only possibility is that the AI has this process in some way.
But well, our current AIs do not produce complete programs, nor fully working ones, nor do they say things without making mistakes. All the people making assumptions about the next generation that will do those things are basically hyping bullshit; and the next generation won't do those things because those AIs don't understand. What doesn't mean that eventually an AI that understands won't appear; of course it will. It just won't be the next generation of those.
There exist program synthesis systems that always produce "complete fully working programs" and that in fact cannot make any mistakes because they are based on algorithms with strong theoretical guarantees (like mathematical proofs) of their correctness. You give them a set of examples and some "background knowledge" and they spit out a program that's consistent with the examples and the background knowledge, no ifs and buts and maybes.
I could point you to several systems like that, if you wish, but first I want to make sure that we both agree what you are saying: a program synthesis system that never makes mistakes "understands". Is that right? Could you also please clarify what such a system "understands"?
Then I'm happy to link you to some systems like the ones I'm talking about.
Yes, I'm saying if it satisfies random requirements and synthesizes full working programs, then it has some form of understanding in it. If creates that understanding through learning, it's very possibly a general form of AI (I'm not certain of this part).
I've never seen any one that arrived at structure by learning. It's always programed in, what puts the intelligence on a very conventional place, and I doubt you'll disagree that it understands what it's doing.
But you're not going to get machine learning without predetermined structure. In Transformers, as in all neural nets, the structure is right there in the architecture of the neural net. So what kind of "learning" without "structure" exactly are you looking for?
Nope, you changed the context from a zero-knowledge AI optimized for text into things that come with knowledge designed in. There is understanding embedded on the things with knowledge designed in, but the understanding is not done by the AI.
Anyway, I'm surely curious about the systems that learn their own structure. You mean they learn the syntax rules? I'd say they have some understanding of the language syntax, but well, unless they can take a requirement and produce working code, they don't understand programing. (And well, if the requirement is in a formal language, they only understand formal transformations, that is, your AI understands compiling. What honestly, is quite limited, but would still be impressive.)
So, my apologies, but I must decline to share any pointers to systems or techniques etc. You seem to have made up your mind before you even see what I'm talking about, anyway.
If AI's outputs were literally indistinguishable from something that truly understands these concepts, then there's no practical difference if it really "understands". This reduces the concept of understanding to some abstract inconsequential property.
Yes, there is a practical difference. It has consequences in terms of legality, blame, and justice, the other does not. If an AI really "understands" things, then it is responsible for its actions and must be taken to court for crimes, and can be sued for mistakes in the output that result in losses (like any lawyer or professional).
Laws, courts, and their punishments are for humans, and they're likely to be absurdly bad at dealing with a personified AI. AI won't feel fear or remorse, may not have self-preservation as a goal, so it may not care about any threats at all. It may not even have sensory input needed to be aware of being punished.