This is wholly philosophical.
Yes, and it always has been. "What is intelligence?" is indeed a philosophical question.
That does not make it meaningless, if that's what you were mistakenly thinking. It means we need valid definitions instead of smuggling in whatever engineering shortcut happens to be convenient.
you have to put your semantics in a frame of reference, an anchor of objectivity. You can’t just say ‘this is objective, we just don’t fully understand it’. There’s no reference to anchor an argument and the argument is meaningless.
Nonsense. The objective anchor is and always has been that intelligence is a cognitive faculty. It involves awareness of reality, a grasp of facts, the understanding of meaning, judgment, and the ability to relate concepts to what they refer to.
You're free to reject that, of course, but then you'd be no longer talking about intelligence in the ordinary serious sense. You'd be talking about output-production.
Mouse traps and smoke detectors collect information and react to them, but they do not store information. The storage of information is a key criteria of intelligence which you are overlooking.
Overlooking isn't the right word. I'm rejecting it as sufficient. Storage does not create understanding.
A hard drive stores information. It does not understand that information. A library stores information. The library does not know what the books mean. A tape recorder stores speech, it does not understand the speech. A database can store and retrieve answers, but it does not therefore possess intelligence.
So adding "storage" does not rescue your definition.
The other missing piece is feedback, altering the information stored from previous inputs based on future inputs, outputs and (possibly) external scoring.
That's not enough. A spam filter updates from feedback, that doesn't mean it understands anything.
Storage + feedback + output construction is still just machinery unless there is awareness, meaning, and grasp.
Basically what you're doing is trying to rebuild intelligence from external functional behavior while leaving out the cognitive faculty itself.
You cannot ask a mouse trap or a smoke detector an arbitrary question. You can ask an LLM any question, even if different models are tuned for specific domains, they can all attempt to address any question.
You can ask an LLM any question because it is a language-output system trained on absolutely massive amounts of human linguistic material. From that, an LLM obtains broad verbal coverage. But understanding is not established. A system can absolutely attempt an answer without knowing what the question means.
This is the issue.
I would say that for practical purposes, ‘understand’ can be defined as the storage of information and the modification of that information with feedback for the purposes of constructing outputs.
No, that simply isn't understanding. It is, however, a proposed recipe for simulating understanding-like output.
I'd like to point out that your own choice of words gives the game away:
for the purposes of simulating it
Simulating understanding is not understanding.
A flight simulator doesn't fly, a weather simulation is not weather, a simulated fire doesn't burn. And a simulated understanding does not understand.
If your practical engineering definition is "the system stores data, updates from feedback, and produces useful outputs", that's fine, feel free to use that definition in an engineering context. But do not pretend that you have thereby established intelligence. You have only abused the definition of "understanding" to the point that a sufficiently elaborate data processing system qualifies. You may call it objective precision, I call it fancy concept vandalism.
