Learning

AIs are trained on trillions of tokens.

Humans don't get that much training.

The gap will close upon AI research progress, but I keep wondering whether some of the gap ultimately cannot be closed.

The idea of sustained "deliberate practice" is a special method for improving skills is interesting --- because the formula is exactly the kind that allows for Magic to happen -- alignment of passion, intention, action onto one's own field of reality. I don't have direct evidence but it would be hard to imagine that there is no Magic effect when people deliberately practice to improve their skills.

So the literal billion dollar question is, what is the equivalent of this for large AI models? Or does this not exist, and we have hence disproved the Church-Turing thesis (which does not carry a billion dollars unfortunately)?

One possible (but far-fetched) solution for the billion dollar question is that humans (the trainers) provide the intention. In order for the effect to happen in the physical universe, it seems that at the very least the intention must be accompanied with some "action", not merely some kind of elaborate ritual that does not have a known physical causal effect on the model training, but it must provide a cause. The simplest way would be for a well-intentioned person, preferably one of the leaders of the project, to intentionally generate sufficient "random" bits to feed into the neural network, either as initialization or random seed. The size of the random bits must be large enough that it is not conceivable that it could have been chosen by luck, and not conceivable that a similar positive training result would be repeatable by exhausting the space of the randomness.

Sounds crazy, and most likely the idea of trying to influence silicon that is outside of your body is too weird for the field of belief to take hold.  But at least the theory is there.  The rituals can be designed in a way to improve the odds.

Not high though.

Another possible route would be to train the AIs with "love". I don't know how to do that but that's what the theory would suggest.

TBH I'm slightly afraid that we might instead just "conclude" we have disproved the Church-Turing thesis... which is sad.