Great Question. The TLDR answer is, only after we teach it, and even then it wil

Great Question.
The TLDR answer is, only after we teach it, and even then it will take a while to tune it. But producing it from scratch is almost impossible to imagine. 😉
I work with the AI’s, meaning the paid versions, every day and it’s amazing how astoundingly dumb they are. 😉 I also have a fairly good understanding of the evolutionary direction and functionality that will be necessary to produce AGI. And it’s incrementally approaching that capacity – despite that I know, and these organizations konw, that the power consumption necessary to convert from single shot and attention based replies to a spectrum of answers, competing together, and then recursively competitively compared with one another.
So, I haven’t seen any genesis of knowledge or understanding at all, such as people produce, no matter how simple. And I have a very hard time imagining that there is any possible difference between the type of instruction I give ai’s today so that they gradually even understand a question I want to ask evolve into the capacity to even interpret those questions until after I’ve trained it to so so.
Now we have a fairly clear idea of how to train an AI to do the work, but there are two considerations (a) the rate of progress means any present investment may not carry forward (b) there are only so many people in our organization, and while the basic construct of all decidability is something we’re comfortable with training, the catalogue of relatively universal criminal, ethical, and moral questions – at least the thirty general categories – and each of those thirty general categories will take 200 to 500 training examples (input-output pairs) both expressions and questions, and it will require multiple cultural variations before it learns the general pattern across cultures. So for our existing staff, it would be almost impossible. Instead we’d need to train people possibly with economics or legal degrees, in the material then have that larger group produce the training data. (c) this kind of thing is expensive for a think tank. 😉

Cheers
CD

Reply addressees: @romanyam