2 ms·
Meh. I mean that. It's a very hard "meh". On the one hand, human-equivalent AI is undoubtedly possible, as is greater-than-human AI, as is (to my knowledge:
by eli_gottlieb 12y ago
Meh. I mean that. It's a very hard "meh".
On the one hand, human-equivalent AI is undoubtedly possible, as is greater-than-human AI, as is (to my knowledge: I have some ideas on how to do it) self-improving AI.
And yes, the "blessing of abstraction" effect of transfer learning does get us out of the Curse of Dimensionality somewhat, allowing the sample complexity for general learners to be a bunch lower than the sum of the sample complexities for a series of individual narrower learners.
And, at its very mathematical fundamentals, statistical learning works, in a way that pre-Winter AI just fucking didn't. I mean, theoretically you could learn a super-duper god-AI in a sufficiently large two-layer neural network... it would just be infeasibly large and intractable to train.
But the only way you get a "hostile Singularity" is if your super-duper god-AI isn't just smart, or smarter than one human, or even smarter than one clever human, but in fact more hierarchically general (able to learn more from the same data), quicker-thinking (able to sample and infer faster), and more accurate (able to approximate intractable distributions more closely from the same amount of computation spent on inference) than whole organized teams of trained, expert humans. And then you have to feed it sufficient training data for it to understand the world more deeply than said teams!
Or unless it just bribes or otherwise corrupts said teams (which is, admittedly, the most likely way for this scenario to happen: humans are definitely easier to buy off than any artificial agent anyone would bother to design).
So a "soft take-off" could happen. But I'm really not sure anyone other than ideological "singulatarians" is putting enough effort into causing it for it to be very probable.
Hard take-offs are very unlikely for reasons of sample and computational complexity: educating a human takes approximately 20 years. Even assuming you can go 100x faster with your AI... 20365/100 = 203.65 days, approximately 4 months of Terminator training in which the designer might notice the AI was "evil" and turn it off.