On Sun, Feb 10, 2019 at 7:46 PM Alan Grimes <[email protected]> wrote: > I've read about half a book that I highly recommend to you: "Our Final > Invention" by Barrat. The book is kinda stop-and-go, with really > engaging factual sections combined with OH-GOD STOP BORING ME!!! > Wired-itis biographical sketches. I think that will help you understand > what we mean by "superintelligence". > > Once you have an AGI, it is almost certain that you will be able to find > a way to improve it, from there it's off to the races...
I looked up the reviews on Amazon. It's a series of interviews by a film maker with AGI experts like Yudkowsky, Kurzweil, Omohundro, Vinge, Dyson, Musk, Bostrom, Tegmark, etc. I am somewhat doubtful that events are happening as fast as many of them claim. Vinge's 1993 prediction of a singularity in 2023 (almost certainly between 2005 and 2030) looks in doubt, at least to me. The premise is that if we can make superhuman intelligence, then so can it, but faster. We don't know how it will do that because we aren't that smart. It will be magic. But I don't buy it. Intelligence depends on knowledge and computing power. An agent can't create another agent that knows more than the parent, so any improvement must come from computing power and increased capacity to learn. Every example of exponential self improvement works that way, from a colony of bacteria evolving drug resistance, to a company investing its profits, to human civilization augmenting itself with language, books, and the internet. In any case, computers already know more than us and can compute faster than us, and have been for so long that we don't even notice. We have had faster than exponential growth in population, the economy, and in global computing power for centuries. The economy is roughly proportional to e^(.0001t)(2100-t), which has a singularity at year t = 2100, but also a good fit to e^(0.03e^0.01t), which is super-exponential but does not have a singularity. There are many other rough fits, with or without singularities. Some keep going up forever. Some peak and go down. Take your pick. The future is really hard to predict. I believe that if Moore's Law is to continue, we will need to soon switch to computing by moving atoms instead of electrons. Clock speeds stalled a decade ago. Transistors are already down to about 100 atoms across, which is as small as you can make them and still distinguish P from N type silicon. They still use 10^5 times as much energy per operation as your brain. That's important as long as the biggest and least reliable component in your smart phone is the battery. Freitas worked out the physics of self replicating nanotechnology in https://foresight.org/nano/Ecophagy.html It is quite interesting that artificial life has nearly the same limits on size, speed, and power as DNA based life. We can make robots as small as bacteria but no smaller. They can replicate in tens of minutes and consume energy on a par with living organisms. We might marginally do better. We already have solar cells that are a little more efficient than chlorophyll. But marginally better is good enough to displace DNA based life. The biosphere encodes 10^37 bits of memory in DNA and performs 10^33 DNA, RNA, and amino acid transcription operations per second. Moore's Law suggests we will get there in the 2080's if it continues at the current rate of doubling global computing capacity every 1.5 years. That is far from certain, of course. Nanotechnology does not need to be self replicating, of course. Airplanes are safe because they are not birds and can't make baby airplanes. Likewise, I think it will be easier to build nanobots in factories like we build silicon chips. Nevertheless, the technology of putting atoms precisely where we want them will get cheaper. Once it is possible for anyone to buy cheap molecular scale 3-D printers, people are going to experiment and build these things, just like cheap computers enabled people to write viruses and worms. -- -- Matt Mahoney, [email protected] ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta6fce6a7b640886a-Ma747632240a5b19a70cf5095 Delivery options: https://agi.topicbox.com/groups/agi/subscription
