On Mon, Nov 4, 2019, 11:35 AM <[email protected]> wrote: > Uppp... hang on how can you even say that when the idea of AGI is even > more proposterous than a quantum computer. Its actually *MORE* powerful, > its not even intractable, its horribly difficult to make something have > *REAL* intelligence. > > Can u see how its proposteriously impossible anyway? >
The proof that AGI is possible is between your ears. Not only is it possible, but given enough planets with the right chemistry, it can arise spontaneously. The human brain is not a quantum computer. It performs non unitary operations, such as writing into memory, which a quantum computer cannot do. So in that respect the brain is more powerful than a quantum computer. But so is your phone. And for that matter, so is a light switch. In my paper on the cost of AI ( http://mattmahoney.net/costofai.pdf ), I outlined the technical problems that need to be solved. A human brain has 6 x 10^14 synapses, which transmit information about firing rate in about 50 ms. Therefore an equivalent size neural network (so far the most successful models for vision, language, and robotics), would require 1.2 x 10^16 multiply-accumulate operations per second, or 24 petaflops. Such computers exist, but the real challenge is to reduce the power consumption from tens of megawatts down to the 20 watts that the brain uses. Otherwise just the cost of electricity makes AGI not competitive with human labor (the most obvious application). This is not possible with transistors, but is possible with nanotechnology, moving atoms instead of electrons. The second problem is that the neural organization of the brain is far more complex than all but the biggest programs. About half of what you know, about 10^9 bits, is inherited, not learned, My own experiments with compressing source code and DNA suggests about 300 million lines of code. This just has to be written, at a cost of USD $30 billion. The alternative is simulating evolution. That required 10^48 DNA base copy operations on 10^37 bits over 3 billion years. The third problem is knowledge collection. Every job is different. It is not enough to train one AGI and make 7 billion copies to take over all of our jobs. The training cost of replacing a human employee is 1% of lifetime earnings. That doesn't go away with robots because the knowledge it needs is unique to individual humans, in our brains and not online. It has to be collected through speech and writing at a rate of 5 to 10 bits per second. Altogether the complexity of the global economy is about 10^17 bits. This is all very expensive. But the ROI on AGI is world GDP divided by market interest rates, or $1 quadrillion. It is going to happen, but only with decades of global effort, automating one little problem at a time. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T252d8aea50d6d8f9-Mdd7ee7d2a90f4b975a53058d Delivery options: https://agi.topicbox.com/groups/agi/subscription
