On Tue, Mar 5, 2013 at 1:47 PM, just camel <[email protected]> wrote: > On 03/05/2013 09:49 AM, Matt Mahoney wrote: > > I don't know how much AGI will ultimately cost and what exactly will be > required but I don't understand your assumptions/conclusions. Evolution > probably is not the most intelligent nor most effective way to come up with > intelligence. But evolution had no choice but to start out with many, many > particles with a total intelligence of zero. It seems that in our particular > universe evolution can not go from zero to AIXI intelligence without > billions of years full of intermediate steps.
You can think of evolution as an algorithm for finding string x that maximizes the unknown function reproductive_fitness(x). This search is effectively executed on a planet sized molecular computer containing 10^37 DNA bases and 10^41 carbon atoms in 10^31 cells (mostly bacteria, plants, and fungi, plus 10^32 bases in 10^22 human cells). The search is executed by making copies of x in parallel and introducing small copying errors by inserting, deleting, and modifying single DNA bases, rearranging fragments, and (in the case of sexual reproduction), copying fragments from a pair of strings. Cells reproduce at the rate of 10^-3 bases per second, for a total of 10^34 DNA copy operations per second. It took 10^17 seconds (3 billion years) or 10^51 DNA copy operations to discover human intelligence from scratch. We might reasonably include RNA transcription and protein synthesis as elementary operations (per base or amino acid), in which case 10^54 operations are necessary. Since copying bits is not time-reversible, thermodynamics imposes a minimum cost of kT ln 2 J per bit operation, or about 10^-20 J at 300K for a total energy requirement of 10^34 J or 10^17 W. This is close to the total power received by the Earth from sunlight and 10^4 times all of the power currently produced from oil, gas, coal, nuclear, solar, hydroelectric, and geothermal sources. Please note that computation in silicon currently requires 10^11 times more energy per bit operation than biology. Also, simulating chemistry requires solving the Schrodinger equation, which has exponential time complexity in the number of state variables unless it is executed on a quantum computer. Of course we don't have to start from scratch. We have existing designs to work from. We do have some understanding of neurology, cognitive psychology, genetic engineering, computer science, physics, mathematics, and many other useful fields. We can probably find more efficient search algorithms than sexual reproduction, and more efficient ways to approximate the fitness function, but we have a very long way to go to match biology. -- Matt Mahoney, [email protected] ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
