August 13, 2007

Predictions on Artificial General Intelligence

Bruce Klein crafted a poll to gain a better perspective on the time-frame for when we may see greater-than-human level AI. There are several interesting answers to the question. The question was interpreted in many ways. I interpreted the question to be when will there be artificial general intelligence which has capabilities that signicantly exceed the human level range.

I have discussed my predictions on hardware for artificial intelligence

There are several papers from the 2006 Artificial General Intelligence workshop


I believe that progress will be getting faster. Examples of computer related capability that is improving faster than Moore’s law are: General purpose GPUs, better nanotechnology, quantum computers and graphene/plasmonic computers. It is the kinds of hardware for computer memory and processing which are the most successful.

An underlying aspect is how useful AGI will be. Quantum computing, which seems to be emerging with Dwave Systems, has usefulness discussions. Even with very good quanutm computing there are still mathematically provably hard problems. AGI pushing those frontiers will also still find it slow going even if the intelligence is faster than human. The greatest impact for AGI would be if they could somehow help circumvent the bad choices and screwups that humans have been making.

Examples of problems caused by bad collective choices
We are not in space and do not have good energy production solutions. This is not because solutions could not be thought of but because the social and leadership system for selecting and organizing around solutions is flawed. Nuclear pulsed propulsion, advanced nuclear thermal should have been developed.

Most people are not rich because of self-defeating behaviors.

Society has poverty because of short sighted corruption (group self-destructiveness)

AGI should still focus on business problems. Thinking Machines did not have a focus on making money and business problems and went out of business.

Predictions on AGI

1. When do we have the raw hardware capacity equivalence or passing of 10 petaflops (but we could get surprised and find we need 1 exaflop) ?

For the 10 petaflop number 2012 for a full real time human brain simulation. (100 billion neurons) 2018 for that simulation to be less than an average annual salary of someone in the developed countries. ($60,000/year at that time)

For the 1 exaflop number 2018 for a full real time human brain simulation. (100 billion neurons) 2023 for that simulation to be less than an average annual salary of someone in the developed countries. ($60,000/year at that time)

2012-2018 for the hardware for greater than human AI.

2. It will take more than hardware. Being able to put the pieces together for really useful AGI will take correct models of intelligence, theories, architectures, algorithms and integration of sensors and more access to information.

Each productive human is pretty highly specialized in the area in which they are making a contribution. The greater than human AI that does most things better than any person or group of people needs to be equal to one hundred to one million human specialists. Narrowly specialized and computers in the general range (some things better and some things worse then people) would make a difference but would not be reaching a different class of capability.

So I think high impact AI that is the range somewhat above and below human level will last from 2012-2035. I think we will be getting towards the real core of the problems re:intelligence that we do not understand yet during this period.

The whole new game level of greater than human AI could be in the 2030-2050 timeframe or a tad before depending upon when we can use molecular nanotech or other technology to make a lot of optical, quantum computronium. A bunch of cheap billion exaflop machines running 100 exaneuron equivalent and connected to molecular nanotechnology sensors systems would be able to compensate for inefficiency in the implemented AGI solutions with a lot of brute force capability.

People will slip-streamed in behind with tight integration with computer intelligence and adopt other enhancements.


Loki said...

Check out my CPS/$1K calculator. Our assumptions are slightly different, but that only changes things by 1 year.

My feeling is that the critical piece is sensors for mapping the brain at a high resolution in real-time. I think such sensors are required to obtain the dynamic behavior of the neurons with only rudimentary understanding of the biochemistry of each neuron type.

Advances in computation power seem a given, but the sensing and modeling process needs directed research.

Snake Oil Baron said...

I suppose that a lot depends on whether the behavior of individual neurons needs to be simulated or if the a simulation of neocortical columns could be achieved while cutting out much of the complexity needed to manifest them in cell form.

From skimming the Kurzweil article that the CPS/$1K calculator page used I could not tell if the brain computation estimate was for the whole brain or just the neocortex but even if it was the later I could imagine a lot of savings on things like knowing how to make bread and cross the street. A colour-blind AI with no sense of smell or taste and no interest in food or sex would still be extremely intelligent with far less computational demands.

bw said...

Loki thanks for the CPS/$1K calculator. Very nice.

For the sensing, see the work with flourescent markers. They can take real time snapshots. they use a rat and replace a section of skull with clear plastic. they they can do detailed tracking of activity. Recent cycling of flourescent markers lets them look at a larger area past the diffraction limit of visible light (200nm) down to 20nm.

Snake oil baron. Your discussion of an AI saving intelligence resources by not needing to think about sex etc... reminds of the Seinfeld episode where George becomes a genius when he stops trying to get sex.

brain said...

Now procesors made with 65nm technology and have ruoghly 100GFLOPS perfomance. Atom size is roughly 0.1nm. So I can predict, what processors making technology can reach 0.65nm (5 atoms size). Now single processor consist of roughly 500M tranzistors. One chip is size roughly 100mm^2. 65nm/0.65nm=100 times shorter. Becouse tranzsistors is in two dimension chip, then after processors making technology reach 0.65nm, single procesor will be 100^2=10000 times faster than "Intel Core2 quad". So processors, which will be maked with 0.65nm technology will has 100GFLOPS*10000=1PFLOPS=10^15FLOPS. To simulate human brain need at least 100PFLOPS. This single chip (which maybe be made at 2020) will needs power of roughly 100W. Our brains needed power is 20W. So conclusion is if chips will be very small and fast and will has tranzistors size of few atoms, they still not will be so eficient like effiecent like human brain (maybe only processors can be effiecent like human brain if chips will be maked with more tranzistors but will less frequently).