Pathways, Timelines and Superintelligence Scenario

Wait but Why summarizes some of the timelines and definitions for artificial super-intelligence. The general consensus among AI researchers is that Artificial superintelligence could arrive around 2060.

Nick Bostrom defines superintelligence as “an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills".

Less Wrong has some visualization and summary of superintelligence pathways

This displays the five pathways toward superintelligence that Bostrom describes in chapter 2 and returns to in chapter 14 of the text. According to Bostrom, brain-computer interfaces are unlikely to yield superintelligence. Biological cognition, i.e., the enhancement of human intelligence, may yield a weak form of superintelligence on its own. Additionally, improvements to biological cognition could feed back into driving the progress of artificial intelligence or whole brain emulation. The arrows from networks and organizations likewise indicate technologies feeding back into AI and whole brain emulation development.

Artificial intelligence and whole brain emulation are two pathways that can lead to fully realized superintelligence. Note that neuromorphic is listed under artificial intelligence, but an arrow connects from whole brain emulation to neuromorphic. In chapter 14, Bostrom suggests that neuromorphic is a potential outcome of incomplete or improper whole brain emulation. Synthetic AI includes all the approaches to AI that are not neuromorphic; other terms that have been used are algorithmic or de novo AI.

NBF believes that the primarily hardware based approaches to superintelligence tend to be more narrow solvers. If very good quantum computer computers are realized. The systems are cracking mathematically hard problems and assisting the process to improve machine learning.

Powerful solvers have human involvement to put in and get out answers and do not have the runaway AI scenario.

Before 2060 there should be
There should be advanced Optalysis style optical computers that have implemented Deep learning.

A startup company called Optalysis is trying to invent a fully-optical computer that would be aimed at many of the same tasks for which GPUs are currently used. Amazingly, Optalysis is claiming that they can create an optical solver supercomputer astonishing 17 exaFLOPS machine by 2020.

Deep Learning + 17 exaFLOP optical computer = 17 ExaFLOP Deep learning system by 2020.

The GPGPUs that implemented the Baidu Deep learning brain may be replaced by new optical computers.

Deep learning is a hot AI (artificial intelligence) field now. It is being used to develop AI that can perform tasks like learning to play video games better than humans

IQ Prediction from Structural MRI using Machine Learning

Stephen Hsu at Information Processing reports on a paper where the authors use machine learning techniques to build sparse predictors based on grey/white matter volumes of specific regions. Correlations obtained are ~ 0.7

A separate UCLA paper show that brain size alone correlates 0.4 with IQ. Also, a notable genetic sequence, located within the HMGA2 gene on chromosome 12, was linked with intracranial volume — in other words, the space inside your skull that marks the outer limit as to how big your brain can get.

PLOS One - MRI-Based Intelligence Quotient (IQ) Estimation with Sparse Learning

Abstract

In this paper, we propose a novel framework for IQ estimation using Magnetic Resonance Imaging (MRI) data. In particular, we devise a new feature selection method based on an extended dirty model for jointly considering both element-wise sparsity and group-wise sparsity. Meanwhile, due to the absence of large dataset with consistent scanning protocols for the IQ estimation, we integrate multiple datasets scanned from different sites with different scanning parameters and protocols. In this way, there is large variability in these different datasets. To address this issue, we design a two-step procedure for 1) first identifying the possible scanning site for each testing subject and 2) then estimating the testing subject’s IQ by using a specific estimator designed for that scanning site. We perform two experiments to test the performance of our method by using the MRI data collected from 164 typically developing children between 6 and 15 years old. In the first experiment, we use a multi-kernel Support Vector Regression (SVR) for estimating IQ values, and obtain an average correlation coefficient of 0.718 and also an average root mean square error of 8.695 between the true IQs and the estimated ones. In the second experiment, we use a single-kernel SVR for IQ estimation, and achieve an average correlation coefficient of 0.684 and an average root mean square error of 9.166. All these results show the effectiveness of using imaging data for IQ prediction, which is rarely done in the field according to our knowledge.

The 15 most frequently selected brain areas by the proposed method.
Colors mainly show different regions. doi:10.1371/journal.pone.0117295.g005

State of cognitive enhancement

A pdf by Nick Bostrom and Anders Sandberg that surveys the state of cognitive enhancement in 2006

Nick Bostrom's website

The paper reviews ways to train ourselves to be more intelligent/expert, drugs for enhancement /nootropic drugs, genetic modification, enhancing devices like computers, and brain /computer interfaces.

I think collaboration and collective productivity as in corporations has been somewhat discounted but communication and tool advancement could also make interesting breakthroughs in that area.

An older paper by nick Bostrom talks about the basic computational power needed for human level intelligence. Our most power supercomputers are in the middle of that estimated range. AI software lags. Access to the supercomputers for this purpose was lagging but there is the brain institute project .

Fairly large scale brain simulation projects have begun. 10,000 neurons were simulated. However, the project is not for artificial intelligence but to study brain structure The Brain Institute at the Ecole Polytechnique Fédérale de Lausanne (EPFL) in Zurich, researchers have built neocortical columns using supercomputing systems from SGI and IBM. They have a IBM Blue gene/L supercomputer with a peak speed of at least 22.8 Teraflops using 8000 processors. They think it will 10-15 years for the hardware to advance to a full brain simulation using their approach If Ovonic cognitive control devices are successfully developed this could happen sooner as they are more neuron like. Also, the use of GPUs and other hardware enhancements could accelerate hardware advancement

Red Herring discusses other cognitive computing projects that are started or being discussed The biggest being discussed is the Decade of the mind project. James Albus, senior fellow at the U.S. National Institute of Standards and Technology, says the NIST plans a project dubbed Decade Of The Mind, which calls for handing out up to $4 billion in funding to companies or universities doing research in mind-based computing.

Artificial Development is building CCortex, an simulation of the Human Cortex and peripheral systems, running on a computer cluster. They do not seem well funded enough to meet their ambitious goals.

There is also Steve Chen's Third brain project to create a biosupercomputer

If we project forward 10 years. It seems a strong possibility is that we could have far better understanding of the human brain and systems that are 10,000 times more powerful and various means to enhance human intelligence by 2 to 100 times without triggering a real superintelligence that is not "strong superintelligence. I foresee "weak superintelligence", which is human intelligence at high speed could provide an evolving pathway to strong superintelligence. It could be a safer path. Many could have access to "weak superintelligence" in the form of tighter coupling to advance computers and nootropic and genetic enhancement. Some in the singularity AI world have indicated that darwinian dynamics would not apply I think the software end is lagging and we will get "weak superintelligence" first and for an extended period. During this extended period darwinian dynamics would be applicable.

If the optimization of intelligence is speeding up and automating normal intelligence with the occasional insight into superior processes, then we would have a broadly advancing wave to strong superintelligence. This would not have many of the dangers that other foresee.

Scroll down slightly from this link and you will see a diagram of AGI plotted as an exponential line against a flat line for human intelligence. Widespread augmentation would make the human intelligence line one that is increasing as well

The danger has been expressed as a scenario where one superintelligence so outclasses all others that it rapidly reaches breakthrough after breakthrough so that its lead rapidly increases and becomes untouchable before any others can detect or respond while response would be effective. Leaving all at the mercy of the one superAI.

In comparing this to money, the superAI danger is like there exists an intelligence motherload (a buried mountain of gold which would be equal to a general theory of intelligence that leads to a far more rapid iterative intelligence improvement). Get to it and you are superrich while others are peasants. Alternatively if everyone (or large numbers) is able to get richer at a fast pace it would be more difficult for one to get dominance.

If we have a world of augmented intelligence then an important element (then as now)is securing vital resources. Getting your intelligence augment contaminated or pirated or turned against you would be bad.

tags:
congnitive enhancement
artificial intelligence
transhumanist
supercomputer
super intelligence

More reading:
Michael Anissimov on friendly AI

Michael Anissimov tracking AGI projects and work

List of AGI projects in 2006