There was a huge news story about AI and AGI which rightly shook the world over the past two days:
What shook me most was a clear statement by Sam Altman, head of OpenAI, depicting a commitment to move ahead with lots and lots of apps making money in the short term without putting much energy into cross-cutting or integrative solutions.
In many ways, the really big issue is whether the human species is capable of working together to develop that level of integration which is necessary to avoid the total chaos and instability (leading to extinction) which is on its way NOW unless we work better and more effectively to use our own natural intelligence, WITH AI and such used as positive tools.
OVERVIEW FOR HIGH DECISION MAKERS
The key acronym AGI, Artificial General Intelligence (AGI), promulgated many years ago by Ben Goertzel, is finally getting the high-level global attention it deserves. The world badly needs all of us to connect better and deeper, to do justice to the interconnected technical and policy issues which AGI is already pushing us into very rapidly.
BUT FIRST: WHAT **IS** AGI?
I have seen many, many definitions for many decades.
I first heard Ben's talk in person in the WCCI2014 conference in Beijing, where I presented my own concept of AGI AT THE LEVEL of mammal brain intelligence. https://arxiv.org/abs/1404.
Phrases like AGI are not defined by God. We all have a right to work with different definitions, so long as we are clear.
=== LIKE SOME OF YOU, I would firmly reject the old Turing test as a definition of what an AGI is. Even Turing himself used much more powerful mathematical concepts when he moved on from early philosophical debates to mathematics that can actually be used in computer designs! (I bcc the friend who showed me Turings Cathedral by Dyson, a great source.) The Turing Test makes me laugh about Eliza, perhaps the first AI-based chat program, developed at KIT decades ago, which showed many of us just how incredibly shaky the Turing test really is.
I would propose that we define an AGI as a universal learning system, which learns to perform either cognitive optimization or cognitive prediction as defined in the NSF research announcement on COPN which is more advanced than any such announcement elsewhere even today:
In other words... universal ability to learn to adapt to any environment, with maximum expected performance, or to predict or monitor any time-series environment over time.
TODAY, I created a googlegroup on QuantumAGI to facilitate easier discussion of the most important players in the real technology creating
a POSSIBILITY of true quantum cognitive prediction or optimization, or function minimization/maximization.
===
Years ago, in the crossagency discussions which created COPN, my friends who ran cognitive science and AI in computer science asked: "Do we want to set the bar so high? " I asked: "Should we really use the word 'intelligent" to refer to systems which cannot even learn anything?" In fact, people with long and deep experience in classical AI knew about Solomonoff priors, one key approach to universal learning-to-predict, which Marvin Minsky himself urged me to study in the 1960s when I took an independent study from him.
The mathematical foundation for the most powerful, universal cognitive prediction now emerging, using classical computing and deep neural networks, is reviewed at: werbos.com/Erdos.pdf. QUANTUM AGI extends that further, simply by doing orders of magnitude better in the loss function minimization tasks at the core of all general effective cognitive prediction methods. EXAMPLES of thermal quantum annealing, in relevant special cases, have already demonstrated that advantage, as shown in papers from IBM and Japan and others at https://drive.google.com/drive/folders/1RKjXCsLMFpo8u_WFhnHEmlVHbo9gmHEu.
==============================
IS IT REALLY SAFE TO UNLEASH AGI AND QAGI ON THE EARTH, GIVEN HOW SCARY THE PRESENT TRENDS ARE??
Many of us, including me, have thought VERY long and hard on that.
Based on the recent talks from Ilya and Altman, etc., I believe that we are presently on course to a very intense and difficult future, similar to the kinds of massive changes in niche which have doomed the world's leading species to extinction again and again over the millennia. We are in the kind of decision situation which meets the technical concept of a "minefield" situation, which we are unlikely to survive unless we build up quickly to a level of collective cognitive optimization beyond ANY of today's AGI or social institutions.
FURTHERMORE.... as in my new book chapter attached (book coming out next month or January from India Foundation), I really doubt that our cosmos lacks intelligence at the level of QAGI already. Keeping up with that level of collective intelligence may simply be ESSENTIAL to our best chances of survival as a species.
YES, there are HUGE dangers if this is developed in the dark. That is why I believe in the necessity of open, transparent international development, including even leadership in the QAGI technology itself in new international venues.
ANOTHER VERSION WITH DETAILS FOR SUBSTANTIVE TECHNOLOGY LEADERS
HOW AGI WORKS --
There are a few different definitions out there about what AGI (Artificial General Intelligence) actually **IS*. YOU ALL can rightly use many ways of handling definitions, because you communicate with different audiences. Please forgive me if I still adhere to many commitments of John Von Neumann, the mathematician whose work underlies MANY branches of science. Von Neumann would tolerate me giving you ONE or TWO useful definitions of AGI, and explaining where it leads.
AGI: universal learning machines, a kind of INTENTIONAL SYSTEM, designed to input some measure of "cardinal utility" U, and to learn the strategy of action or policy which will maximize the expectation value of the future value of U. In modern neural network mathematics, the best way to name these is to call them "RLADP" systems, Reinforcement Learning and Approximate Dynamic Programming. Even today, the old book "Neural Networks for Control" by Miller, Sutton and Werbos from MIT Press is an important source for learning what this means in practice, and understanding where key places like Deep Mind are really coming from. These are systems which LEARN TO DECIDE, in an agile way.
BUT THERE IS NO ESCAPING the essential importance of "where does U come from?" This is basically just a modern reflection and extension of the most ancient problems of philosophy; Von Neumann's concept of U traces back clearly to utilitarians like Jeremy Bentham and John Stuart Mill, and back from there to Aristotle's Nicomachean Ethics, which I remember reading at age 8 when I found it in my mother's old schoolbooks.
BUT: a more practical definition: modern AGI, in practice involves THREE elements, three types of universal learning machine. There is RLADP, which learns to exert decision and control (which has be applied to anything from monetary transactions to weapons control to words to energy systems). There is learning to predict or model or describe the state of the world, which FEEDS INTO making better decisions. And there is the "simple task" of learning to minimize some function F(W) with respect to weights W.
THE problem of survival for humanity is an example of an RLADP problem, where we try to maximize the probability of human survival, which of course requires further definition and refinement. FOR NOW --
THE OPENAI debate reminds me that the problem of human survival or exaltation is a specific TYPE of RLADP problem, which mathematicians would call "highly nonconvex." Concretely, it is a MINEFIELD problem, where the paths of possibility ahead of us mostly hit explosive "unexpected" sudden death -- but also with aspects of "needle in a haystack" where there are GOOD possibilities we might miss. SOLVING such problems requires a lot of caution and foresight, which is why stronger work in foresight is essential to human survival. SUCH RLADP problems end up requiring solution of highly nonconvex function minimization or maximization problems.
Early in this century, NSF organized the most advanced research effort ever in probing this mathematics AND connecting it to the intelligence we see in mammal brains: https://www.nsf.gov/
Following that program, I often say "cognitive optimization" to refer to RLADP and intelligent function minimization/maximization. "Cognitive prediction" refers to that other universal learning capability, which is advanced further in werbos.com/Erdos.pdf and in Buzsaki's recent book on the brain as a prediction machine.
I attach my paper in press from the India Foundation, and another in a book now available by Kozma, Alippi, etc, giving even more details.
Quantum AGI, as I define it (THE canonical definition created in my published papers and patent disclosure), simply ENHANCES these three universal learning capabilities -- RLADP, prediction/modeling and function minimization -- by HARNESSING the power of quantum physics AS DESCRIBED BY THE GREAT PHYSICIST DAVID DEUTSCH OF OXFORD.
You could call this "quantum cognitive optimization" and "quantum cognitive prediction."
The foundation which all QAGI is built on is minimization or maximization of nonlinear functions.
It was initially developed (by me) to address minefield or needle in a haystack types of problem, though it looks as if the new types of quantum computers will also give many other improvements.
Here is a metaphor: if you had a million haystacks or gopher holes in your big back yard, to FIND the best needle in a haystack (or deepest gopher hole), WHY NOT HIRE A MILLION SCHRODINGER CATS to work in parallel, and report back which is best?? A million times faster than one-at-a-time search!!
Deutsch's Quantum Turing Machine is not a brain or an AGI; just a faster type of old sequential computer, a Turing machine.
DWAVE was a HUGE mental leap forward, which would FIT the vision I just described... BUT ONLY if the function minimization at the core of the system is replaced by the kind of hardware which ACTUALLY harnesses these cats. (DWave is like paying for a million cats, but putting them on a leash, locking them up on a patio or a restricted sidewalk. Strong efforts at energy conservation have that effect.)
The papers in our Project Amaterasu folder and recent emails describe how Deutsch's physics works here, and how to build the hardware.
No comments:
Post a Comment