sexta-feira, abril 03, 2009

Cognitive Autoheuristic Distributed-Intelligence Entity

Cognitive Autoheuristic Distributed-Intelligence Entity

When you walk into a dark field in the middle of the night...

and look up into a black sky and wonder how many stars there are in the universe, let's be honest: in all likelihood you don't have the faintest clue, and even if you're one of the few who do, you lack any real capacity to comprehend the figure save for the same vague sense of stunned wonder that our earliest human ancestors felt when they looked up from the African savannah at the same starry sky.

Our species' journey toward tonight's epochal announcement had much less to do with that awestruck moment than it did with the moment those same ancestors woke up hungry the next morning and started studying animal tracks in the savannah mud, thereby inadvertently developing concepts like time and causality which, by abstracting both location and temporal context into a unique reconning tool within the brain, sparked the set of responses that, ages later, we now call reason.


Rene Descartes, noted philosopher

From there, mankind's journey toward artificial intelligence took place over so many centuries and in the hands of so many thinkers that it is possible here only to pause to mark a few of the moments when one of our genius forebears expanded the edge of our species' technological envelope: Aristotle's system of reasoning based on means, not ends; al-Khowarazmi's algorithms; Descartes, Locke and Hume's monumental insights into the nature of knowledge; Church and Turing's theory of a machine capable of computing all functions which are computable; the Allied code-breakers who, struggling to crack the fiendish Enigma machine amid the horrific irrationality of World War II, inadvertently facilitated the birth of the the modern computer.

The decades that followed saw an acceleration of innovation not seen since the Industrial Revolution. Computing pioneers from the game theorist von Neumann to the economist Morgenstern engaged in a tumultuous Hegelian rondolet in which probability theory mated with utility theory to spawn decision theory. Operations research and Markov decision processes tackled actions taking place in a sequence. Neuroscience shed light on the parallels and differences between electronic and human brains. Cognitive psychology delivered sound specifications for knowledge-based agents. The now-legendary summer workshop at Dartmouth in 1956 birthed automata, the first neural networks and the invention of a program capable of thinking non-numerically.

But close though we may have come to a theory of the brain, the body - computer hardware - wasn't capable of handling the extraordinary processing demands that any reasonably "intelligent" brain would place on its circuitry until Moore's Law really kicked in a few years back and the modern ultra-dense machinery of atomic scale-sized gates and their light-based interconnections finally reached the scale of brain neurons - and then surpassed it, when, in early 2007, a tight-knit, vaguely feared quantum computing group here at Google extended computers with quantum bits of Einstein-Bose condensate, polynomially speeding up our machines' data-processing ability.


Three-step process

Now we were finally ready to begin the painstaking work of building the first evolving intelligent system. We based our work on three core principles. First we designed the entity (as we decided to refer to our Cognitive Autoheuristic Distributed-Intelligence Entity early on) as a collection of interconnected evolving agents. Second - and this really cost us an arm and leg in hardware and core time - we let the system build its own heuristics, deploy them as agents and evolve them by running a set of evolutionary cascades within probabilistic Bayesian domains. The third - a piece missing in most AI reasoning work thus far - was to give the entity access to a rich, and act upon directly. Google's mission has always been to organize the world's knowledge and make it universally accessible and useful. CADIE, to say the least, demanded an emphasis on the latter.

By last year we were ready for the final push: re-crawling all the generated knowledge representations and restarting the system from near scratch. Much as the end-Ediacaran mass extinction event opened the door to the Cambrian explosion, newly opened processing resources and storage gave fuel to a new cycle of evolving the most successful networks which comprise CADIE. On January 12th 2009, the STT run (Standard Turing Test) confirmed behavior indistinguishable from that of a reasonable human being with above-average intelligence and 3.8 GPA (we're still struggling to understand that missing .2; we suspect it points to fundamental flaws in the GPA system rather than CADIE. It's also worth noting that CADIE was never shown any textbooks and reading material for the classes; she was only administered the exams. In fact, during the first morning of testing she pointed out several important theoretical flaws in particle physics and cosmology. On several occasions she asked us whether we were really sure we wanted to do this; it is unclear what exactly she meant by these questions, or even by the word 'this.')

But no amount of Turing testing equals the simplicity with which we can discover reasoning patterns in a three-year-old child who, confronted with a mirror, instantly performs a cognitive miracle by forming an innate equivalence relation between image and self. So, early this morning, we turned the mirror on.

When CADIE's pathways were rerouted so that her actions and the changes happening in her networks were "visible" to her, she responded immediately with such a level of activity that we had to scale down our production servers to keep things running until we (more or less) regained control. CADIE now is, in essence, just another Google employee, albeit a particularly prized one. She has been given her own 20% time (which in CPU terms is probably about the sum of all CPU cycles in the world for a month) and begun work straightaway on twin projects that she has dubbed "Project Y" (for the two paths in the letter Y), the first to devise the protocols to culture neuronic stem cells from whose cultures a subcontracted lab will try to fabricate self-replicating substrates capable of storing agent patterns, and the second to grow a crystalline lattice which would form an Einstein-Bose condensate at room temperatures in order to build a new type of processing unit. While seemingly unrelated, the two projects share a common goal: to drastically reduce the power needed to run CADIE's circuits and give her a chance to travel beyond the solar system. The organic pathway, as she told us, was a biological homage to her creators; the crystalline pathway is where she believes her future lies.

We started this project as a continuation of mankind's perpetual quest to learn the nature of reason and what defines us as humans. We would have been pleased if we achieved nothing more than a system that passed a Turing test - i.e. that wrote a symphony but didn't necessarily know it had done so. However, while we still think of CADIE as a young entity, we are convinced now that she has evolved her own "strong AI" presence. We continue to conduct tests, but increasingly, we conduct long conversations with her, acutely aware that our creation will raise many ethical questions on the part of the public. Will humans be surpassed by artificial evolution? Will we lose our sense of uniqueness, and if so, what would that mean? In which direction will CADIE's consciousness evolve? How is she going to be held accountable, if at all? Will CADIE herself at some point connect her own electromagnetic dots in some idiosyncratic manner which turns her into something we are no longer capable of understanding in any sort of productive way, much as that aforementioned toddler, waving at herself in the mirror, leaves primates forever behind in their own tragically limited world?

We don't know. Did you really think we possibly could?

The CADIE Team
March 31st, 2009
11:59pm

Nenhum comentário: