Next:	Acknowledgements		Bookmark
Up:	Levels of Organization in General Intelligence		Monolithic
Prev:	3: Part III: Seed AI

Conclusion

"People are curious about how things began, and especially about the origins of things they deem important.  Besides satisfying such curiosity, accounts of origin may acquire broader theoretical or practical interest when they go beyond narrating historical accident, to impart insight into more enduring forces, tendencies, or sources from which the phenomena of interest more generally proceed.  Accounts of evolutionary adaptation do this when they explain how and why a complex adaptation first arose over time, or how and why it has been conserved since then, in terms of selection on heritable variation.  [...]  In such cases, evolutionary accounts of origin may provide much of what early Greek thinkers sought in an arche, or origin - a unified understanding of something's original formation, source of continuing existence, and underlying principle."
            -- Leonard D. Katz, ed., "Evolutionary Origins of Morality" [Katz00]

In seeking the arche of intelligence, I have striven to neither overstate nor understate its elegance.  The central shape of cognition is a messy 4D object that casts the thousand subfields of cognitive science as 3D shadows.  Using the relative handful of fields with which I have some small acquaintance, I have tried to arrive at a central shape which is no more and no less coherent than we would expect of evolution as a designer.

I have used the levels of organization as structural support for the theory, but have tried to avoid turning the levels of organization into Aristotelian straitjackets - permitting discussion of "beliefs", cognitive content that combines the nature of concept structures and learned complexity; or discussion of "sequiturs", brainware adaptations whose function is best understood on the thought level.  The levels of organization are visibly pregnant with evolvability and plead to be fit into specific accounts of human evolution - but this does not mean that our evolutionary history enacted a formal progress through Modalities, Concepts, and Thoughts, with each level finished and complete before moving on to the next.  The levels of organization structure the functional decomposition of intelligence; they are not in themselves such a decomposition.  Similarly, the levels of organization structure accounts of human evolution without being in themselves an account of evolution.  We should not say that Thoughts evolved from Concepts; rather, we should consider a specific thought-level function and ask which specific concept-level functions are necessary and preadaptive for its evolution.

In building this theory, I have tried to avoid those psychological sources of error that I believe have given rise to past failures in AI; physics envy, Aristotelian straitjackets, magical analogies with human intelligence, and others too numerous to list.  I have tried to give some explanation of past failures of AI, not just in terms of "This is the magic key we were missing all along (take two)", but in terms of "This is what the past researchers were looking at when they made the oversimplification, these are the psychological forces underlying the initial oversimplification and its subsequent social propagation, and this explains the functional consequences of the oversimplification in terms of the specific subsequent results as they appeared to a human observer."  Or so I would like to say, but alas, I had no room in this paper for such a complete account.  Nonetheless I have tried, not only to give an account of some of AI's past failures, but also to give an account of how successive failures tried and failed to account for past failures.  I have only discussed a few of the best-known and most-studied AI pathologies, such as the "symbol grounding problem" and "common-sense problem", but in doing so, I have tried to give accounts of their specific effects and specific origins.

Despite AI's repeated failures, and despite even AI's repeated failed attempts to dig itself out from under past failures, AI still has not dug itself in so deep that no possible new theory could dig itself out.  If you show that a new theory does not contain a set of causes of failure in past theories - where the causes of failure include both surface scientific errors and underlying psychological errors, and these causes are together sufficient to account for observed pathologies - then this does not prove you have identified all the old causes of failure, or prove that the new theory will succeed, but it is sufficient to set the new approach aside from aversive reinforcement on past attempts.  I can't promise that DGI will succeed - but I believe that even if DGI is slain, it won't be the AI dragon that slays it, but a new and different dragon.  At the least I hope I have shown that, as a new approach, DGI-based seed AI is different enough to be worth trying.

As presented here, the theory of DGI has a great deal of potential for expansion.  To put it less kindly, the present paper is far too short.  The paper gives a descriptive rather than a constructive account of a functional decomposition of intelligence; the paper tries to show evolvability, but does not give a specific account of hominid evolution; the paper analyzes a few examples of past failures but does not fully reframe the history of AI.  I particularly regret that the paper fails to give the amount of background explanation that is usually considered standard for interdisciplinary explanations.  In assembling the pieces of the puzzle, I have not been able to explain each of the pieces for those unfamiliar with it.  I have been forced to the opposite extreme.  On more than one occasion I have compressed someone else's entire lifework into one sentence and a bibliographic reference, treating it as a jigsaw piece to be snapped in without further explanation.

At this point in the ritual progress of a general theory of cognition, there are two possible paths forward.  One can embrace the test of fire in evolutionary psychology, cognitive psychology, and neuroscience, and try to show that the proposed new explanation is the most probable explanation for previously known evidence, and that it makes useful new predictions.  Or, one can embrace the test of fire in Artificial Intelligence and try to build a mind.  I intend to take the latter path as soon as my host organization finds funding, but this may subtract from the time available to mend the gaps in the present paper.  Hopefully my efforts in this paper will serve to argue that DGI is promising enough to be worth the significant funding needed for the acid test of building AI.

In today's world it is commonly acknowledged that we have a responsibility to discuss the moral and ethical questions raised by our work.  I would take this a step farther and say that we not only have a responsibility to discuss those questions, but also to arrive at interim answers and guide our actions based on those answers - still expecting future improvements to the ethical model, but also willing to take action based on the best current answers.  Artificial Intelligence is too profound a matter for us to have no better reply to such pointed questions as "Why?" than "Because we can!" or "I've got to make a living somehow."  If Homo sapiens sapiens is a noncentral and nonoptimal special case of intelligence, then a world full of nothing but Homo sapiens sapiens is not necessarily the happiest world we could live in.  For the last fifty thousand years, we've been trying to solve the problems of the world with Homo sapiens sapiens intelligence.  We've made a lot of progress, but there are also problems that we've hit and bounced.  Maybe it's time to use a bigger hammer.

Next:	Acknowledgements
Up:	Levels of Organization in General Intelligence
Prev:	3: Part III: Seed AI