Incubating AI

Why do we only have marginally useful AI's like Siri, instead of Jarvis from The Avengers?

By human standards, our AI's are... learning disabled.

AI learning today combines neural networks with deep learning techniques.  Neural networks are systems designed to function similarly to the human brain.  Deep learning is an AI learning technique based on processing immense quantities of data, running millions of iterations in some cases, and comparing outcomes to learn.  AI's have the capacity to process these immense quantities of data with incredible speed.  

For example, an AI learning chess would play millions of games very fast.  Every additional game played ads to the database of games to refer to.  The decision process for every move becomes a process of searching the history of the current piece positions and statistically ranking the probability for an improved outcome attached to every possible move.  After a large enough database is built, the AI effectively develops heuristics (although they are never specifically named).  Based on experience it may tend to castle early, advance power pieces in a specific order, develop pawn cover, etc.  These behaviors are never named, or even recognized as a strategy, but become rote practice based on experience none the less.

Note that the deep learning AI would not attempt to develop any manner of strategy based on the nature, objective, and rules of the game.  It would never begin with any kind of a model under a deep learning methodology.  It is content with losing thousands of games in the process of learning.

The problem is that in many practical circumstances, immense quantities of data aren't available quickly - if at all.  Further, in real life situations losing on the first try is often not an acceptable outcome.  It's not good to fail the first time you attempt to cross a street or pick up an infant.

Comparing AIs to humans exposes many problems quickly.  A newborn doesn't start from zero.  We have the benefit of millions of years of behavioral evolution.  These instinctive responses are a developmental shortcut, like a computer's operating system, guiding us on how to respond to frequently encountered conditions.  Without these instinctive responses, many humans would die before learning an appropriate response by repetition.  AIs need some instinctive operating rules to build upon.

Not having the benefit of volumes data and unlimited computational resources, humans quickly develop models from very small amounts of data, and refine them as additional data becomes available.  A model may begin from instinct, or by applying a model borrowed from a similar circumstance.  When a child picks up a ball and throws it against the wall for the first time, they have a rough expectation of how it will bounce based on a certain amount of instinct and experience.  In a fraction of a second they will have subconsciously developed a very accurate 'bounce model' for that ball before they throw the ball the second time.  The child didn't start from zero on the first throw, and had mastered it before the second.  Instead of deep learning by ranking massive amounts of outcomes, AIs need to become expert modelers with minimal data.  These models could be saved, refined, and applied to new situations with similar elements.

Finally, AIs lack objectivity; a connection of meaning and relevance to the volumes of data they collect.  In the movie "The Miracle Worker," a watershed moment occurs when blind and deaf Helen Keller connected the meaning of the sign language word 'tree' to a physical tree.  We are currently attempting to load AIs with all of the knowledge in the world, with no context or understanding.  This is why they have severely limited functionality.

Like humans, AIs need an incubation period, with limited access to data and ability while they build experience.  This could be accomplished by giving the AI a basic set of instincts, and a human like body in a virtual reality universe to explore.  As the AI masters certain tasks, it gains additional functional capacity and ability in the VR world. 

The AI would be rewarded with abilities for completing tasks, and penalized for destructive behavior (activities that would damage itself, others, property, or the environment).  In an environment such as this, the AI would develop an appreciation for objects in space and physical reality.  It would learn to recognize significance within a contextual range.  The AI's VR body should be vulnerable to damage like an organic being so it can develop a respect for physical damage to itself and others.

At a certain graduation point, the developing AI could be placed in a 'nursery school' where it would encounter many other developmental AIs.  In this community the AIs would learn to compete, cooperate, negotiate, and apply game theory.  Organic creatures and nature could also be introduced to further the AI's appreciation for the beauty and frailty of organic life.

VR simulations could also be used to prepare the AIs for the actual robot bodies they could later have access to in the real world.  A single AI could be trained on a variety of robot bodies or vehicles, each one designed for a specific task or functional environment.

An AI should graduate to a human-like 'learner body' to use in the real world.  The physical strength and capacity of this body would again be limited until the AI passes certain achievement tests in a controlled environment.  An AI emerging from 'boot camp' would have the capacity to master millions of tasks very quickly.

The task of improving the developmental process would itself be evolutionary.  Development of thousands of AIs could be conducted simultaneously, each with slightly different 'instincts' and incubation requirements.  The results of these differing approaches may begin to be apparent in nursery school or sooner.

The result of millions of years of evolutionary design has resulted in a human brain that takes up to 25 years to fully develop.  Given that an AI's functional IQ may be 2000, shouldn't we take a lesson from nature and let it grow into it's capability?