For the past twenty years, the approach towards creating machine intelligence has been reminiscent of the way humans took flight; First, our scientists looked to nature for design templates, then set to work mechanically duplicating the aerodynamic capabilities of a bird, substituting raw power in all the places where they failed to match elegance of design. The result? A line of hulking machines that can fly faster than any bird, but could not begin to approach their flexibility of movement. Then the helicopter was designed to improve dynamic agility, but the architecture sacrificed speed in order to achieve it. Similarly, computer science has designed systems such as "Deep Blue" - a machine so intelligent that it can beat any living human in a game of chess, but for all its intelligence, it is unable to learn how to play a simple game of checkers.

An archetypical example of this kind of AI is a software package developed by UCLA professor Richard Korf to solve the famous Rubik's Cube (citeseer.nj.nec.com/cachedpage/157215/1). Since this puzzle has frustrated millions of fairly intelligent humans, this program is often seen as a dramatic illustration of the potential capabilities of machine intelligence: despite the enormous complexity of the problem (a state space on the order of 10^19 possible configurations), Korf's software can solve almost any initial state in under twenty moves.

These are brilliant and sophisticated algorithms, but the brilliance and sophistication lies in the skill of the programmer, not in the inherent capabilities of the program itself, which relies on brute force computation (evaluating hundreds of billions of configurations per run), a huge (42 Mbyte) lookup table database, and vast computer resources (several weeks on a workstation). Certainly these programs are not thinking, not learning, not adapting, and not behaving in any way that really resembles intelligence. In fact they are not qualitatively any more intelligent than an old hand held calculator that can instantly evaluate nine digit multiplications. These are computer programs, written in ordinary computer languages, performing more or less ordinary computations, written by extraordinarily smart programmers.

Today almost all of AI research is focused on the one domain in which its practitioners have achieved success: Expert systems. But referring to these systems as a branch of AI is a popular miscategorization, for nothing distinguishes these methods from regular computer programming or statistical analyses. Many of their capabilities are impressive, but all of them are highly specialized to perform a single task, and none of them has any kind of generic capacity to learn new tasks or function in new environments. Expert systems can be extremely valuable as long as they are focused on solving problems that are narrow and deep (e.g., medical symptom analyzers), and general information systems are useful as long as the information requests are wide and shallow (e.g., search engines).

The widely publicized Cyc project (www.cyc.com) is another example of today's brand of artificial intelligence. Cyc is a growing database of over a million facts that collectively embody much of the common sense knowledge possessed by humans. Each fact, or rule, is manually entered into the database by its programmers and endows Cyc with tidbits of knowledge such as "water is wet" and "when you let go of things they usually fall." By making connections between input data and the various facts in its database Cyc is able to draw inferences and even assist in providing solutions to some planning problems.

An extensible warehouse of knowledge such as Cyc will undoubtedly prove to be a useful tool for many applications. But the dizzying claims of its creators that Cyc has a capacity for understanding and reasoning, and may eventually surpass human capacity for intelligence, ethics and compassion are the kind of rhetoric that badly damages the credibility of AI research, and probably misdirects resources and attention away from real attempts to design artificial minds.

One branch of AI that does try to create systems that more closely resemble our brains in their internal workings is the field of artificial neural networks (ANNs). These systems have achieved some success in their ability to recognize and classify data inputs, most notably in hand writing recognition, and are very robust with respect to sensitivity to noisy data. However, the functional resemblance of these networks to biological neural networks is only superficial; these models are strikingly unbrainlike in their architecture and operation. They are nothing essentially more than correlation algorithms that determine the best match between input data and a precoded set of already known data.

Moreover, they are incapable of ongoing experiential learning. There are two phases of network operation: programing (also called training), and execution. When the programing phase is complete, learning ends, and the processing capability of the ANN is permanently frozen at its current level of competence. If the network has been trained to optically recognize characters, for instance, the introduction of a new character into the alphabet would require that the network be wiped clean and trained from scratch. For this reason, ANNs have not made very much progress over other approaches to AI. While the programming process is different (entering a sequence of coded instructions vs. applying a statistical correlation function), the result is the same - a static system that executes an exact and predetermined series of calculations.

Some ANN modelers incorporate an evolutionary process into the training phase, but like their conventional counterparts, once the training is complete, these networks will stop learning anything new. The Robokoneko project (www.genobyte.com/robokoneko.html) is a prominent example of this training method. Robokoneko is a virtual cat that will someday be capable of walking, sitting upright, playing with a ball, and a few other kittenlike behaviors. The cat will be controlled by an artificial brain of 75 million neurons. Its behaviors will be acquired through an evolutionary process in which a large virtual population of these brains is subjected to many generations of mutation, replication, and finally artificial selection by the trainers for those with the most desirable behaviors. After tens of thousand of these generations it is hoped that Robokoneko will have sufficiently developed skills. But as soon as this training is done, the evolutionary process comes to an end, and the best-behaving brain is chosen for the real cat. That cat's brain will not evolve any further, and so will never be able to learn anything new or improve upon its current performance. Biological evolution doubtlessly does have an analog in cognition and learning, but that process must be much more autonomous than in the Robokoneko project, and must not ever terminate.

Because so many of the capabilities usually associated with human intelligence are unattainable to ANN technology, many AI researchers argue that neural models alone will never be sufficient to simulate intelligence. Tragically for the AI community, criticisms against ANNs have been mistakenly but universally extrapolated as fundamental flaws that would apply to any neural-based learning system. By failing to develop neural models any more complex or interesting than statistical correlation calculators, ANN researchers have unwittingly erected their own straw man for traditional AI researchers to attack. The result is a self-reinforcing failure of progress in the development of new neural models. But we know from the example of our own minds that some kinds of neural systems definitely are capable of intelligence. Biology is an excellent guide: Through studying the brains of intelligent living organisms, we have learned that highly robust and adaptive phenomena emerge from the hierarchical self-architecture of much simpler elements.

Current approaches to AI have failed because none of them is essentially different from the original computing paradigm: programming a highly specific series of predefined instructions for the computer to follow in precise manner and exact order. Intelligence as we understand it cannot be achieved using this approach to problem solving. The cognitive functions of the human brain are literally defined by adaptive self-organization, and without this critical foundation, attempts to create a conscious system will continue to miss the mark.

Intelligence depends on the ability to successfully integrate data from the environment into previously stored classifications and generalizations - and the ability to generate new classifications and generalizations and integrate those with successively higher hierarchical levels of metaclassifications. The mind performs a very sophisticated kind of pattern matching, perceiving and recognizing patterns from the external environment and from within its own reservoir of accumulated experience. This pattern matching is an evolutionary process: new patterns are created, old patterns mutate and adapt, and successful patterns become further integrated into a rich ecology of interacting and multidimensional patterns.

Since the mind's capacity for pattern matching is a robust, self-architecting, evolutionary and hierarchical process, we think any intelligent system - artificial or otherwise - must be composed of elementary components, and elementary, though dynamic, interactions between them. The capacity to store concepts and interrelate them will be emergent from these dynamic interactions. As networks of interacting components begin to interact with each other, and successively more intricate interactions emerge among those, so emerges higher cognitive functioning in the form of successively more intricate memory traces and information processing.

The emergence of complex phenomena from the interactions of simple components is illustrated in every facet of biological evolution. An evolved sequence of interactions among DNA molecules forms a more complex and more dynamic functional unit, the gene. An adaptive and coordinated assemblage of genes unites to form an even more complex and functional unit, the genome. Organisms assemble to form populations, populations to form species, and species to form ecosystems, with each level attaining new functionality emergent from that below it. Every level of the biological hierarchy is an active target for natural selection, and selection among competing entities at any one level has the potential to cause sorting (differential survival and reproduction) at any other level, with cascades of upward and downward causations rippling through the entire network of life.

The result is a highly adapted, highly complex, highly robust and fault tolerant ecosystem, designed from the bottom up, with every interacting component - from DNA molecules to groups of species - seamlessly integrated into the dynamic whole. As a paradigm for engineering design, this model has never been surpassed.

Just as in the evolution of life, the elementary components in our brains emerge to form highly adapted, highly complex, highly robust and fault tolerant functional systems - our minds. Coordinate interactions among neurons form a more complex and functional unit: a simple neural circuit. Closed circuits of linked neurons form an even more complex, functional and dynamic unit: feedback loops. Higher circuit dimensions (metaloops), made up of conflations of interacting loops, represent higher levels of integration and coordination of stored memories. As the network forms loops and pathways that interact with each other by virtue of their shared connections, metaloops themselves become interrelated in higher hierarchical dimensions of metaloops. Ultimately, in a network of sufficient complexity both in terms of the number of neurons and connections and the number of interrelated loops and pathways, every loop will become incorporated into a single unified cohesive metaloop, embodying the collective repertoire of patterns - memories, classifications, and generalizations - gleaned throughout its existence. And this processes continues and grows throughout the life of that mind.

Only with an emergent hierarchical system based on elementary components can true machine intelligence ever be realized. Yet the world of computing remains transfixed on the opposite approach: Top-down design of phenomenally complex programs and algorithms that are very adept at specific tasks, but will never develop any capabilities beyond them. These programs can solve a Rubik's Cube or tell you that if you drop a glass of water the floor will become wet, but they will never acquire such knowledge on their own. And even when successful, these complex programs are notoriously fault sensitive and unstable.

An intelligent system must be neural in that it must consist of elementary components and dynamic interactions between them, and must manifest the emergence of successively more complex, dynamic, and adaptive structures and interactions from these elementary components. As with all forms of biological evolution, adaptation must occur autonomously and continuously. But while biological adaptation is an emergent process observable in large populations of ephemeral components that live, die, and become replaced, neural learning is an emergent process of gradual changes in the connections between the neural loops and metaloops. While these connections undergo many changes, the neuron itself is relatively permanent. Learning occurs within an individual mind, which responds to contingencies of positive and negative reinforcement, automatically adapting to its environment, automatically learning and optimizing new behaviors, automatically maintaining internal consistency. The process is often slow, but the outcome is extremely powerful.

Slowness of learning, however, will soon be a handicap unique to the animal kingdom: biological neurons fire at maximal rates on the order of a thousand times per second; silicon neurons made up of transistors can fire at clock speeds of a billion times per second. This is a difference in learning and processing speed of a factor of one million. And with the steady march of technological advancement, this difference in learning speed between humans and the robots built by them is certain to become astronomical.

Animals are further limited in their intellectual capacity by their number of neurons, number of connections per neuron, their organization and circuitry, and perhaps countless other factors. While many of these parameters have undoubtedly been shaped and optimized by millions of generations of evolution, they surely face an enormity of biochemical obstacles and resource constraints that silicon neurons will not. An artificial mind could enjoy boundless architectural flexibility in the number of neurons, the number of connections between them, and their dynamics of interaction.

This means that unlike every other AI project ever envisioned, a neural learning system could immediately be scaled to any desired size and any desired performance level. The bottom-up design approach means there is no ceiling: Through the emergence of complex relationships formed between simple components, anything is achievable. Elucidating the fundamental nature of these simple components is the only challenge separating today's technology from an overnight revolution.

The potential for commercial application is without end. Virtually every where current computer technology fails, an artificially intelligent system will excel. The top-down design of today's technology will pale to the bottom-up design of the most complex and adaptive systems every known. And truly intelligent machines may well come to surpass even the biological models on which they were based.

We have begun developing and training our networks within easy domains such as mazes, puzzles, and games. But our goal is not to construct a program that can solve a Rubik's Cube - instead, we intend to evolve a conscious system which, when confronted with a Rubik's Cube, wants to solve it, and understands what needs to be done. We seek to answer the question of consciousness - to explain the metaphysical nature of being - not by attempting to manufacture a sentient robot, but by giving birth to a virtual embryo, nurturing it to maturity, and setting it free to make order out of chaos.

Our goals are ambitious, but we firmly believe that immersed in a high-intensity, intellectually synergistic, singularly focused environment, the right team can make this happen. We must be ready to think outside the box. We must be ready to reconstruct the box. We know why others before us have failed. And we continue to believe that the fundamental building blocks of our minds are not beyond our intellectual grasp.