The failures of artificial intelligence (AI) are a remarkable disappointment to science, industry and government. In this paper, we explore both the successes and failures of AI.
A machine that acts like a human has interested even the non-scientist. The discussions and dreams that began in the 1930s still continue after 80 years. The ENIAC tube computer of John Mauchley in the 1940s at last suggested an avenue that could make this possible.
Academia wrote multiple papers and proposals written in the 1950’s for this “artificial intelligence” (AI). They considered how AI might look, be created, be ruled or governed, be recognized and even be applied. Novels created AI of exceeding stature and expectation.
Now two decades into the next century, AI has yielded only minimal progress. It produced few compelling stories to match the promises of the 1950s. It has done little, much less enter the world where it must obey the “three prime laws of robotics”.
This paper poses questions; obviously, I have some answers in mind. My goal is to first show awareness: “Houston, we have a problem!” It is rather unfair to critique a problem without also posing the likely solution.
Subsequent papers examine worthwhile areas of Artificial Intelligence and explore the most-likely alternative to AI. Perhaps there is something less artificial, less program-scripted, something more human than artificial. There is.
Get one thing from the start: The human brain runs on carbon-based sugar and consumes some 65 Watts of power. It resides in a fixed volume of space. Like the heart that never ceases pumping for a lifetime, this brain never stands down or reboots.
That is your baseline, something meticulously and intentionally designed. Suggesting that this assembly self-assembled from slime scrapes the depths of human arrogance. How, then, can it be recreated?
What Has AI Research Produced?
First, let’s examine some of successes from AI research. In fairness, it has produced a small number of useful concepts and products, including:
- Fuzzy logic of the 1980s
- A general system of self-organizing “neurons” that maps inputs and expected results to create deterministic outputs.
- Genetic algorithms, permutations of machine-generated code. It selects the code that has produced the most optimum results.
- A general correlation model based on the brain’s neural cortex that has come to be known as Machine Learning. Variations of it can recognize patterns, whether visual, aural or data-centric. This has become an invaluable asset for data reduction and interpretation.
- The Darpa-funded “Web 2.0”.
- The Web 2.0 development of RDF and its “Triple Store” methodology for data storage. A shift from relational databases, it is also at the heart of IBM’s Watson computer. It is a computationally-intensive system consuming somewhat more than 65 Watts.
- System’s based on Machine Learning – such as created by Google – for vastly improved language translation.
- To artificial intelligence, ChatGPT brings a level of English-based interaction. It currently has its quirks but holds promise with additional engineering.
Each above area suggests some small area of human brain function. Some of it – safe to say – has revolutionized segments of the data industry, a wonderful outcome. But, it’s not really intelligence.
Are we serious about this? Then honesty is needed in the assessment. Most of artificial intelligence seems like raw computing power to me, adequately governed by Moore’s Law. None of it is seems governed by the “three prime laws of robotics,” nor is even serious enough to intrude on the applicability of those ‘prime laws.’
Why Has AI Failed?
AI’s record of failure is seen in Darpa’s history of technology requests. Darpa – and the CIA – consistently ask for software and analytics that distill information into a form suitable for presentation to the human analyst. The analyst then assesses it and makes recommendations or decisions. (This is the essential process of Command and Control, aggregation of sensor data for the human analyst.)
Understand this: Were a true machine intelligence available, the Department of Defense (DoD) and like agencies would be letting it do the final analysis, making decisions. They would be seeking industry creation of tools to analyze the validity of the machine’s answer. They would be seeking tools to assess relevance of the supporting back-stories that such machine-based agents would provide.
Contributors to the Failures of Artificial Intelligence
The failures of artificial intelligence to produce substantive machine intelligence is a continuing matter. The main contributing causes are institutional. Other causes stem from the nature of the undertaking.
- One clear starter is DoD’s (make that largely Darpa‘s) monetary focus on “bio-mimetic” AI systems. Academia has taken the money and duly produced. They haven’t produced real AI, though.
- A psycho-mimetic model is required, not a bio-mimetic model. The bio-mimetic model has led researchers into the weeds. Remember, this is a silicon-based emulation, not a carbon-based emulation.
- Human brain function is difficult enough to visualize. Current AI methods have become so complex as to not be understandable. Recently, renewed calls are being made for an “understandable AI”.
- There is relatively low core agreement among academics as to what constitutes human thought process.
- There has been institutional focus on the validity, use and need for “rules” that define how the brain “infers” new information from old. This is a limiting model. Rules-based implementations and their maintenance are complex and brittle. The diligent care needed for them makes achieving the human-like goals difficult. The supporting tools have gotten in the way of the needed outcomes.
- Emulated emotion is conspicuously absent, as if that would drive towards better assessments and decisions by the machine.
- There appears a lack of (academic and medical) understanding of emotion and its subsequent impact on the processes of memory, motivation and decisions.
- There has been no agreed common metric of human personality, such that it be emulated. Common temperament definitions are available but largely dismissed in emulating human intelligence.
- Human emulation intersects the domains of psychology, biology, linguistics, software, electronics and (effective) education. This intersection is largely missing. America demonstrates little ability to educate humans very well, and much less so, a machine agent.
Anyone playing on the forefronts of human-like machine intelligence would do well to examine the implications of humility. There is remarkable lack of it, particularly in the advisers to government agencies. Human ego (such as the old not-invented-here impediments of 1980’s industry) appears to play a significant blocking role in the creation of human-like intelligence.
We have explored the history of artificial intelligences and the issues surrounding its development. Is it possible to significantly change the trajectory of machine intelligence?
Alternatives to Artificial Intelligence
Focusing on the last 20 years, 1998 brought Darpa funding that explored graph theory. They showered the universities with it, seeking an alternative to the relational database. The alternative would be built on graph theory.
A major outcome was the formulation of the Resource Description Framework (RDF), on which a new “Web 2.0” was to be built. Their alternative to the SQL query language became the SPARQL language.
Darpa’s hopes for basing machine intelligence on it were built upon OWL, an “inferencing” language layered atop SPARQL. The simplistic RDF structure did not readily support inferencing with OWL without creating complex rules.
The initial restrictions placed on RDF definition were both its strengths and its weak point of failure.
The academic fascination with “inferencing” is a good example of stopping near the “6% solution” in a like manner that Machine Learning does. Of themselves, neither are of significant value in building human-like mind processes. Interestingly, they are more useful for data reduction than for determining the implications of data or events. They are wonderful platforms for powerful sensory capacity, but act as external senses, not as a brain.
We proclaim that RDF restrictions killed both OWL and RDF-based word ontologies. The triple-store and its “inferencing” are dead. University interest and taxpayer funds might revive them, but markets generally knows when to let an idea go.
Separate from governmental sponsorship, during the last 10 years, industry has created machine intelligence mechanisms that fully support emotion. This too has been based on some form of graph theory.
The initial systems were rather brittle. For example, if the supplied word ontology was not precisely correct, they would not recognize the nuances of English or what they imply. Even so, they demonstrated the capacity to autonomously analyze events for their implications.
Trained primarily with just the English backstories about events of that nature, they were able to outpace their human analyst counterparts.
One system under development shows promise. It is an agent with linguistics-based Human-Level Intelligence (HLI), and again rooted in graph theory. Its use and interconnect of a graph-node based neuron is a fundamentally different architecture and node form than the earlier efforts such as those at Neuric.
The HLI agent has a very robust parser and readily deals with cultural idioms and cultural matters. LBI can prevail in the face of the failures of artificial intelligence.
Summary
In this paper we have discussed the successes and failures of Artificial Intelligence. We highlighted some of the causes for the failures, both institutional-based issues and the “nature of human nature.”
Future papers will discuss matters peculiar to the alternative “Human-Level Intelligence” (HLI) systems most likely to replace AI. Artificial is always dropped when it becomes obvious you have the real thing.