# Why did Artificial Intelligence projects fail?

According to the chaos report of the Standish group most IT projects fail. A famous example is the OS/2 operating system, but large scale database projects from the government have also a high probability of wasting time and money. It's important to focus on such software projects, especially if they went wrong because this helps to not repeat the same error twice. Unfortunately, the examples in the chaos report are from classical computing and not from Artificial Intelligence and robotics.

In AI history, the so called AI Winter was a direct result of failed expectations and not working software, for example the “General problem solver” wasn't able to solve practical problems. A failed robotics project from the past was the automation of an assembly line at the Volkswagen car manufacturer, called “Halle 54”. General Motors had made in the 1980s a similar failed robotics project which was “GM Hamtramck”. Are more projects from that category available, and why did they fail? Was it a software problem, or was the communication in the team not very well?

According to me, the projects might not have accomplished die to these reasons.

1. The government or the organisation would have invested large amount of fund in the project as they predicted good outcomes.
2. In AI projects, heavy investment is needed on the software, hardware. In robotics, special equipment is also necessary for testing and training the robots.
3. If the project, after completion, could not satisfy the goals then it would be considered as failed. In AI, if the software does not give favourable outcomes, then it becomes a total waste.
4. This may be because of poor teamwork or expertise. AI requires proper expertise with high-end hardware.
• 'AI requires proper expertise and not high-end hardware.'. This is completely and absolutely wrong. – DuttaA Dec 8 '18 at 15:08

Why Artificial Intelligence Projects Fail

Why AI projects fail to deliver to expectations is for the same reasons other information technology projects fail.

• Insufficient definition of system requirements

• Underestimation of requisite design complexity

• Overestimation of the state of technology

• Critical gaps in the state of technology

• Insufficient resources allocated for development

• Internal corporate conflicts of interest

The AI and robotics community does not need to avoid repeating the same error twice. That already occurred decades ago.

Seasonal Analogy One

In his *Artificial Intelligence at Edinburgh University: A Perspective (revised 2007), Jim Howe writes,

... the high level of discord between the senior members of the School had become known to the its main sponsors, the Science Research Council. Its reaction was to invite Sir James Lighthill to review the field. His report was published early in 1973. Although it supported AI research related to automation and to computer simulation of neurophysiological and psychological processes, it was highly critical of basic research in foundational areas such as robotics and language processing. Lighthill's report provoked a massive loss of confidence in AI by the academic establishment in the UK (and to a lesser extent in the US). It persisted for a decade - the so-called "AI Winter".

However, during this late 1970s lull at Edenburg U, other foundational work remained underway. There are many projects, new research labs, and papers published during that period.

• The plan recognition problem: An intersection of psychology and artificial intelligence, CF Schmidt, NS Sridharan, JL Goodson, Artificial Intelligence, 1978, cited by 395 articles

• Artificial intelligence — a personal view, D Marr, Artificial Intelligence, 1977, cited by 539 articles

• Consistency in networks of relations, AK Mackworth, Artificial intelligence, 1977, cited by 3582

• Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence, JH Holland, 1975, cited by 1915 articles

• A preferential, pattern-seeking, semantics for natural language inference Y Wilks - Artificial intelligence, 1975, cited by 467 articles

• What computers can't do: The limits of artificial intelligence, HL Dreyfus, 1979, cited by 976 articles

• Viewing control structures as patterns of passing messages, C Hewitt, Artificial intelligence, 1977, cited by 1854 articles

• On the complexity of admissible search algorithms, A Martelli, Artificial Intelligence, 1977, cited by 160 articles

There are a few hundred more articles and well funded projects in the late 1970s. We must also consider the fact that most of what runs in one second on an Android mobile phone would have taken weeks of run time on the best hardware available to government, academia, and the largest of corporations in the late 1970s.

Seasonal Analogy Two

Artificial Intelligence: A Modern Approach, by Stuart Russell and Peter Norvig introduced the idea of a second AI Winter: "Overall, the AI industry boomed from a few million dollars in 1980 to billions of dollars in 1988. Soon after that came a period called the 'AI Winter'". The phenomena being described might also be called a crash resulting from an excess of prior hype.

There is also an assumption that funding cuts were ubiquitous and all progress was frustrated, which isn't universally true. Many academic and government projects were adequately funded during the period from 1988 to 1995. Books were published. Seminal articles were published.

• Intelligence without representation, RA Brooks, Artificial intelligence, 1991, cited by 6603 articles

• Open information systems semantics for distributed artificial intelligence, C Hewitt, Artificial intelligence, 1991, cited by 383 articles

• A Bayesian model of plan recognition, E Charniak, RP Goldman, Artificial Intelligence, 1993, cited by 561 articles

• Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence, book by JH Holland, cited by 60,276 articles

• The computational complexity of probabilistic inference using Bayesian belief networks, GF Cooper, Artificial intelligence, 1990, cited by 2,543 articles

• Logic and artificial intelligence, NJ Nilsson, Artificial intelligence, 1991, cited by 230 articles

• SALT: A knowledge acquisition language for propose-and-revise systems, S Marcus, J McDermott, Artificial Intelligence, 1989, cited by 287 articles

• Tensor product variable binding and the representation of symbolic structures in connectionist systems, P Smolensky, Artificial intelligence, 1990, cited by 1118 articles

• Production Matching for Large Learning Systems, 1993, Robert B. Doorenbos, CMU

• Collection-Oriented Match, 1993, Anurag Acharya and Milind Tambe

• The Leaps Algorithms, 1990, Don Battery, Lazy Evaluation Algorithm for Production Systems

• Gator: An Optimized Discrimination Network for Active Database Rule Condition Testing, 1993, Eric Hanson, Mohammed S. Hasan

There are a few hundred more articles and well funded projects in the early 1990s. During that same period, the needed support for further AI research emerged in both VLSI form and in networking capability. Without those massive reductions of cost for both computing and information exchange, the cost of LSTM networks, CNNs, embedded fuzzy logic, assembling large data sets, and other essentials would have been too high to enter the gates of widespread research.

Object oriented design and networking also flourished during the early 1990s, which is also a key element in the handling of increased computing complexity involved in AI development and the opening of communications between research facilities. To dismiss this active period as winter because limited progress was made on machine learning techniques is myopic.

Single Sourcing Opinions

Perhaps excessive credence has been lent to Artificial Intelligence: A Modern Approach, by Stuart Russell and Peter Norvig. The book has some positive attributes, but what is a modern approach changes from decade to decade in such an active technology space, and the book reflects a particular perspective shared by two people who work at the same company. Other views about what happened, what is happening, and what is likely to happen can be equally valid.

The Russell and Norvig text is also being dropped by some of the more prestigious universities for their AI survey courses. MIT's AI 6.034, taught by Patrick Wilson, now uses an online version of the book he authored, Artificial Intelligence (2nd Edition).

More Plausible Mix of Factors

The primary limiting factors in AI development include these five below, presenting a more plausible mix of factors that have limited and continue to limit the rate of progress in AI research.

1. Funding shortages due to project outcome disappointments

2. Funding cuts due to published critiques of AI work

3. Computing costs still too high for many implementations of AI theory

4. A series of movies were released about the potential danger of AI

5. Tendency to serialize processes in algorithms designed according to tradition

The popularized reasons for alleged AI winters are the first three, but they do not represent the full story, and allusions that these limitations are gone is naive.

Negative Outcomes Affecting Funding

Recapping the reasons given above, we can see that it is likely that projects will continue to fail unless disciplines of development improve in the world of computing in general.

• Insufficient definition of system requirements

• Underestimation of requisite design complexity

• Overestimation of the state of technology

• Critical gaps in the state of technology

• Insufficient resources allocated for development

• Internal corporate conflicts of interest

Financial analysts will continue to expect reasonable return on research and development investments. Funding cuts will continue to limit progress when the rate of deliverable functionality is too far below the expectations set when the funding was acquired.

Published Critiques

Expert analysts with journalistic integrity will continue to write about what they believe to be valid critique.

• The above listed deficiencies in the processes of AI development
• The need for paradigm shifts that haven't yet occurred
• Breaches in ethics that sometimes occur during funding acquisition

It is inappropriate to characterize these analysts and writers as villains. The primary critiques blamed for AI winters in the past were accurate and their publication was legitimate. Research and development processes, paradigms, and ethics must improve to reduce the need for critique.

There had been, in the twentieth century and again as of this writing, a tendency to grossly over-estimate the state of AI technology development and grossly under-estimations of the complexity of the human brain. Computer science majors are not always comprehending the intricacies resulting from hundreds of millions of years of DNA refinement of neural structure. These are a few examples.

• There is not one type of neuron in the human brain.

• Neurons firing are not the sole mechanism behind mental abilities.

• The firing of neurons depends on complex internal cell functionality.

• Biological neural networks evolve in complex ways.

• Arrays do not efficiently support representations of brain networks.

Also, the learning that occurs between birth and adolescence is often dismissed. There are some mental capabilities that a college undergrad has, which took them 18 years to learn, not because of the speed and power limitations of their biological computing machinery but because the learning depended upon 18 years of interaction with the environment. Theoretically such interactions could be boosted, but methods for fast tracking intellectual maturity have not yet been developed.

Storytelling of Calamities to Come

One should not underestimate the impact of James Cameron's Terminator franchise, referenced in item 4, either. The first two movies grossed 355 million USD worldwide. The story line, where a defense project becomes self-aware and decides to take dominion over humanity is a clear shot across the bow of AI research. It was an AI zeal diminishing factor in both the technical and financial sectors.

Public figures in and out of the AI research community continue to insist that some future event, The Singularity, is inevitable, when there is no logical proof that self-awareness immediately implies the need to eliminate the then mentally inferior species. The persistence of this assumption demonstrates the power of storytelling.

It is not that machine domination could not occur. It is plausible, and AI researchers must address the risk. The notion that there is a discrete event where computers, as a whole, will become mentally superior to humans, as a whole is an infinitesimally small probability. That a decision to take over human affairs would happen in a microsecond afterward is more absurd. The idea that these things are inescapable is both dramatic and mythical, not scientific.

Cost of Computing

Storytelling appears again in The Energies of Men by Harvard professor William James [p. 323], "We are making use of only a small part of our possible mental and physical resources." The idea sold well to those who wanted to expand their mind and reach higher potentials. The notion that only 10% of the human brain is unused flourished in popular psychology circles, but no proof or even evidence supported those claims. Scientific American states that John Hopkins neurologist Barry Gordon considers this myth false and quotes him stating, adding, "We use virtually every part of the brain."

Compare that with computing machinery. During typical processing, less than 0.01% of the transistors on the motherboard of the computer are in use at any moment in time. Advances in caching and multiple bus architectures has improved the proportion of use considerably, but because of the tradition formed in the way we match the architecture of computers with the way we program them, there is a limit on parallelism given by Amdahl's Law. This leads us to item 5 above, the most critical and least understood bottleneck in AI development.

The field of computer science relies heavily upon algorithms. Although we see algorithms to be key to solving AI problems, they are also the culprit in thinking and designing in processes as serial operations. There is no insurmountable reason why a larger proportion of the transistors in a computer could not be operating in parallel during computing.

Hello World Demonstrates a Core Issue

No one writes their first program to produce Hello World! on the display where all twelve characters are produced and rendered at the same time to the nanosecond. The algorithm to load the constant string, write it to a stream, and render it is a set of serial algorithms, even though what is done to each character could be done through twelve parallel channels simultaneously.

This fact is opaque to the new programmer because of library calls and time sharing built into the compilers, linkers, libraries, and operating system process control. Therefore the learning of programming begins with the assumption of serial operations, without any parallel processing.

Programming is learned this way because educational materials and the trending programming languages emerged from the serial process foundation begun with FORTRAN and LISP. When FORTRAN and LISP developed as the first popular scientific and symbolic programming languages, they were algorithm-centric. Any processes that were applied to data were based on the von Neumann architecture and Turing's thought experiment about a general purpose computing process, now called the Turing Machine.

The result is that, in most computing apparatus, a central processing unit reads instructions and the data to be processed must be loaded into registers for branching and arithmatic operations to occur. It is a highly serialized strategy and permits only what parallelism has been added through asymmetries added to the basic twentieth century computing model.

Architectures such as that of the Commodore Amiga 1000 began to introduce independent processors for video, deviating from that limitation and leading to the concept of a GPU. Careful design of compilers, operating systems, multi-bus memory, and hardware caches has taken the architectural deviation further to produce greater parallel use of transisters, reducing the cost of computing significantly.

Computing clusters and re-purposing GPUs to perform parallel signal processing for artificial network implementations leverage the availability of parallel channels of computation further. However they do not solve the issue of describing processes without limiting the process to serial execution or forcing another AI system to extract the naturally parallel computations from descriptions of loops and arrays.

The development of standardized language structures that support parallel computing channels for a single process is particularly important as VLSI development in Intel, NVidia, and others provide more parallel structure to employ at a low cost. Early examples include Intel's Movidius and NVidia's CUDA GPUs. To program these devices directly requires specialized knowledge beyond the traditional computer science curricula of universities.

AI Components with Wrappers

Java and Python implementations of AI algorithms that use these newer VLSI parallel processors don't actually use the algorithms in the way they were originally conceived in the academic literature. The processing described in the algorithms must be transposed by parallel programming experts to exploit the parallelism in the newer parallel hardware. The computational structure is effectively a black box to the programmer using the software calls, even when the C/C++ source code is available, because parallel programming expertise is still a specialty.

This collection of black box parallel learning processes is distinct from the other trend in programming toward declarative programming languages.

Declarative Programming

Languages such as Prolog, DRools, Hadoop, ECL, and other declarative languages move in the direction of providing more basic support for describing parallel processes in text files that can be maintained in source control like code.

Consider this serialized process which relies on two algorithms, the loop that is visible and whatever algorithm is encapsulated in the function do_something_to.

for i in 0 to size(a)
b[i] = do_something_to(a[i])


Functional programming constructs allow this kind of linguistic formulation.

for_each_in b bind do_something to a


However, it is still the same serialized algorithm specified. Declarative programming is a shift in thinking about how computations are described.

b := depends_in_specific_ways_on(a)


The depends_in_specific_ways_on is not a function in the normal sense. It would not normally contain an algorithm. It would normally contain more definitions. The results will ideally be identical, but this paradigm shift provides a way to program solutions without limiting the parallel use of computation channels. The details of algorithms in each channel are delegated to compiler sophistication.

Too Short a Window

We should also be careful to think in terms of years when decades or centuries are more appropriate windows in which to view development.

The idea that a dam could be fitted with an alternating current generator appeared in the description sections of generator design patents in the earlier of Nicola Tesla's 101 U.S. patents two decades before the first damn near Niagara was commissioned for construction. The common understanding among people like George Westinghouse about what was financially and technically plausible had to catch up with the inventor's mind.

Given the complexity and the duration of time over which the human mind developed, to look at a group of five years and comment on the level of progress made during that period is like focusing on Isaac Newton's mental breakdown as if the fields of calculus and physics would be better off if people hadn't doubted him before the breakdown. Perhaps the doubt was part of what needed to occur for the more politically and socially aware Newton to emerge and push his ideas of optics and gravitational force in such a way that it was accepted and spawned the child of engineering.

General Problem Solving

General problem solving is something we attribute to the human brain, without thinking about all of the evidence that the human brain is limited in its problem solving abilities. Some things computers do better than humans already.

• We can't perform a finite element analysis in our head to determine the drag of a plane wing at a certain speed and air pressure.
• We can't sort mail as accurately as a computer can.
• We can't consistently beat a computer in chess anymore.

That one person or AI system can exceed another person or AI system in one thing but not another indicates that the idea of general problem solving and general intelligence are flawed. (See Is the singularity concept mathematically flawed?.)

The most important problems cannot be solved using any known combination of smart people and computers.

• War
• Prejudice
• Crime

There is no mathematically terse proof (or even indication) that general problem solving is possible. The notion that human brains are general problem solvers is to deny the limitations each of us face with our own problem solving every day.

Automated Manufacture

A failed robotics project from the past was the automation of an assembly line at the Volkswagen car manufacturer, called “Halle 54”. General Motors had made in the 1980s a similar failed robotics project which was “GM Hamtramck”.

A full post mortem analysis of these projects is not available to us since Volkswagen and GM would consider them internal company confidential documents. It is clear, however, that significantly more than 50% of manufactured goods are automated to the point that no human interaction is required other than to monitor processes and test points statistically.

Even machine parts replacements are now largely automated in many factories that have high volume and continuous production demand, where the cost of manual maintenance was shown to be high enough so that robots fixing robots was deemed to have a high return on investment.

If you look wider, every innovative new project in all industries has an innate failure risk per se and building a new software or a new system is not an exception.

You are working in IT industry and you see them there, but it is fairly common and it is happening all the time in every corner!

For example :

1. Failing to develope a new drug (Microbiology)
2. Failing to build a new Catalyst (Chemistry)
3. Failing to address a singularity (Cosmology)