An AI Winter

In 1973, the United States and United Kingdom governments decided to withdraw funding from AI projects. The result was an AI winter, where new development largely came to a slow crawl. An AI winter is a period where general interest in the AI space tends to wane, both in terms of research and funding, often as a result of disenchantment in what the community expected it to achieve as compared to the actual progress that was made.10 In hindsight, people attribute this slump more to unrealistic expectations rather than an actual lack of progress.11 Buoyed by the early wins over the course of the decade, scientists had come to expect that a lot more could be achieved in the next few years than what was actually possible given the resources and technology available at that time. For instance, in 1970, some researchers were talking about creating a machine with the same level of general intelligence as the average human being by the end of the decade—something the world’s best minds are still working on five decades later. Once investors started to realise that this was not feasible, interest started waning, and so did the funds available for research.12

Lack of computing power was one of the biggest hurdles to pursuing deeper research within AI. Teaching computers how to learn required them to be able to handle a large volume of data, which simply wasn’t possible for the machines of that time. In the 1970s, computers were not powerful enough or fast enough to store and process all the information they needed to complete these tasks. Despite advancements being made in computing and computer science, the late 1970s were slow when it came to new developments in the field of AI. While processors and storage systems kept getting better and more powerful, research focus had shifted away from AI as researchers failed to achieve the lofty, and possibly unreasonable, targets they had set themselves.

However, this period saw the focus shift to other technologies, which ultimately resulted in the development of new software systems and RPA. RPA allows companies to automate low-value mundane tasks, which frees up employees to work on more value-added activities. Data entry, invoice processing, basic customer service tasks and repetitive tasks like payroll processing were among the earliest tasks that were automated. A large proportion of mid- and small-sized organisations have automated these tasks, showing just how widespread process automation has become. In addition to speeding up the process, the use of automation has also improved efficiency to an unprecedented level, dropping error rates to near zero. This was among the earliest instances of technology doing a human’s work, and not surprisingly, there was a great deal of consternation among employees about their jobs being taken away by machines.

The 1970s and the 1980s saw the advent of programmable logic computers in industrial automation, further enhancing the role of technology on the shop floor. Industrial robots made an appearance in Europe and the United States of America. In 1969, Victor Scheinman invented the ‘Stanford Arm’, which could be used to carry out basic tasks, eventually paving the way for robots to perform more complex tasks like assembly and welding.13 European firms like ABB Robotics started producing industrial robots, leading to a further spread of industrial automation. This was the start of an era of significant productivity gains and efficiency on the shop floor as companies started doing more with less and speeding up production. Assembly lines started getting automated, and people started getting trained on how to use the machines effectively to do their jobs better. Parallelly, researchers kept making incremental progress in AI, with the period between 1980 and 1987 being another boom era for AI. The early 1980s saw the first commercial deployment of an expert system, which brought a lot more attention to what they were capable of.

This was the R1 or XCON, which was developed at Carnegie Mellon University by John McDermott. XCON, or expert configurer, was introduced by the Digital Equipment Corporation in 1982 to configure computer orders and improve their accuracy based on customer orders. Four years later, Digital Equipment said that it had saved US$40 million a year through the use of the R1 system.14

What struck a final death blow to the exuberance around AI in the 1980s was the decision of Defense Advanced Research Projects Agency (DARPA) to stop funding AI research in 1987. Prior to that, the government agency had been among the biggest funders of AI in the United States, having invested significantly in research—US$100 million in 1985 alone. However, by the late 1980s, DARPA started slashing funding for this project after failing to see any significant progress. The withdrawal of support from DARPA eventually resulted in the start of the second AI winter, which lasted well into the mid-1990s.15

At around the same time, the Japanese government shut down its Fifth Generation Computer Systems initiative. Japan had been among the biggest investors towards commercial use cases of AI and expert systems. However, the heavy costs associated with the project coupled with the results not matching up to expectations resulted in the Japanese government pulling the plug on the initiative after a decade. This marked the end and failure of the parallel processing approach to AI.16

Computing devices from companies like Apple and IBM were starting to hit the market with far better processing capabilities than the specialised LISP systems. As these desktop computers began gaining market share, it resulted in interest rapidly declining in the LISP systems for AI. As with the first AI winter, this period, too, saw a lull in funding and a drop in broader interest in AI projects, even as researchers continued to make progress on some key parameters. One among these was the development of probabilistic reasoning, which led to a revival of interest in AI projects, opening up a new approach towards AI.

Meanwhile, industrial automation continued to make deep inroads into all areas of the manufacturing process across industry sectors and geographies. The drop in microprocessor prices and spread of personal computers helped this spread further, leading to the advent of RPA. The evolution of RPA meant that automation was now impacting office workers and not just the shop floor. So far, automation had been synonymous with the use of machines to automate processes. The progress made in computing, coupled with the basics of automation, resulted in RPA, where over time, bots could be programmed to carry out repetitive and routine tasks. In addition to cost and time savings, this also freed up people to work on more value-creating tasks. The way these bots have evolved, it is now possible to program the same bot to carry out multiple tasks as well. These bots are trained to interact with the IT systems or applications in the same way a human would and are increasingly learning to respond better to natural language inputs. They continue to get better, with AI systems now being integrated into them.17

Meanwhile, for researchers working on AI, the shift towards probabilistic reasoning marked a fundamental shift in how AI systems addressed problems. This now allowed the AI system to handle uncertainty and ambiguity in decision-making, something that had been missing in the past. Taking off from the work done by Joseph Weizenbaum with ELIZA in the 1960s, Richard Wallace, a computer scientist, created a chatbot called ALICE (Artificial Linguistic Internet Computer Entity). ALICE had been programmed with sufficient basic knowledge to be able to converse with another human. One significant difference was that it had also been fed a sample data collection of natural language, which allowed it to elaborate the meaning of a phrase through specific terms.18

The most significant development in the 1990s towards the growth of AI was the development of NLP capabilities. This, in turn, led to progress across various areas of AI development. This growth was fairly widespread and not just restricted to a few universities or AI Labs. In 1997, Dragon Systems in Massachusetts, United States of America, released NaturallySpeaking 1.0, a speech recognition software prototype.19 This software included standard natural language and was the first truly accessible computer dictation software. Microsoft’s Windows integrated this into its offerings, allowing users to dictate rather than type their inputs. The biggest win for AI in the 1990s came from IBM, with then world champion Kasparov walking off from his match with the Deep Blue supercomputer in 1997. While this might seem trivial in the context of what computers and supercomputers can do today, in the 1990s, this was a massive event.