Renewed Interest and Big Wins

Renewed Interest and Big Wins

The turn of the century brought with it a renewed interest in AI and possible applications of this technology. The early wins of the late 1990s had once again started to generate interest in AI the world over. Apart from researchers and academics, companies, too, were getting curious about how they could use this new technology to their advantage. The early 2000s also saw a huge spike in interest in automation software as companies, flush from beating the Y2K bug, were keen on continuing to invest in software that could boost productivity and efficiency. The focus shifted from industrial automation, which was largely focused on bringing in efficiencies in manufacturing, to intelligent automation.

While the two terms, automation and AI, are increasingly being used interchangeably, it is incorrect to do so. An automation system does not necessarily have to use AI or machine learning to function. Similarly, not all AI tools are used to automate a process. No doubt there are overlaps, but by and large, the two are separate kinds of technology. When a service provider like Amazon or Netflix starts to provide personalised suggestions based on your personal history, it is an instance of automation using AI to work effectively. Amazon using machine learning to automatically tag key characteristics in an item on the website once again is a combination of AI and automation. AI programs started getting better in the early 2000s and had enough intelligence to respond to basic questions.

The development that really made the world sit up and take notice was Apple introducing its voice chat assistant Siri in 2011.20 Suddenly, AI-based conversations and assistance were accessible to pretty much anyone who had an iPhone, which was a reasonable share of the population, especially in the United States of America. AI moved from being something researchers talked about in academic journals to something a layperson had in his or her pocket.

Advent of Machine Learning

The biggest boom around AI development happened after 2010. With the widespread use of the internet all over the world, people began generating massive amounts of data, which, in turn, were available to researchers to train these neural networks. Algorithms could be trained on more accurate image classification and recognition based on the thousands, if not millions, of training samples available.

Even as these developments gained pace, some researchers started looking beyond just the tech progress that was being made. Most notable among these was Margaret Boden, a professor of cognitive science at the University of Sussex, who started exploring the philosophical aspects of AI. Her work delved into the nature of creativity in machines and the boundaries of AI.21 Around the same time, hardware development picked up pace. The launch of very-high-efficiency graphics cards and processors helped in the acceleration of learning algorithms. Computers that would earlier take weeks to process certain data samples could now do it within minutes. This was unheard of in the history of computing.

In 2012, Google Search Labs (Google X) could train its AI system to recognise cats on a video. Carrying out what today seems like a simple task, what with cell phones accurately recognising and tagging different members of your family in your photo album, took over 16,000 processors back in 2012. It was still a giant step when it came to the progress made by AI systems, with a machine being smart enough to recognise something and distinguish it from other similar samples.22

All this was what was happening behind the scenes. What consumers and the average person saw was Watson, a computer answering system created by IBM, beating two of the greatest Jeopardy champions in 2011. Jeopardy is a US quiz-based game show, where contestants are provided answers and have to guess the questions. Soon after, AlphaGo, Google’s AI system that specialised in Go games, beat Fan Hui, the then European champion, and Lee Sedol, the world champion, in 2016. If that wasn’t enough, it beat itself—AlphaGo Zero.23

All these advancements can be traced back to one core development—a complete shift in how AI systems were being trained. The early 2000s saw a huge improvement in machine learning systems, which further distilled into deep learning systems over time. Fundamentally, a shift from expert systems to deep learning is a shift from creating coding rules for computers to allowing computers to learn and draw inferences through correlation and classification on the basis of substantial amounts of data.

Deep learning techniques have since been refined and enhanced, especially with increased amounts of data at our disposal. In 2003, Geoffrey Hinton at the University of Toronto, Ontario; Yoshua Bengio at the University of Montreal, Quebec; and Yann LeCun at the New York University, New York City, started a research programme to bring neural networks up to date. Simultaneous experiments were conducted at Microsoft, Google and IBM to show that these deep learning techniques resulted in halving the error rates for speech recognition.24

The experiments done on the impact on image recognition showed similar results, further validating the need to shift from expert systems to deep learning techniques. The challenges with text recognition and understanding are harder, as we have seen with some of the issues with GenAI tools even in 2024. While the AI system may be smart enough to understand the instructions being issued, it often struggles with contextualising them.

The broader field of AI can now further be refined into artificial narrow intelligence (ANI), artificial general intelligence (AGI), artificial super intelligence (ASI) and GenAI.25 ANI is the use of AI for a fixed, specific task. Among the best examples of ANI is Deep Blue, IBM’s chess-playing computer, which was capable of playing chess extremely well but was unable to do anything else.

ANI systems are designed to specifically solve a given task or problem and will be able to do so with a very high level of precision. The early stages of autonomous vehicles could be considered ANIs since they were only trained to drive and not have any other features like self-learning based on traffic rules. This is another feature of an ANI system: they tend to be incapable of self-learning and will be unable to use their training to perform another task, even if it is very similar to what they have been trained for. These systems are finding rapid deployment across the world as they are the easiest entry point for enterprises that want to adopt AI-based systems. Some common use cases include scanning and analysing medical reports to come to an initial diagnosis or analysing data to predict demand.

AGI systems are designed to be self-learning, such that they can then apply their knowledge to a range of evolving tasks and situations. Most experts believe that we are at an early stage of AGI, and it remains to be seen just how far machines can go with AGI. A truly effective AGI would be akin to a fully functional human being, able to learn, understand and react appropriately in different situations based on its training. Unlike an ANI system, which can scan medical reports and summarise and perhaps even provide a diagnosis, an AGI would then go on to prescribe a preferred line of treatment tailored to the individual patient based on their history and other parameters specific to that person. Currently, AI systems are still not capable of doing this and, at best, are being used by doctors to provide an initial diagnosis, which they then go on to validate based on their experience and then prescribe the way forward for the patient.

ASI refers to an AI system more intelligent than a human being. This is one area where experts are currently split down the middle on whether the evolution of an ASI will be good for humanity or an unmitigated disaster.

Generative AI, or GenAI, refers to AI systems that can analyse existing data and create new data based on this. This includes image generation systems like Dall-E 2 and text generation systems like GPT-3 and GPT-4, apart from audio, video and code generation.

What we are seeing at present is a coming together of AI and automation at an unprecedented scale. With RPA, once a developer or programmer defines a specific set of instructions for the robot or bot to perform, it can then perform this repetitive action at high speed and in high volumes in an almost completely error-free manner. There are different kinds of automation from the basic automating of tasks through RPA and business management tools to process automation where software is used to automate entire workflows or processes and even provide business insights and solutions. In AI automation, which is what we see in chatbots and virtual assistants, the system integrates the AI system with the chatbots, which allows them to understand generic questions and trawl through the available data to provide the appropriate responses.

At present, most enterprises have adopted some form of basic process automation in their day-to-day operations. Enterprises operating in service-oriented fields like financial services, hospitality, retail and healthcare have been among the early adopters of AI automation, introducing dedicated chatbots and customer service assistants. Several companies now allow you to complete entire transactions without having to ever directly interact with a human being. For instance, if you want to return a pair of jeans you bought online, a series of clicks will take you through the process, and your return request will be filed. Even buying insurance online can be done with zero human interaction. If a customer wants to buy insurance, the system will use GenAI to understand the request and extract the relevant information and then use RPA to generate the quote. This will then be sent to the customer, and if they accept it, another system will create a customer profile and generate the insurance and send it back to them. All this is done without any human interaction on the insurance firm’s side.

Combining AI with automation closes the loop on the progress made in the respective fields over the last few decades. The efficiency that’s being brought about by combining automation and AI will only increase going forward. The applications could cover a much wider range of tasks than what is currently possible, and all of this, in turn, will have a positive impact on worker productivity and overall efficiency. Setting aside the doomsday scenarios of machines completely taking over the world, most experts believe that this will not cause mass unemployment. Automation has slowly become more prevalent over the last few decades, and each time, there have been concerns over the impact this would have on employment. It was once believed that the introduction of automated teller machines or ATMs in banks would render the bank staff jobless, but five decades on, we can see that that isn’t the case. While it does render some specific tasks and roles irrelevant, what automation primarily does is transform jobs. This is apart from the new jobs that automation would create. Leaders, or aspiring leaders, would do well to keep this in mind as they craft their people strategies.