In late November 2022, the world's first generative AI chatbot was released to the public, marketed as a revolutionary technology and the next great leap in human advancement. Within just two months of its launch, ChatGPT crossed 100 million active users, one of the most rapid technology adoption rates in history. Today, an estimated 1.5 to 2 billion people globally use generative AI in some form.
For anyone witnessing this development from the periphery of the tech world, artificial intelligence may appear to have emerged out of nowhere. A technology that entered public consciousness less than six years ago has now become the go-to assistant for work, education, creativity, communication, and information. But the vision of artificial intelligence is nearly 70 years in the making.
"We're wired for linear thinking. So exponential growth always feels like it came out of nowhere," says Ashwin Vasudevan, founder and CEO of Migratex and co-founder of Hectae Analytics and Software Solutions.
Yet the pace of AI's rise surprised even people working in technology, including Vasudevan himself.
"Before ChatGPT, I genuinely believed we were still at least two decades away from large-scale AI adoption. I expected progress to continue, but mostly in the background through specialised applications. I was very wrong. I underestimated how quickly advances in computing power, data, and model architecture would converge,” he added.
The Origins of Artificial Intelligence
The intellectual foundations of AI emerged in the aftermath of World War II.
In 1950, Alan Turing published the paper, Computing Machinery and Intelligence, posing a question that would lay the foundation for modern AI: ‘Can machines think?’
Rather than wrestling with philosophical dilemmas, he proposed a practical test called the Imitation Game. If a machine could hold a conversation convincing enough to fool a human interrogator, he argued, it deserved to be called intelligent. It was a radical idea at a time, when computers were primitive machines occupying entire rooms, and struggled to do basic arithmetic.
But Turing's proposition landed in fertile ground. It planted a seed that researchers have been watering ever since.
Six years later, in 1956, a group of scientists gathered at Dartmouth College and formally coined: Artificial Intelligence. These early pioneers believed that human intelligence could eventually be mapped, coded, and replicated through logic. Their approach, known as Symbolic AI, or “Good Old-Fashioned Artificial Intelligence” (GOFAI), treated the computer as a master logician. Feed it enough "if-then" rules, the thinking went, and it would eventually think like a person.
The Long Winters
From a theoretical standpoint, scientists predicted machines with human-level intelligence would be a reality within a few decades. But they were wrong.
The “AI winters” were periods when excitement around artificial intelligence collapsed, funding dried up, and researchers struggled to translate theoretical promise into tangible results. There were two major AI winters, and through years of trial, error, and patience, the two periods of decline ultimately laid the groundwork for the breakthroughs that would follow in the decades ahead.
The scientific community at the time was divided into two groups. One group of scientists kept building their rule-filled logic machines, another group took a completely different approach, looking at the human brain for inspiration.
A researcher named Frank Rosenblatt built an artificial neural network called the Perceptron in the 1960s. Perceptron was an early computer system designed to recognise patterns and learn from data. Think of it less like a machine being told every rule beforehand, and more like teaching a child through examples, show it enough pictures of cats and dogs, and it gradually begins to recognise the difference on its own. It was the first real sign of a machine that could genuinely learn, beyond following instructions.
However, in 1969, two prominent scientists published a paper that was formal proof that the Perceptron had fundamental limitations, there were problems it simply could never solve, no matter how long it trained.
In 1973, the British government commissioned the Lighthill Report, officially titled Artificial Intelligence: A General Survey, a highly influential paper written by applied mathematicians to evaluate the state of academic artificial intelligence research. The report delivered a scathing and deeply pessimistic assessment of the field, arguing that many of AI’s major promises remained largely unfulfilled.
Almost overnight, research funding vanished, scientists moved on, and AI entered a period of stagnation that would come to be known as the first AI winter.
The 1980s brought a brief thaw.
Companies began to invest heavily in Expert Systems, a form of artificial intelligence that encoded the knowledge of human specialists into vast rule based databases. Unlike earlier neural networks that learned by recognising patterns from data, Expert Systems functioned by memorising enormous instruction manuals, following predefined rules instead of identifying patterns independently. Expert Systems relied entirely on predefined “if-then” instructions written by programmers.
One early success, a system called XCON, saved its corporate owner millions by automating complex computer configuration tasks. But, just as before, the limitations became impossible to ignore. These systems were brittle, brilliant within a narrow lane, yet helpless the moment reality became unpredictable or messy.
By 1987, the market collapsed. Winter returned.
Permanent thaw
Despite the protracted AI winter, a small group of researchers continued working on neural networks, convinced the technology had been abandoned too early. Among them was Geoffrey Hinton, often referred to today as one of the “godfathers of AI.”
In 1986, Hinton, alongside David Rumelhart and Ronald J. Williams, helped popularise a training method called backpropagation, a technique that allowed neural networks to improve themselves by learning from mistakes. Although the idea existed earlier, their work demonstrated that multi-layered neural networks could, in theory, learn increasingly complex patterns from data.
In the years that followed, Hinton’s work on backpropagation was a powerful theory waiting for a powerful machine. But the progress remained slow. Neural networks were promising, but researchers faced a fundamental problem: computers simply were not powerful enough to train them at scale. For much of the 1990s, AI remained a niche academic pursuit, overshadowed by more immediate technological advances.
The explosion we see today was finally enabled by the "Big Three": massive datasets, advanced hardware, and refined algorithms.
The internet was generating unprecedented amounts of data. Computing power was becoming exponentially cheaper. And graphics processing units (GPUs), originally designed for rendering video games, proved remarkably effective at training neural networks. NVIDIA’s CUDA platform, released in 2006, gave researchers the tools to run deep learning models on this hardware.
The breakthrough arrived in 2012. Researcher Fei-Fei Li had spent years assembling ImageNet — a database of over 14 million labelled images, effectively building the world's largest visual classroom for machines. At the annual ImageNet competition that year, a team led by Hinton and his students unveiled a deep neural network called AlexNet, which dramatically outperformed every competing system, cutting error rates by an astonishing margin.
This was an "unequivocal turning point," proving that deep learning was the superior path for artificial intelligence.
But language remained a stubborn challenge.
Machines could identify objects in photographs with increasing accuracy, yet understanding human language required something far more complex: context. Words derive meaning from the words around them, and earlier AI systems struggled to maintain that context over long passages of text.
Researchers needed a new breakthrough. They got one in 2017.
Attention Is All You Need
In June of that year, eight researchers at Google published a paper with an unusually confident title: Attention Is All You Need. It introduced a new architecture called the Transformer.
“Before this paper, models processed language sequentially, word by word. This paper suggested looking at all the words and understanding how one word connects to another word,” said Vasudevan.
Previous language models processed words sequentially, one after another — slow, and prone to losing track of what had been said earlier. The Transformer solved this through a mechanism called "attention." Rather than reading text the way a tired person reads a long document (by the end, the beginning is already fading), the Transformer could examine all words simultaneously, understanding how each related to every other. Context, at last, could be preserved.
Transformers could be trained on vastly larger datasets, process language far more efficiently, and scale in ways no previous system could. The more data and computing power researchers provided, the more capable the models became.
The paper would go on to become one of the most influential research publications in the history of computer science. Almost every major generative AI system today — ChatGPT, Claude, Gemini — traces its foundations directly back to it.
In 2018, OpenAI released GPT-1, demonstrating that Transformers could generate coherent text. The pieces had finally come together: the architecture (Transformers), the data (massive datasets), and the computing power.
Reflecting on the launch of ChatGPT, Vasudevan said his focus was never solely on the technology itself, but on what it could do in the real world. "What interests me most is application, not research. The underlying technology is fascinating, but the real question is simple: what problem can this solve for someone today?"
Today, an estimated one to two billion people use generative AI in some form. What appears to be an overnight revolution is actually the culmination of decades of breakthroughs, failures, and persistence.
India and AI
The country produces some of the world's largest pools of engineering talent and has become a major hub for technology services, yet it remains a relatively small player in the development of AI models. "India's strength has historically been implementing technology rather than inventing foundational technologies. That carried into AI. The talent and ambition are there. What's missing is the ecosystem: compute, funding, and institutional patience," says Vasudevan. India's absence from the frontier of AI development is, in this context, both an economic and a strategic problem.
Challenges for India:
AI development requires enormous computing power and specialised chips known as GPUs, which India currently cannot manufacture at scale.
India has only three of the world's top 500 supercomputers, compared to 215 in China and 113 in the US.
For decades, India’s tech success was built on selling skilled labour rather than creating intellectual property (IP). This "services model" prioritised providing software solutions to global corporations, leaving India as a consumer of products rather than a creator of platforms.
India produces a high number of engineering graduates, but its PhD pipeline is significantly smaller — one-third of the US and one-fifth of China.
Around 7% of graduates pursue higher education abroad, and India struggles to attract or retain top-tier AI researchers, leading to a persistent "brain drain."
The scale of capital is vastly different; in 2024, US firms invested approximately $109 billion in AI ventures, while India’s private equity investment in AI (as of 2019) was only around $1 billion.
But
IndiaAI Mission was launched with an outlay of Rs 10,372 crore – to democratise technology and supporting indigenous models. Models like Hanooman (a multimodal LLM) and Bhashini (a translation platform) are designed specifically to handle the nuances of more than 22 Indian languages
Through the India Semiconductor Mission (ISM) 2.0, India has committed Rs 1,000 crore towards strengthening its semiconductor ecosystem, with a focus on domestic chip design and manufacturing and an ambition to emerge as a top semiconductor nation by 2035.
The Future of AI
Recently, Pope Leo XIV released Magnifica Humanitas, the first encyclical of his papacy, focused on the rise of artificial intelligence. In the document, he argues that "technology is never morally neutral" and warns that societies driven solely by efficiency and optimisation risk overlooking human dignity, wisdom, and the common good.
Drawing on the biblical story of the Tower of Babel, he writes that humanity faces a choice between "constructing a new Tower of Babel" or building a future in which technological progress serves human flourishing. The encyclical also raises concerns about the impact of AI on work, human relationships, democratic institutions, and the concentration of power in the hands of a few actors.
The concentration of power described by the Pope has a historical precedent. When a small number of actors control the infrastructure through which the rest of the world accesses knowledge, work, and communication, the arrangement begins to resemble something older than Silicon Valley. The colonial analogy is imperfect, as it always is, but the structure is familiar: data, labour, and markets flow upward toward a concentrated centre, while the populations that generate much of that raw material have little say in the systems built from it. The models are trained on the world's languages, cultures, and behaviours — and then sold back to that world, pre-loaded with the assumptions of whoever built them.
"One future is defined by AI improving productivity, healthcare, and education. The other is build around concentration of power, displacement without transition plans, and eroded trust. Both are already unfolding. The question is which one wins. AI colonialism is already happening,” concludes Vasudevan. “We're just not calling it that yet."