The Rise of Artificial Intelligence

[Part 1 of a Three-Part Series respectively on the Past, Present, and Future of AI: A Non-Technical Exploration!]

Problems which can be defined in language can be coded in a software but where precise definition is not possible, like identifying cat in an image, how to tell what a cat is? So only way left is as humans learn, you can’t tell a human child what a cat is but if you show him 2-3 examples of cat he will identify the next cat. Machines too needed such models that can learn from examples. These models, which learn through exposure to data, are known as AI models. Typically, they are mathematical structures with interconnected nodes and weights that adapt and improve as they analyze more data.

For decades, numerous AI models were developed, but none of them worked. Their efficiency was too low and the reason for this remained a mystery—especially when similar approaches seemed to work remarkably well in humans.

In 2007, Fei-Fei Li had an insightful idea: unlike a blank slate, human children don’t start learning from scratch. Even if they haven’t seen a cat before, they have prior knowledge of other animals, like dogs or rabbits, and can transfer this knowledge to identify a cat after just a few examples. AI models, on the other hand, begin with no prior knowledge. For instance, if you show an AI two images of cats, it would attempt to find common patterns between them and might mistakenly include irrelevant objects, like trees, as part of its “cat” identification. Fei-Fei Li hypothesized that by providing AI models with a vast number of labeled images, the likelihood of irrelevant items appearing in multiple images would diminish. This approach would allow the AI to learn more efficiently and accurately identify a cat. To test her hypothesis, Fei-Fei Li gathered 13 million labeled images and launched the ImageNet, an annual competition, in 2007. The competition invited AI researchers to train their models on these labeled images and evaluate their accuracy. Initially, the best-performing models achieved only around 25% accuracy. However, by 2012, accuracy levels had dramatically improved, reaching 98%. The winning model, AlexNet, was based on a neural network architecture developed by Geoffrey Hinton, Yann LeCun, and Yoshua Bengio in the 1980s, marking a turning point in AI research and image recognition.

In 2015, the accuracy of AI models in image recognition surpassed human-level performance, which led to a pivotal moment in the history of artificial intelligence. This achievement solidified the belief that AI had truly “arrived,” with neural networks now being synonymous with AI.

Following this breakthrough, major tech companies raced to recruit the leading minds in the field. Google hired Geoffrey Hinton, Facebook brought Yann LeCun on board, and Baidu hired Yoshua Bengio, all of whom had contributed immensely to the development of neural networks.

In 2024, Geoffrey Hinton was awarded the Nobel Prize in Physics for his pioneering work in neural networks, recognizing his significant impact on the field.

AI has traditionally been highly effective in niche domains, such as identifying objects in images, but it lacked the broader understanding of the world. This AI was referred to as Artificial Narrow Intelligence (ANI) because it could only excel at very specific tasks.

However, by early 2017, the concept of Artificial General Intelligence (AGI) began to take shape. AGI is a hypothetical form of AI that could understand the world in a manner similar to humans. While humans took millions of years to develop this understanding, it was hypothesized that AI could achieve a similar comprehension much faster if humans could transfer their knowledge directly to AI models.

The key challenge was that human knowledge was primarily stored in text form, predominantly in English. Transformers were the best model available at that time for learning from text, and it was hypothesised if somehow we can train transformer on all human knowledge it may get understanding of the world as humans have. It was an expensive hypothesis to test.

Enter Sam Altman, who took on the monumental task of training a Transformer model on a massive dataset scraped from the internet, encompassing vast amounts of human knowledge. And voila! It worked.

These trained transformers are known as Large Language Models (LLMs). The term “language model” is somewhat misleading, as although these models primarily use language as a medium, they don’t just process language—they understand the world.

In November 2022, Sam Altman launched ChatGPT, which had this advanced capability. The world celebrated this as the arrival of what many perceived to be Artificial General Intelligence (AGI).

However, this milestone marks only the first level of AGI development. It was a significant step in the evolution of AI, but there is still much further to go on the road to AGI’s full realization.

Stay tuned for?Part 2,?where we will explore the present state of AI.