Beyond the Buzz: The Story Behind AI’s Emergence

The emergence of artificial intelligence has become more than just a buzzword; it is a wave that has touched almost everyone, including people outside of the tech world. But have you ever thought, what has happened suddenly that draws the entire industry's attention to AI? Was it a Big Bang revolution, or is there a deeper story behind Artificial Intellgence's rise to the limelight? The purpose of this article is to explore the historical roots of AI and provide a glimpse into the future.

First thing first, is the idea of "thinking machines" having intelligence new? Certainly not—efforts to bring this concept to life have been in the making for decades. By and large, if we consider the entire AI journey, there are three major milestones. The word "artificial intelligence" was first used, in 1955, by scientist John McCarthy. He introduced the idea of 'intelligent machines,' which could match or surpass human intelligence.

To achieve this, researchers identified two major dimensions. The first one is Symbolic AI, and another one is Connectionism AI. We often refer to Symbolic AI as GOFAI (Good Old Fashioned AI), which is based on rules and logic, whereas Connectionism AI is inspired by how neurons in the brain are organized in layers and communicate with each other.

To dive further, symbolic AI uses symbols and applies predefined rules. The way the human brain uses symbols to understand and remember, for example, if we hear the word "tree," our mind recollects everything that is associated with a tree, such as its roots, leaves, color, type of tree, etc. These symbols or information are often organized in networks that define their relationships. Basically, at the core of symbolic AI, there is an expert system, which uses if/then rules to guide decision-making and behavior. Symbolic AI dominated research from the 1950s until 2012; during that time, many systems were developed, like ELIZA (1960), SHAKEY (~1972), Perceptron (1957), and Backpropagation (~1986). While these innovations contributed valuable insights, none were able to create systems that could rival or surpass human intelligence.

From 2010 onward, research institutions like UNC Chapel Hill, Princeton University, and the University of Michigan organized the ILSVRC (ImageNet Large Scale Visual Recognition Challenge) competition, which benchmarks the efficiency of algorithms for image classification. In this, participants were given a large dataset of images, and they had to correctly identify and categorize them.

In 2010 the best-performing models had an error rate of around 25%, meaning they could only classify images correctly 75% of the time. However, in 2012, a major breakthrough occurred when Alex Krizhevsky, a student at the University of Toronto, introduced a deep neural network called AlexNet. AlexNet, consisting of eight layers, dramatically improved performance, reducing the error rate to 15.3%—a significant leap that marked the rise of connectionist AI (deep learning).

This initiated a wave of innovation.

• In 2013, GoogLeNet was introduced by Google, which had 22 layers of neural network. This further reduced the error rate to 6.7%.
• In 2015, ResNet (Residual Neural Network) was launched by Microsoft, which had 152 layers. This achieved an error rate of just 3.57%. This surpassed the average human error rate of 5.1% in image classification.

This was a defining moment; for the first time, machines were able to outperform humans in image recognition.

Alex’s experience left a lasting impression that influenced developments in the years that followed. In 2017, Google published a groundbreaking paper, "Attention is All You Need." This was the major milestone and introduced the Transformer architecture. Transformer architecture revolutionized the field of Natural Language Processing (NLP). Before this paper, it was challenging for machines to understand the meaning of entire paragraphs. However, the Transformer model enabled machines to understand context at a much deeper level, marking a significant leap forward in NLP.

The impact of transformer architecture was so profound that it laid the foundation for many innovations, and well-known models such as BERT (2018), GPT (2018), and T5 (2019) are all built based on that. The implementation of LLM enabled machines to become capable of understanding written and spoken human language effortlessly.

While applications such as ChatGPT could communicate with humans effortlessly, they lacked the ability to perform any action. This limitation paved the way for AI agents in 2020. Unlike traditional chatbots, these agents don’t just provide information; they can also act on it. For example, while ChatGPT might tell you the best places to visit in the USA and recommend hotels, an AI agent could go a step further and actually book a hotel for you on your specified dates.

AI agents, although capable enough to perform actions, were limited to handling specific tasks such as managing calendars, checking weather, or similar tasks. We can consider AI agents as virtual assistants that follow instructions and execute but don't think independently.

This limitation led to the emergence of agentic AI. In 2025, NVIDIA introduced a toolkit, AgentIQ, which is used to support the development of autonomous, reasoning-capable AI agents that can operate across both organizational and individual contexts. Essentially, agentic AI is the framework, whereas AI agents are the building blocks within the framework. As well, agentic AI is a bigger concept of solving issues with limited supervision, whereas an AI agent is a specific component within that system that is designed to handle tasks and processes with a degree of autonomy.

Despite the evolution of agentic AI, the communication between agents was not well structured. This gap was addressed in 2025 when Google introduced A2A (Agent-to-Agent). A2A is a protocol that allows AI agents to securely communicate.

In the same direction, in May 2025, Anthropic launched MCP (Model-Context-Protocol). MCP is an open standard that connects AI agents, LLMs, and platforms. MCP enables not just communication, but also context sharing and seamless operation across ecosystems.

For example, a traditional agent might be hardcoded to check the weather. If we want it to book flights, we need to manually add that functionality. Whereas, an MCP-enabled agent, without being reprogrammed, can dynamically access a weather tool, a flight booking API, or a CRM system. It simply sends a structured request via MCP.

There has been a lot of progress in areas like MCP, agentic AI, and A2A. But when you see the basics, what are we really trying to achieve? Ultimately, building machines that can think and reason like humans.

Look at the journey of artificial intelligence so far, and you soon realize that the dream has always been to build machines whose smarts match and even exceed human intelligence. In this regard, large language models (LLMs) are a major step in that direction. You can think of them as the brain of the system, capable of understanding and generating instructions.

ChatGPT which is a client of the GPT model, can communicate, explain, and respond, but it can't do any work by itself. To actually get things done, we need something that can take action, like limbs in a human body. That is where the concept of AI agents comes in. These agents are designed to perform tasks autonomously, following the instructions from the LLM "brain". As we move forward, these agents also need to collaborate or communicate with each other (like human limbs coordinating together) to handle more complex tasks. The communication between agents (or limbs) is named agent-to-agent (A2A) communication. The agent interactions, not just with each other but with the LLM and external systems, are known as Model-Context-Protocol (MCP).

Just to conclude, we are still in the early or weak phase of AI. We are making progress in simulating human intelligence across various domains like healthcare (diagnosis and surgery), law, transportation (like self-driving cars), and more. But there is still a long way to go before AI can fully match the depth and flexibility of human cognition.

#bert #agent #ai #transformer #models #learning #Models#MCP#A2A