An Introduction to Artificial Intelligence
Artificial Intelligence (AI) has transformed the technology industry; some predict it will bring the next productivity revolution, similar to industrialisation in the 20th century. It is increasingly being used in applications used by businesses or consumers. The news every day mentions billions of dollars being invested by tech companies into AI-enabled applications and the underlying infrastructure of data centres.
This article marks the beginning of a three-part series focused on artificial intelligence. It digs deeper into AI and answers fundamental questions: What is AI? Who are the leading players providing AI technology? What are some real-world examples of AI usage?
AI in a business context: Should you adopt AI in your business? <to come>
Using Microsoft CoPilot in your business. <to come>
Introduction to AI
AI is the ability of software systems to 'simulate' human cognitive functions, a kind of human-like intelligence - learning, reasoning, and problem-solving. It is not new; some form of AI already existed before, like manufacturing robots, and Amazon or Netflix recommendations. Traditional AI uses rules-based algorithms; Google recommends travel deals if I search for flight tickets. A smartphone camera auto-corrects its focus and flash settings based on whether it is sunny or dark. These are examples of narrow AI.
AI evolved with the launch of ChatGPT in 2022, enabling software to converse in natural language. ChatGPT understands my prompt, analyses it, and uses its vast knowledge combined with its existing memory about my previous interactions (personal context) to produce an intelligent response. While software assistants like Amazon Alexa and Apple Siri could respond in natural language before, this was the first time a system could understand, interpret, and respond in natural language with coherence and general-purpose capabilities. The innovation made ChatGPT the fastest adopting technology system ever, with 1 million users in 5 days.
The competitors were not far behind. Google launched its Gemini along with the AI mode in search. Grok by Elon Musk's X and Claude by Anthropic are other generative AI systems.
AI systems evolved from traditional rules-based approaches to modern 'statistical and probabilistic' approaches. The generative AI is a system that makes the best prediction for what should come next when preparing a response, and it's this high probability of picking the next word correctly that produces a coherent sentence, paragraphs, or an entire article.
These AI firms continuously evolve their offerings for more accuracy and faster performance. A race towards Artificial General Intelligence (AGI) is a form of AI on par with or surpassing human intelligence. AGI that can do research and even make discoveries, perhaps a new medicine for a nasty disease. Sam Altman, OpenAI CEO, has stated that AGI could be a reality in the near future.
A little bit about the science behind AI
AI is a broader term that defines all types of artificial intelligence, such as generative AI, computer vision, and natural language processing (NLP). In AI, machine learning (ML) is used to develop 'models' of these different types. A model is a software programme made of algorithms and data. The ML methods, such as Deep Learning, are used to develop more complex models. As with many things that we humans create, the inspiration behind Deep Learning is the functioning of the human brain itself. Deep Learning uses neural networks, which are multiple layers with each layer made of various neurons, or nodes. A layer takes input from its preceding layer, performs work and passes output to its succeeding layer.
Consider a prompt 'Generate an image of summer school holidays on a beach in Wellington, NZ' to DALL-E, OpenAI's image generation model. This model will first create an image pattern with an outline of a beach - the sea, children, and sandcastles. The model then refines this image with increasing accuracy as processing goes through its layers. One layer may create an image of a person, another of a sandcastle, and ultimately, generate the whole picture. It is like building a house; first comes the architectural plans, next comes the construction of foundations, and then there are the fittings, like the kitchen and so on.
The diagram depicts the relationship of AI and its constituent members. We do not need to become data scientists, but knowing different terms and their positions in the AI discipline helps.
Model development
The generative AI, the AI that generates something new like text, images, audio or video, is based on large language models (LLMs). These are all-encompassing language models trained on huge data volumes - books, blogs, websites. A model is only as good as the data it is trained on. Thus, data plays a critical role in model development. Data can be labelled or unlabelled. An example of labelled data is an image of a cat with the label 'Cat'. Unlabelled data has no labels, like a thousand images of cats and dogs; to software, these are all just images. Model developers, like OpenAI, have signed many agreements with content producers to use their data for model training.
Labelled data is expensive, as a human is required to create the label. This means going through the thousand images and labelling each as a cat or a dog. Labelled data improves model accuracy, whereas unlabelled data is used for large-scale pattern recognition. An AI model processes all one thousand images and groups them into two categories: first, being Cats and second, being Dogs.
The training process involves a model learning the patterns in language to recognise the characters and the underlying sentiment. A model that can identify positive, negative, or neutral sentiments. This pattern recognition can be understood from an example of a model identifying digits in an image. Consider that this model is trained on the MNIST database, a database of thousands of images with each image having a handwritten numeric character. The computer program can identify the patterns that make each numeral and then use its learnt knowledge to determine the digit from a new set of images successfully.
Similarly, pattern recognition is used with computer vision models to identify shapes and other objects in images or videos (thousands of images or frames). Such models are being used for self-driving cars to identify obstacles on the road correctly. Robotics is another field that combines multiple models - Computer Vision for object identification, NLP models for conversations and so on.
AI is not sentient; it does not possess consciousness or human-like awareness. It is all a software programme that can create new patterns based on the patterns it has learnt. The MNIST data example above shows how a software programme is developed that can identify or even create new digits.
Agentic AI
The larger, complex AI models like LLMs are the foundation of many applications, with ChatGPT being one of the most prominent. Applications take LLMs as foundation models, enriching them with configuration, coding, and integrations for specific tasks.
For example, consider the prompt “top news stories in NZ today” in ChatGPT. The GPT LLM is trained on historical data—it does not inherently know today’s headlines. To answer such a query, ChatGPT can call on an AI agent to search the web for current news, then merge the retrieved information with its reasoning capabilities to generate a response.
Microsoft Copilot is also built on foundation LLMs, enhanced by AI agents. These Copilot agents can fetch data from Bing Search, an organisation’s internal data sources (like SharePoint, Teams, or Outlook), or integrate with external systems to provide richer, context-specific responses.
This leads us to agentic AI, combining foundation LLMs with AI agents to achieve specific goals. Agentic AI can orchestrate a variety of agents to not only fetch information from the web but also take actions such as sending an email, booking a calendar meeting, or reserving a hotel room.
While generative AI is reactive, it produces content in response to a human prompt, and agentic AI systems are proactive. They can be fully autonomous, triggered by human input or events, and then carry out multi-step actions to accomplish goals.
Examples of agentic AI systems include:
GitHub Copilot – assists developers by proactively suggesting and generating code.
Microsoft Security Copilot – helps security teams detect, investigate, and respond to threats by drawing on AI agents and organisational data.
ChatGPT with Actions and Custom GPTs – where the model can call APIs, search the web, and perform tasks.
AutoGPT and BabyAGI – early experimental open-source frameworks for autonomous agents.
AI hallucinations and bias
AI is trained on data, and data quality can vary—for example, data from a newspaper. While a national newspaper could be considered to have factual, trustworthy information, its reporting still will have its biases - left or right-leaning news media. Other types of biases include gender bias or racial bias.
This leads to generative AI systems to generate responses that aren't factual, simply made-up information presented as fact. That's the reason there are disclaimers accompanying most AI systems today.
Summary
AI is a field of computer science that develops software systems with human-like intelligence. The different types of AI - Computer Vision, NLP, and Generative AI - are developed using machine learning techniques such as deep learning. These software systems are called models, a combination of algorithms and data. The models are trained using vast quantities of labelled or unlabelled data. Through training, AI systems develop the ability to recognise patterns. This acquired knowledge empowers them to identify and classify new patterns, generate novel data, and perform predictive tasks. The LLMs are the complex, all-encompassing models, also called foundation models, that are used by applications. Agentic AI extends the foundational LLMs to achieve goals through AI agents, where each agent performs a specific task. GPT 4 is OpenAI's LLM, which is used with ChatGPT and sold as a service to other vendors. Microsoft has exclusive licensing rights to use GPT models as a foundation for its apps, branded as CoPilots.