Introduction to Generative AI

Generative AI is transitioning from an industry buzzword to mainstream reality at a rapid pace. This article introduces generative AI at a high-level, laying the foundation for understanding the technology and its applications. It delves into the evolution of AI, its current capabilities, and the accompanying ethical considerations. The article ends with insights into the future of generative AI and its potential impact on our lives.

The History of AI

Understanding the history of AI provides a broader context for generative AI.

The roots of AI can be traced back to early philosophers and mathematicians who aimed to mechanize reasoning. However, the groundwork for modern AI was established in the 19th and 20th centuries, epitomized by George Boole’s Boolean algebra and Alan Turing’s concept of thinking machines.

In 1943, Warren McCullouch and Walter Pitts introduced the first artificial neuron, a mathematical representation of a biological neuron. This marked the beginning of neural networks, which are now fundamental to modern AI.

In 1950, Alan Turing released a paper titled “Computing Machinery and Intelligence”, suggesting a test for machine intelligence. This Turing test is still used today as a way to think about the evaluation of AI systems.

The term “artificial intelligence” was first introduced in 1956 during the Dartmouth Summer Research Project on Artificial Intelligence, marking the onset of AI research.

Numerous discoveries during this period spurred an AI boom in the 1960s, propelled by funding from the US Department of Defense for potential military applications. Leading figures like Herbert Simon and Marvin Minsky optimistically predicted that machines would achieve human-level intelligence within a generation. However, the intricacies of AI proved more challenging than anticipated, resulting in reduced funding and research, leading to what’s termed the “AI winter”.

The 1980s saw a revival in AI interest due to the commercial success of expert systems, which were rule-based systems emulating human reasoning. These systems found applications in diverse sectors, including healthcare and finance. Yet, this resurgence was temporary, with another “AI winter” setting in by 1987.

During the 90s and 2000s, machine learning (ML) became the predominant approach in AI. The amount of data that became available during this period was instrumental to the success of ML. Unlike traditional rule-based systems, ML algorithms discern patterns directly from data, leading to a range of applications such as, email spam filters, recommendation systems like Netflix, and financial forecasting. Machine learning shifted the focus of AI from rule-based systems to data-driven systems.

A significant shift occurred in 2012. Enhanced computational power (boosted by GPUs), data availability, and advancements in neural network algorithms gave rise to deep learning, a subset of ML. Deep learning quickly outpaced other ML techniques, leading to a surge in AI research, funding, and applications. By 2022, global investments in AI were approximately $91 billion, accompanied by a substantial increase in job opportunities and specialists.

Today, the applications of machine learning-based AI are ubiquitous, ranging from basic tasks like spam filtering to complex ones like autonomous vehicles and medical diagnostics. Generative AI has emerged as a subset of ML, and has garnered significant attention due to its ability to create content, such as images, videos, audio, and text.

What is Generative AI?

AI/ML engineers employ various tools and techniques to convert data into machine learning models, which then make predictions or categorizations. For instance, a model trained on an image dataset of cats and dogs can differentiate between the two based on learned patterns.

ML models cater to diverse applications: video security systems detect humans and potential break-ins, voice assistants like Siri and Alexa process speech to respond to user queries, autonomous vehicles identify objects and make decisions, and the healthcare sector utilizes ML to spot anomalies in medical images, among other uses.

Considering its pervasive use, let’s term this “traditional AI” or “traditional ML”. Such AI classifies or predicts content, taking an input to produce an output, such as identifying whether an image has a cat or dog, determining the best route to a destination, or estimating the likelihood of a tumor in an X-ray image.

Generative AI, a subset of ML, utilizes neural networks to create content. Trained on extensive datasets such as images, videos, audio, or text, these models generate new content based on identified patterns. Different generative AI models cater to varied content types: for instance, image generation models like OpenAI’s DALL-E rely on extensive image datasets, while text generation models like OpenAI’s ChatGPT are trained on vast text datasets.

Generative AI can craft a plethora of content. Image models can create diverse images, emulating specific artists or art movements. Similarly, text models can mimic specific authors or genres, producing text ranging from technical to creative, or even generate code in various programming languages.

Encountering generative AI models for the first time might seem magical. They appear to conjure the requested content out of thin air. Asking a text generation model to write a poem or a story, or an image generation model to create a painting or a photograph, can be a surreal experience. Consider the following examples:

The following images were generated using DALL-E, an image generation model. The model was prompted to “Generate an image of a bustling renaissance-era city populated by anthropomorphic animals.”

Notice the intricate details such as the ornate architecture and the different species of animals. In less than a minute, the model produced two unique images, each with a distinct style.