Thoughts on the Artificial Intelligence Boom

11/09/2024

By Carlos Martinez Cagnazzo, Chief Technical Officer at LACNIC

With all the excitement and controversy  around artificial intelligence, I think now is the ideal time to concentrate on what really matters:  Where can AI truly make a meaningful impact? Where does it fall short? Can it solve everything and beyond? No one denies that AI is a remarkably powerful tool, but its impact is stronger in some areas than others.

To provide some context, AI is a relatively new technology, but it is one of those groundbreaking innovations that comes along once in a decade, once in a generation, or even once in a lifetime. In short, it is a revolution that impacts almost everything—from the economy, work, and finance to art and creativity, as well as security, privacy, and governance.

But let’s return to the basics. What exactly is artificial intelligence? There is an informal definition that I really liked from Mark Smith, a physician who gives presentations on artificial intelligence with a focus on medicine. He describes it this way:  “Any activity performed by a machine that, if done by a person, would make us say, ‘Wow, that’s smart.’”

There are two different periods in the development of this technology. The first occurred between the 1940s and 1950s. During that time, there was a strong emphasis on linking “intelligence” with logical-mathematical thinking, such as mentally adding two 20-digit numbers, multiplying two 20-digit numbers, or playing chess. However, both are rule-based tasks that are easily handled by a computer and algorithms.

But this did not make the computer intelligent. The perception of the concept of “intelligence” began to evolve, shifting towards more complex aspects like language use, learning from experience, connecting concepts, categorization, classification, and the use of analogies.  This transition started to emerge in the late 1990s and early 2000s.

When we talk about algorithms—whether it is for making a breaded cutlet or powering platforms like Instagram or Uber—we are referring to a precise and step-by-step recipe aimed at producing a specific result. But is that artificial intelligence? Not really, because these algorithms are tailored to a single purpose and do not learn from their experiences; they do not remember previous actions.  The first major approach to artificial intelligence is developing a system designed to be more general in problem-solving and capable of adapting to variations in those problems.

A  Brief History

Looking back, it is clear that the pursuit of “automation” has been consistent throughout human history.  We can already observe it in the early history of this technology, with 17th-century automata, the fascinating story of Jacquard’s looms with their “programmed” designs or the analytical engine of Charles Babbage and Ada Byron.

The 1940s brought us the modern Theory of Computation by Alan Turing. By the 1950s, the term “Artificial Intelligence” was first introduced in 1956 by John McCarthy, Marvin Minsky, and Claude Shannon. In 1957, Frank Rosenblatt introduced the concept of the “neural network,” which laid the theoretical groundwork for artificial intelligence.

There has also been a historical evolution in the idea of “building an intelligent system,” which is important to highlight given the common confusion between the terms “artificial intelligence” and “machine learning.”

In the 50s and 60s, when the logical-mathematical approach was at the forefront, creating an intelligent system meant replicating how we think and reason.  This led to the development of symbolic AI, rule-based systems, and the so-called “expert systems.”

In the 80s and 90s, machine learning began to take shape, driven by the ability to process large amounts of data. For the first time, vast volumes of digitized information became accessible, allowing systems to learn from this data. From the 90s onward, the time has come to replicate the brain’s functionality through neural networks, leading to the rise of today’s prominent technologies, like language models such as ChatGPT.

It is interesting to note that many common applications, which people might not associate with artificial intelligence, actually rely on AI combined with neural networks. These include image classification, Internet search, numerical data processing in time series, categorization, and clustering. Recommendation systems on platforms like Netflix, Amazon, and Spotify are examples of artificial intelligence, as are voice recognition and synthesis tools such as Alexa and Siri. Other AI-driven applications include navigation apps such as Waze, credit card fraud detection systems, and services like Google Photos and Google Translate.

The Era of GPTs

Although artificial intelligence has been evolving for some time, it truly captured widespread attention on November 30, 2022, with the release of ChatGPT version 3. This version reached 100 million users in just over two months—a milestone that took Netflix ten years and Instagram two and a half years to achieve.

Returning to the story, we have already mentioned that “neural networks” are designed to replicate how the brain works, but how exactly are they trained? In supervised learning, a neural network is trained by minimizing error or maximizing reward, starting with a labeled dataset.  In contrast, unsupervised learning enables the network to identify hidden structures and patterns on its own, without relying on prelabeled data.

What is happening with the operation of Large Language Models (LLMs), the standout technology of the moment? Firstly, LLMs are massive neural networks; for example, given a sequence of words, an LLM statistically predicts the next word in the sequence. It is worth noting that these models are trained on vast datasets, which include almost all the text available on the Internet.  A common example of an LLM is the phone’s autocorrect feature.

What does the acronym GPT stand for, which gives these models their name? First, it is called “Generative” because it creates content that did not exist before by predicting the probability of the next word. “Pre-Trained” refers to the fact that it generates this content based on extensive training with large volumes of text. Finally, “Transformer” describes the neural network architecture employed.

In short, the artificial intelligence behind these models is remarkable because it not only generates new content (sequences of symbols that did not previously exist) but also “understands” what we ask of it and intuitively grasps what we need. Also, it processes unstructured information, making it more versatile and less tailored to specific tasks.  This is why, when you interact with ChatGPT, the natural language interface gives the impression that you are “talking to someone.”

In conclusion, I would like to emphasize why, even though the concepts of artificial intelligence have been around for at least 60 years, we are now in such an exciting phase.

First and foremost, the existence of the Internet plays a fundamental role, as it underpins many other factors by providing the vast volume of data necessary to train GPT models.  Additionally, cloud computing offers substantial computational power on demand, and hardware like GPUs, originally developed for different purposes, also accelerate key mathematical operations. Last but not least, there is strong financial interest in investing in artificial intelligence applications. 

While AI is often associated with risks like job displacement, model biases, social trust issues, and challenges to the concept of “truth,” it also offers significant benefits. AI is highly efficient, enhances productivity, improves communication clarity, encourages innovative thinking, and supports better decision-making.