Artificial Intelligence (AI) is revolutionizing the industry as a whole and changing the future of work. As with any emerging technology, the language is changing alongside the technology itself. This piece is intended to unpack some of the most commonly used terms and concepts in the AI space. From Artificial General Intelligence (AGI) to Generative Adversarial Networks (GANs) to Large Language Models (LLMs).
In a recent discussion, Sam Altman, CEO of OpenAI, described AGI as “the equivalent of a median human that you could hire as a co-worker.” AGI aims to develop general intelligence of a kind that will be a major step towards making massively superhuman capabilities in terms of economically important tasks. This ambition is unambiguously stated in OpenAI’s charter. Altman didn’t explain what AGI means in this context. Its significance in the AI ecosystem cannot be overstated.
To really understand AI, you have to know what’s going on behind the curtain. GANs and LLMs are two different but important fields in AI research and development. Understanding these ideas allows people to better understand what is involved in AI systems. It further reveals their transformative promise for industry, academia and government.
Generative Adversarial Networks: A Dual Approach
At the core of GANs is the astounding application of two neural networks competing against each other. In each iteration, one of the neural networks is solely responsible for generating that output based on its training data. This generative model produces new data instances that follow the same distribution as the original data.
That generated output is then fed into a second neural network. This one acting as a classifier. This new classifier jointly scores the authenticity of each generated data instance wrt real data instances. That feedback loop between these two models is what pushes the generator to make constant iterative improvements. It continues to improve its outputs, leading to more and more lifelike outcomes.
This two-pronged strategy has been highly successful in applications such as generating art, as well as photorealistic image synthesis. The generator and discriminator are locked in a perpetual arms race. This interaction significantly increases the efficacy of GANs.
Deep learning systems like GANs too, are primarily driven by data. They require huge quantities of data points to generate acceptable output. Usually, millions of such examples are required to adequately train these models. Graphical processing units, or GPUs, paved the way for deep learning’s success. They deliver the computational power necessary to process massive datasets quickly and accurately.
Large Language Models: Mapping Language
Large Language Models, or LLMs, are another breakthrough innovation in AI technology. Realized as billions of numerical parameters, LLMs are extraordinary transformers of human language. They learn relationships between words and phrases by encoding patterns found in extensive datasets, including billions of books, articles, and transcripts.
As a result, it produces a very powerful model of language. This complex and transient landscape of meaning is what allows LLMs to generate believable and contextually appropriate text. Although these models can selectively generate disinformation, they primarily produce dialogue that convincingly mimics human conversation. This feature makes them extremely useful for applications like chatbots, content generation, and language translation.
LLMs are based on a remarkably strong architecture that models complex linguistic structures. This computational design gives them the ability to produce text that ranges from a catchy slogan to a detailed story. Because of this, their versatility has led conversational agents to become a staple across multiple industries, such as education, entertainment, and customer service.
Transfer Learning: Building on Previous Knowledge
The other major underlying premise of AI development is something called transfer learning. It’s the practice of using a pre-trained AI model as the starting point for developing a new model. This method gives researchers and developers the opportunity to build on what’s already been learned and save significant time training for new applications.
By starting with a model that has already learned valuable insights from large datasets, developers can fine-tune it for specific tasks with fewer data points. This saves their team time — potentially months — while decreasing the computational resources needed to train their models.
Given these challenges, transfer learning is most impactful in situations where a lack of available data creates obstacles. By creatively adjusting existing popular models, researchers are able to achieve state-of-the-art results with small datasets. In a way, this technique beautifully encapsulates the collaborative nature of AI research, where innovations are layered upon each other to push the field forward.