Insights

Do Generative AI tools like ChatGPT live up to their hype?

Adnan Masood, PhD. Chief AI Architect, UST

Every week brings fresh news about a new development in generative AI, the most exciting technological advancement in recent memory. Every day, there is a steady drumbeat of new AI-powered product launches or enhancements. Various product management tools, search engines, and social networks have announced they are incorporating AI into their solutions within the past month.

Dominating the headlines is OpenAI, whose ChatGPT-3, which launched in November 2022 and became the fast-growing consumer technology in history, procuring 100 million users in two months. It has raised $11 billion, 91% of that coming earlier this year via a stand-alone investment by Microsoft. It demoed the next iteration of this technology, ChatGPT-4, on March 14. While the specifics of GPT-4 - which underpins ChatGPT-4 - have not been revealed, the model is much larger and clearly demonstrates improvements in reasoning, problem-solving, and human benchmark performance compared to ChatGPT-3, which is based on GPT-3.5.

The attention paid to AI demonstrates to businesses and the public how this technology can and will fundamentally change how we communicate, learn, and work. There will be an arms race for multiple startups and incumbent companies to build the prevailing model that will be the preferred choice of companies, a potential multi-billion-dollar business. Right now the focus is on AI content and text generation, but there will be many other applications following suit.

Executives are paying close attention to how this technology can remake their business. Early results are very promising. But before we get into that, we need to discuss what generative AI is, who the major players are, and why it is such an exciting technology.

What is Generative AI

It is an ensemble of sophisticated machine learning and deep learning algorithms to generate novel content across various modalities - such as audio, code, speech, images, text, simulations, videos, and even XR experiences by analyzing and interpreting pre-existing content.

Generative AI can be integrated with other technologies, such as natural language processing (NLP) and computer vision (CV), to create powerful AI-driven solutions to help businesses improve their decision-making, automate tedious tasks, and create new revenue streams.

Generative AI offers a wide range of opportunities to businesses of all sizes and across industries to help streamline their operations, expedite technology enhancements, and reallocate resources where needed most.

Common tools like ChatGPT build their knowledge base through "self-supervised learning". That means it is not given specific prompts but learns from the data itself. It is a powerful tool for building conversational AI because of its ability to generate human-like text that is fine-tuned to specific tasks and integrated with other technologies to create other use cases. Its responses often "sound" like a human response because it mimics the style in which it was trained.

By analyzing the reams of text data it was trained on and processing the patterns inherent in the language, text-based generative AI tools can often produce a correct and well-articulated answer. While ChatGPT is dominating the headlines, other notable generative AI models are also garnering attention.

Who is pioneering Generative AI

Academia: As with many business solutions, the foundational technology behind generative AI began in the labs. Pioneer researchers and “fathers of the Deep Learning revolution” Geoffrey Hinton, Yoshua Bengio, Yann LeCun, as well as important contributor Ian Goodfellow, made significant contributions to the field of deep learning and generative models. These included the development of Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and transformers, which are widely used in various AI applications. Their work won the 2018 Turing Award.

OpenAI: An AI startup founded in 2015 by several series entrepreneurs, including Y Combinator alum Sam Altman and Elon Musk. Altman serves as CEO. OpenAI’s core technology is Generative Pre-training Transformer (GPT), which powers the free-to-use ChatGPT-3. Consumers who pay for the “plus” service will get access to ChatGPT-4, as will other businesses that pay for the soon-to-be-launched API.

Microsoft is now using ChatGPT-4 to power a beta version of its search, which responds to queries instead of directing users to a series of links. It also has created an image-related AI technology DALL-E, which uses a version of an image-oriented modified version of GPT-3.

Google Bard: Like ChatGPT, Bard is a natural language model-based generative AI tool based on Google’s proprietary Language Model for Dialogue Applications (LaMDA). It recently announced it will soon make Bard a major component of its Workspace suite, which includes Google Docs and Sheets, Gmail, and more.

Google previously released Bidirectional Encoder Representations from Transformers (BERT); a frontrunner that demonstrated a successful range of Natural language processing (NLP) tasks based on unannotated text. Like ChatGPT, Google also has a beta version of its search powered by Bard. But it currently only allows researchers to access it.
Hugging Face: a community where AI developers share open-source models, which is now working with Amazon to make it easier for open-source models to get widespread business adoption.

Facebook LLaMA: Facebook’s newly announced 65-billion-parameter large language model was trained by text from 20 languages. It is the successor to its RoBERTa solution, similar to BERT.

Stable Diffusion: A deep learning model that creates images from text prompts. It is a product of collaboration between different partners, including Stability AI and CompVis LMU.

Midjourney: Like Stable Diffusion, it is a text-to-image model that produces images based on prompts.

More contenders will invest in or build a startup in this competitive space, including Elon Musk (again), Marc Andreessen, and others. Get your free CIO's Guide to Generative AI.

How the Major Generative AI Tools Differ

Generative AI tools differ in what they do. A simple explanation: OpenAI’s ChatGPT returns text responses to text prompts, while DALL-E returns graphics or images to text responses.

They also differ by how many parameters or variables a generative AI model has. The parameters influence how the model takes inputs and turns them into outputs.

By way of explanation, ChatGPT has 20 billion parameters compared to the 175 parameters of GPT-3, which ChatGPT is based on. But more parameters do not necessarily make a more powerful model. OpenAI has yet to announce how many parameters the just-released GPT-4 has.

Other important factors include the data quality, model architecture, how the model is tuned and improved, and, importantly, the costs to businesses depending on the complexity of their needs.

Addressing Common Misconceptions about Generative AI

It is not sentient or conscious or has feelings. Despite some salacious articles about the technology getting “upset” or professing “love” for the interlocutor, ChatGPT is not thinking about its answers. It provides the best guess answer to the query. Since it is trained on content created by humans, it can appear to mimic the tone of their responses, including seeming angry, sad, or love-sick.

It is not 100% accurate. While it is impossible to know whether some of the viral claims captured in screen grabs are real, there have been several documented instances of ChatGPT, for instance, getting basic facts wrong like what year it is, and what theJames Webb telescope pioneered.

It is not providing real-time answers. The training process currently requires an end date in the data to parse and provide a valid answer. For ChatGPT, it is September 2021. So if you ask ChatGPT who won Super Bowl LVII (i.e., The Kansas City Chiefs), it cannot tell you the answer.

It is not intended to replace humans. True, generative AI can eventually do a lot of manual tasks currently performed by humans. But like the industrial revolution automated many laborious tasks previously handled by humans, generative AI will do the same with the information revolution. It will free up humans to do more analysis or pursue other careers where the human touch is more important.

It is not just about chat and search. This misconception is due to the emergence and popularity of ChatGPT. While, yes, a new way to search for information is one of the most powerful ways to use generative AI, many more business cases are being explored. Indeed, we anticipate using the technology to make knowledge management instantaneous is one of the most exciting applications. It will also revolutionize development projects, expediting the move to low-code/no-code solutions.

It is not equitable. It is well understood that models such as ChatGPT encapsulate biases. It is important to address these biases to ensure that the models generate fair and unbiased output. AI biases can arise when the training data is biased towards certain groups and perspectives, which can prioritize overrepresented groups. This can lead to biased outputs in the language generated by the models. Feedback loops can perpetuate biases by training the models with biased output.

In our next post, we will discuss some of those practical business applications. But now that you have a better understanding of generative AI, the emerging players, how to tell the difference between models, and understand some common misconceptions, you are ready to continue learning more about this revolutionary technology.

Visit UST Generative AI solutions to learn more.