ChatGPT finally gives AI its moment in the sun

ChatGPT is an AI chatbot, trained using petabytes of data to generate human-like responses to prompts from actual humans. Developed by the research laboratory OpenAI, it's not the first system of its kind. But what sets ChatGPT apart – and made it an overnight sensation – is that OpenAI released it with a user interface that allows the public to interact with ChatGPT directly, for free. Within four days of its debut, ChatGPT had over one million users. By January 2023 that number had shot up to 100 million, making it the fastest growing consumer application ever. That popularity, in turn, has awakened organizations to the potential of ChatGPT and its peers.

At the 18th Annual Emerging Technology Summit, Trevor Upton, Mosaic Sector Head for Transportation and Semiconductors at KeyBanc Capital Markets, moderated a panel discussing the opportunities and challenges of AI applications like ChatGPT. Upton was joined on the panel by Robert Claus, Chief Technology Officer at DataChat Inc.; Maya Mikhailov, Founder of SAVVI AI; and Glenn Ko, CEO of  Stochastic.

The panel kicked off by looking under the hood of ChatGPT.

Transformer models change the AI game

The "GPT" in ChatGPT stands for Generative Pre-trained Transformer. As Stochastic’s Ko explained, a transformer is a type of deep neural network architecture that Google introduced in 2017. It differs from earlier deep learning models in its ability to train itself rather than requiring costly and time-consuming human intervention. This is where the "P" in GPT comes in. Once a transformer model pre-trains itself on a foundation of general knowledge, only a small amount of additional "labeled" data is required to fine-tune it for specific uses. For this reason, transformer models are also known as foundation models.

"If you're doing email classification, you take a pre-trained model, provide some examples of placing e-mails into the proper categories, and deploy it," said Ko. "That makes it easy for people to produce a high-precision, top performing model without having to construct these really large data sets." That differentiator changed the landscape of AI, with transformers soon replacing all of major AI deployments in big tech organizations. It is now the most widely used architecture.

But just because transformer models represent a vast improvement over their predecessors doesn't mean they're perfect.

You get out what you put in, for better and for worse

A large language model like ChatGPT gets its context from the data it's trained on. Releasing an application built on one of these models requires first figuring out exactly what you want it to do. Once that use case is defined, not only do users have to teach pre-trained models how to serve that intended purpose, they must also add parameters that govern how the system behaves.

"You want to be really careful that it's not going to go do random things that it gleans from the original training data but that don't apply to your application," explained DataChat’s Claus. "For example, before releasing ChatGPT to the world, OpenAI had to teach it not to use profanity."

Adding these guardrails to transformer models, though, is like babyproofing your house – you may keep your child from getting into specific types of trouble, but they may still find a way to cause mayhem.

Data curation will determine whether chatbot deployment is a dream or a disaster

Since November, some of ChatGPT's weirder interactions have grabbed headlines, like when a New York Times reporter thought it was trying to convince him to leave his wife. "It's putting together a combination of words that it thinks you want to hear," said Mikhailov. "It looks like it has intent. It looks like it loves you and like it can replace your spouse, but it's just a large language model. It is not a sentient being."

While most of these exchanges have been inconsequential, they show the system's unpredictability and the organization's lack of control over the output. Both issues raise concerns for business leaders as they grapple with how to leverage tools like ChatGPT. That's why data curation – the ongoing process of determining what information the AI is given access to – will make or break the widespread adoption of these applications by businesses.

"If you train a large language model on all the emails and texts that your company has generated for the last decade," said Mikhailov, "that system is going to show you things that you’d like to forget." In addition to old skeletons and HR complaints, chatbots trained on poorly curated data may dredge up regulatory and confidentiality headaches.

"I was talking to someone recently about the kind of data that they were putting into some of these models, and I found out that they're inputting data that may be proprietary to their organization," said Mikhailov. "Just think about the data that you're feeding it, and don't just say, 'I'm going to feed it everything and hope it comes up with some insights.'"

If you want to know how to use ChatGPT, ask ChatGPT

One of the most popular uses for AI recently has been to accelerate the conceptualization of creative works. Graphic designers, writers and artists are increasingly turning to tools like DALL-E, another OpenAI creation, to shave time off of projects.

"That's where I see the industry heading, and that applies to pretty much any task where you're generating content," said Claus. "A model can get you 80% of the way there and then you just have to fill in the rest of the gap."

But how do you decide whether your organization needs to build its own models?  

"When we help companies launch AI models, we ask them, 'What sort of data matters in your industry?'" explained Mikhailov. "Sometimes they say, 'I want to build a model, but I don't know what's important.' Well, you know what you can do? You can ask ChatGPT. 'Hey, ChatGPT, what are some things that might affect ecommerce cart conversion rates?' And it will tell you a pretty plausible answer. And that gives you a place to start."

With ChatGPT, AI is just warming up

Almost as soon as it launched, people went crazy for ChatGPT, and that energy quickly spread to corner offices around the country.

"At the end of December, all of a sudden, we were inundated with emails from people saying, 'We need an AI strategy,'" said Mikhailov. "These models have been around, but they've always been behind the scenes. But ChatGPT really connected on a very visceral level and all of a sudden the world woke up to, 'Whoa, that's AI.'"

Business leaders may not know yet whether or how they'll use the technology, but many will see AI's potential to change the competitive landscape of their industry. "Companies are looking around and saying, 'I see now that this brings efficiency, that this can make our workforce faster and smarter, that this could upskill people. What is our strategy as a company to compete in this new world where people are using tools like ChatGPT to compete against us?'" concluded Mikhailov. "I think this moment is the beginning of an AI transformation across all industries."

To learn more

About our industry expertise, visit www.key.com/technology or to discuss this topic and more as it relates to your business, connect with our KeyBanc Capital Markets technology experts.

For more information on KeyBanc Capital Markets, visit key.com/kbcm