CERAWeek: Large Language Models Fuel Industry-wide Productivity

Will AI be bigger than the internet? Some experts at CERAWeek by S&P Global argue that not only will it make a greater impact, but it will also free up employees from tedious busy work.

Bill Vass, vice president of engineering at Amazon Web Services (AWS), said during a March 20 session at CERAWeek that he sees a future where the large language models (LLMs) that make generative AI possible are as ubiquitous and helpful as spell check.

“These large language models, they're going to be like spell check. If you start thinking of them that way, that they're going to be in everything you do, they're going to be part of every design plan, every document you write, every presentation you do,” he said.

And in much the same way that spell check has not put writers out of business, generative AI coding capabilities won’t replace software engineers, he said. Amazon CodeWhisperer, for instance, handles the drudgery of writing software code, he said.

As a software engineer, Vass said, “the hard part is figuring out what you want to do. And once you've figured it out, it’s busy work to do the coding.”

LLMs have accelerated user uptake of generative AI, reaching 100 million users within two months, said Schneider Electric CEO Peter Herweck, compared to the seven years the World Wide Web took to reach 100 million users.

“AI has now arrived in everybody's living room,” Herweck said.

He said the use of AI is now a question of how quickly a company can adopt it into its operations and products.

“I think it's big. It's bigger than the World Wide Web, particularly with the large language models and generative AI of course in combination with many other AI models,” Herweck said.

Vass said the cloud has made it possible for people to leverage AI capabilities. Permanent cloud storage of data is one reason, he said.

“I really encourage you to save all your data. That data is gold. The greater amount of data you have, the greater density models you can build, the more parameters and the more weights between those parameters, the more accuracy you will get” from generative AI, which is based on statistics and math, he said. 

LLMs could also make existing systems more productive, Herweck said.

For example, a user with 100 turbines can use Schneider’s software and voice recognition to request a comparison of turbine performance. While the capability to unearth that information has resided in the system for a long time, it was difficult to extract, to the point that only about 20% of the system’s capabilities were actually used, he said. 

“Now if you add some of those large language models to it, all of a sudden you can unleash value that is already implemented in your sites,” Herweck said.

World of possibility

Vass said the industry is moving towards being software-defined and away from having systems defined by programmable logic controller and supervisory control and data acquisition systems.

A major benefit of transitioning is that software-defined systems enable full-fidelity digital twins and pave the way to better optimization, he said.

“Once you have that, you can apply machine learning to optimize it and then the cycle starts over again. That optimization, the software goes back into the real world, generates more data and that cycle just continues,” Vass said.

AWS is pulling massive amounts of data from suppliers, logistics, manufacturing and customers.

“They feed all of that along with a bunch of synthetic data—all the possibilities into a large language model watching and optimizing their entire enterprise. So they're moving to a software-defined enterprise that's ML-enabled in the large language model. That's the most aggressive thing I've seen,” he said.

Human intelligence and documentation needs to be part of a company’s digital transformation journey, Herweck said.

“You need to, in your digital transformation, take the people with you at the end of the day, and also try to … document how decisions are made by AI,” he said.

That documentation is crucial because of the possibility of liabilities if wrong decisions are made, he said. “Make sure you have very good documentation that can be done automatically, but do documentation so … people in the future can go back and learn from mistakes,” Herweck said.

A work in progress

While there are many potential uses for AI and generative AI, some areas still need work.

“I can't overemphasize the cybersecurity need for that, and the control of your data and the ability to maintain encryption and have it in a virtual private cloud,” Vass said.

Attribution also needs some focus.

“We really would like intellectual property providers to want to contribute to the models and get compensated when their intellectual property is used in the models rather than be conflicting with that,” he said.

Overtraining a model on a specific domain can create catastrophic forgetfulness, which is an inability to cognitively respond in those areas, he added.

Hallucinations, or incorrect output, remain problematic and are a reason Herweck believes human intelligence should remain part of the workflow.

For all these reasons, Vass said, a multi-model environment would be useful.

“I think you're going to have models checking models,” he said.